Chat or Agent: When to Hand AI the Work, and When to Just Ask

TL;DR: We’ve crossed from asking AI questions to handing it real work. The skill that matters now isn’t picking the smartest model. It’s knowing when to work in a chat window, when to work in an agent that touches your actual files, and when to put two models on the same job so one does the work and the other checks it. The rule I use: if the output ends up in a folder, use an agent. If the output is a thought, stay in chat. When the stakes are high, use both.

The copy-paste shuffle you already know

Here’s a workflow a lot of us fell into without noticing.

You run a piece of analysis in one model. Say a research summary in Claude. The answer’s good, but you’ve learned not to trust the first pass on anything that matters. So you copy the whole thing, the original question and the output, and you paste it into a second model. ChatGPT, usually. And you ask the obvious question: what did this miss?

It finds something. You paste that back into the first model. Refine. Then back again. Four or five round trips, two browser tabs, a lot of copying. By the end you’ve got a better answer than either model gave you alone. You were the wire between them. The relay runner carrying the baton back and forth by hand.

I did that dance for the better part of a year. Then it stopped being necessary.

From asking to handing off

Two things happened at once, and they’re easy to miss if you only ever open AI in a browser tab.

The first is that these tools stopped suggesting and started doing. For two years, AI handed you text and you did something with it. Now the agent versions act on your real work. Claude Code and Anthropic’s new Cowork can read a folder on your machine, write the files, run the steps, even take over your screen when there’s no cleaner path. OpenAI’s Codex can copy a project into a sandbox, work on it for an hour while you do something else, and hand back the finished change. You stop describing what you want done. You hand off the doing.

The second is that the manual relay I described got automated. I’ve wired my Claude and Codex setups together so one can call the other without me in the middle. I run analysis in Claude, and in the same command it sends the work to Codex for a second read and brings the feedback back. I’ve connected Google’s Gemini the same way, through its new agent interface, so I can pull a third model in when a question is worth it. The copy-paste shuffle that used to eat an afternoon is now one instruction. Three models on one problem, no tabs, no baton-passing.

That’s not a niche trick. It’s where the whole market is going. Codex is running about four million developers a week. Claude Code passed two and a half billion dollars in annualized revenue, more than doubling since January. Anthropic says the quiet part out loud: 2025 was the year this changed how developers work, and 2026 is the year it changes knowledge work. One industry analyst put the shift cleanly. The bar moved from generating content to executing workflows, and the three big AI labs are all racing to own the part that does the work.

So the question for anyone running a firm isn’t really “which model is smartest” anymore. It’s “where should this work happen, and who checks it.”

The split that sorts most of it

I’ve landed on a rule that holds up across almost everything I do. Two buckets.

If the durable output belongs in a folder, use an agent. A client summary built from forty PDFs. A spreadsheet. A deck. An updated matter file or knowledge base. Code. Anything you’ll save, version, and come back to. This is what Claude Code, Codex, and Cowork are built for. They work against the real state of your files instead of handing you a copy to paste somewhere.

If the output is mostly a thought, stay in chat. Advice. A first draft. Brainstorming. Turning a pile of research into something you can act on. A quick comparison. The language for a client conversation before it ever becomes a document. Claude on the web and ChatGPT are faster and more fluid here, and more comfortable when you’re thinking out loud or shaping prose line by line. My early-morning writing happens in a chat window, not a terminal. That’s the right tool for that job.

And when the stakes are high, use both. One model implements, the other reviews. This is the automated handoff I described, and it’s where the real edge is. Draft a firm policy in one model, have a second one tear it apart for gaps before it reaches the committee. The two together catch what either one alone would wave through.

Here’s the concrete version for a firm. Need to turn a stack of discovery documents into a structured summary you’ll file and reuse? That’s folder work. Hand it to an agent. Trying to decide how to frame your AI policy to a roomful of skeptical partners? That’s a thought. Stay in chat and think it through. Writing the actual policy that will govern the firm? High stakes. Draft it in one model, break it in another, then have a human read every line.

Where this bites

None of this is free, and the risks are exactly the ones a law firm should care about.

More power means more blast radius. An agent that can write files can write the wrong ones. Cowork runs in an isolated space on your machine and asks before it deletes anything, which helps. But the screen-control piece is still an early research preview, and Anthropic itself says don’t point it at anything touching client finances or health records yet. These tools have already had real security holes. Cowork shipped with a data-leak bug that surfaced days after launch. Codex had a critical flaw where a poisoned code branch name could hand an attacker your access token. For most people that’s a footnote. For you it’s confidentiality and privilege. Treat agent access to client data as a decision the firm makes on purpose, not a default you back into.

Two models agreeing is not the same as two models being right. A second opinion shrinks your blind spots. It does not replace judgment. Models can be confidently wrong together, in stereo, and a tidy consensus can talk you out of the doubt you should have kept. The human still owns the call.

And not everything needs the heavy machinery. Agent sessions burn through usage far faster than a chat does, and they take longer to set up. Half the time you don’t need a sandbox and a review loop. You need an answer. Reaching for the workbench when a question would do is its own kind of waste.

What to do Monday morning

You don’t need to rewire anything to get value from this. You need to start sorting.

1. Split your recurring work into two piles: the work that ends as a file and the work that ends as a thought. Run the file work in Claude Code, Codex, or Cowork. Keep the thinking in a chat window. That one sort decides most of your tool choices for you.

2. Take one deliverable that matters, an AI policy, a client memo, a valuation summary, and run it through two models on purpose. One to draft, one to tear it apart. Then a human signs off. See what the second pass catches that the first one missed.

3. Before any agent touches real files, set the rails. Decide which folders it can write to, what it has to ask before deleting, and draw a hard line: nothing client-confidential or privileged goes near an agent until your firm has vetted exactly how it’s handled.

The part that’s still yours

We spent two years getting good at asking AI better questions. That skill still matters. But it’s not the whole game anymore.

The tools can do the work now. They can check each other’s work. You can put three models on one problem before breakfast. What none of them can do is decide what “done” means, or whose name is on the result when it leaves the building. That part didn’t get automated. It got more important.

The relay runner can finally put the baton down. The call at the finish line is still yours.

Share Intelligence by Intent

Six Percent: The AI Hallucination Number That Surprised a Room of Family Lawyers

TL;DR: I asked a room of family law attorneys to guess how many of the world’s tracked AI hallucination cases involve family law. Their guesses ran from 20 to 90 percent. The real number is 6. The more useful finding is who’s behind those cases, and what it tells you about where the risk is heading.

Bring up AI hallucinations in a room full of lawyers and you can feel the air change. People sit up. A few arms cross. Everyone has read the headlines by now: the sanctions, the lawyer who filed a brief built on cases that never existed.

I was at the AAML National Family Law Conference in Las Vegas this week, giving the AI keynote, and hallucinations came up. They always do. A few weeks back I was on a panel with the LA County Bar Association, and there it was again. It’s the question that won’t sit down.

So in the keynote, I tried something. I asked the room to guess what share of the tracked hallucination cases worldwide involved family law. Hands went up. Twenty percent. Fifty. Someone said ninety.

The real number is six.

Three things worth saying out loud

Early in the talk I put up a slide with three things I think are worth saying before anyone gets comfortable. Here’s the first, and it’s not the controversial one: the risk is real. Mata v. Avianca was not the last sanction, not even close. Real attorneys are still filing briefs full of cases that don’t exist, and they’re paying for it with real money and real public discipline.

The second is the one that makes people shift in their seats. Avoidance is not safety. I watch this one land differently than people expect. Your opposing counsel is already using these tools. Your clients assume you are too. Refusing to touch AI doesn’t shield you from a malpractice claim. Using it badly is what gets you there. That reframe does more work than any horror story about a sanctioned lawyer.

And then the quiet one underneath both. The bar is moving. Technology competence is part of the duty of competence now, and the question stopped being whether you use AI. It’s whether you use it well.

Where the number comes from

None of this is a hunch. There’s a research fellow at HEC Paris, Damien Charlotin, who has been tracking these cases globally, and doing the best job of it I’ve seen. He keeps a running database of them. The charts are worth an hour of your time. And the part I appreciate most is that he doesn’t paywall any of it. The whole dataset is free to download, so I did. I encourage you to go explore his site!

As of this week, he’s tracking 1,659 decisions. The curve is steep. Two years ago you’d see a handful of these a quarter. Lately it’s a few hundred. Most of the fabrications are case law: confident citations to opinions that were never written. And in the minority of cases that name a tool, one comes up far more than any other. ChatGPT. Not because it’s worse than the rest, but because it’s the one people reach for. A general-purpose assistant, doing a job that really needs a legal research database.

Who’s getting caught

Here’s where it got interesting for the family law crowd. Of those 1,659 cases, 99 are family law. That’s the six percent. But I wanted to know who was behind them, so I split the 99 by who filed the bad document.

Sixty-eight of them, 69 percent, came from people representing themselves. Not lawyers. Self-represented litigants.

That tracks once you think about who those people are. Someone handling their own custody case reaches for what’s free and in front of them: a search engine, a consumer chatbot. They don’t have Westlaw or Lexis. And they have no way to look at a clean, confident citation and sense that the case behind it doesn’t exist. An attorney builds that instinct over years, reading enough opinions to feel when something’s off. A parent fighting for custody on their own has no reason to have built it.

The harder problem is coming for the courts

And that’s the part that isn’t really about lawyers at all. If self-represented litigants are the main source of these fabrications, and the tools keep getting easier to use, the volume of confident, fake-citation-laden filings landing on court dockets is going to climb. Judges and clerks are the ones who’ll absorb it. I don’t have a clean answer for how the courts handle that. I just know it’s coming, and faster than most of the system is ready for.

If you’re a family law attorney, six percent should be a relief. It mostly isn’t you. But don’t let it turn into comfort. The number is low today. It’s also climbing across every field, and the bar for competent use keeps rising whether or not your practice has caught a sanction yet. Low risk is not the same as no risk. And it’s a long way from no responsibility.

What I’d do Monday

If you take three things back to your firm, make it these.

1. Match the tool to the task. A general-purpose assistant is great for a first draft, a summary, a way to think out loud. It is not a citation engine. When you need case law, use a database built for it. Most of the fabrications in the data come from using the wrong tool for the job.

2. Put a human on every output, and make the check a real one. Not a skim. Someone opens each cited case and confirms the holding says what the brief claims it says. That last part matters more than people think. A meaningful share of these cases aren’t invented cases at all. They’re real cases bent into saying something they never said, and a fabricated cite and a misread one fail the same way in front of a judge. Put it in writing as a policy, not a good intention.

3. Build a real final pass. A fact-checking step, a skill, a checklist, something that catches a fabricated cite before it reaches a judge. It backs up the human review. It does not replace it.

Here’s the thing the headlines keep missing. These tools are getting better fast. They draft, they summarize, they can read through tens of thousands of documents before you’ve finished your coffee. That’s real, and it’s worth using. But none of it touches the part that matters. Clients aren’t paying you to produce text. They’re paying for your judgment. Your read on the case. The thing you catch that the model never will.

That part isn’t going anywhere. It’s the reason someone hires you instead of asking the chatbot themselves.

TL;DR: California’s State Bar has proposed AI amendments to six Rules of Professional Conduct. None of them are new duties. All of them make clear that competence, confidentiality, candor, and supervision already cover how your firm uses AI. Public comment closes August 6, 2026. The smart move is to get a written AI policy in place now, because one of the proposed rules turns that from a nice-to-have into a managerial expectation.

You saw the headlines. A lawyer files a brief. The cases look perfect. Clean citations, real-sounding holdings, the works. Except the cases don’t exist. The model made them up, nobody checked, and now there’s a sanctions order and a name in the press.

And you thought: not at my firm.

Here’s the thing. That instinct was right. It just didn’t have rules behind it yet. California is about to fix that.

What California actually proposed

On June 12, 2026, the State Bar’s ethics committee approved a second round of proposed AI amendments to the Rules of Professional Conduct. This traces back to a letter from the California Supreme Court in August 2025, directing the Bar to consider folding its 2023 generative AI guidance into the rules. The comment window closes August 6.

The changes touch six rules: competence, communication with clients, confidentiality, candor toward the tribunal, and the two supervision rules covering managing lawyers and nonlawyer staff.

One detail matters more than the rest. Every change lands in the comments, not the black-letter rules. The operative text doesn’t move an inch. Which tells you what this really is. Not a new AI rulebook. A statement that the duties you already carry apply to AI, full stop. The “the rules don’t mention AI” defense is going away. And that is the whole point.

Getting the facts right is now explicit

Start with the obvious one. The candor rule says verify your authorities before you file. No fabricated cases, no misquoted holdings, no cases pulled out of context, and that includes anything AI touched. Fine. Expected. You already knew that.

But the competence rule reaches further. The duty to verify follows the citation, not the courtroom. A fake case in a demand letter is the same violation as a fake case in a brief. Same for the memo to your client and the email to the other side. Anywhere a cited authority leaves your firm, someone had to confirm it was real. The hallucination problem was never just a courtroom problem.

Your AI tools just became a confidentiality test

This is the one with teeth.

Under the proposed comment, putting client confidential information into an AI system can count as revealing it, when there’s a substantial risk the information gets accessed, retained, or used in a way that breaks confidentiality. So the associate who pastes client facts into a consumer chatbot to summarize them might be making a disclosure, not saving an hour. The tool that trains on those facts is the exposure.

The test is whether that risk is substantial, judged against three things: whether the tool retains or trains on your data, the security around it, and whether anyone else can reach the information. It’s not a ban, and it doesn’t require zero retention. Most serious commercial tools hold your data about thirty days for abuse monitoring, then delete it. That window isn’t the problem. Training on your inputs is. So is data sitting where other users or third parties can get to it. A vetted enterprise tool that stays out of training, controls access, and purges on a schedule is fine. A free account someone signed up for last Tuesday, learning from every word typed into it, is not. That gap used to be an IT preference. Now it’s an ethics question.

When do you tell the client?

The communication rule adds something your partners will argue about. Your duty to keep clients informed now includes weighing whether to tell them you’re using AI on their matter.

It doesn’t order you to disclose AI use to every client on every file. People will read it that way, and they’ll be wrong. The rule asks you to make the call on purpose, against how new the tool is, the risks and benefits, the scope of the work, and how sophisticated the client is. A bank’s general counsel and a first-time divorce client are not the same conversation. The only wrong answer is not having made the call at all.

Supervision is a leadership problem now

This is where it gets close to home for firm leadership.

The managing-lawyer rule says you must make reasonable efforts to put internal policies in place governing AI use. In writing. The nonlawyer rule says your staff’s AI use is your responsibility to instruct and supervise. So the paralegal running a chatbot, the assistant using a tool you’ve never heard of, the contract reviewer who found something free online: that lands on the supervising lawyer. Not in theory. By name.

If your firm has no written policy and no idea what tools your people are using, this is the rule that should get your attention.

Yes, these are only comments

I can hear the pushback. These are just comments. They aren’t even final.

Both true. The comments don’t carry the same force as black-letter rules, and the proposal still has to clear public comment, the committee, the Board, and eventually the California Supreme Court before it binds anyone. So nothing here is enforceable tomorrow morning.

But that misses the point entirely. The underlying duties, competence, confidentiality, candor, supervision, are enforceable today. They always have been. The comments just take away your cover for pretending AI lives somewhere outside them. Waiting for the rule to pass is waiting for permission you already don’t have.

One more gap worth naming. The Supreme Court specifically asked about agentic AI, the tools that plan and act on their own with little human touch. These proposals mostly treat AI as general technology and don’t take on autonomous agents directly. The Bar put its agentic-AI thinking in the updated Practical Guidance instead. That’s where the harder questions are still being worked out. If you’re moving toward systems that draft and file with minimal human review, plan for that conversation too, not just this one.

What to do Monday morning

You don’t need a task force. You need three things done.

1. Write the policy and train your people to it. A short AI use policy people actually read beats a long one that sits in a drawer, and it should set your default for when you tell clients about AI use. Your nonlawyers need to know what they can use, what they can’t, and what never goes into a tool. The managing-lawyer rule is about to make all of this an expectation, so build it now while it’s still your idea.

2. Vet your tools before client data goes in. Check retention, training, and security, and get the answers in writing. If the vendor can’t explain in plain English what happens to your data, that’s your answer.

3. Verify every citation, wherever it’s headed. Court filings, client memos, letters to the other side. If AI touched it, a human confirms the authority exists and says what you claim it says.

And if you have a view on the rules themselves, you have until August 6 to say so. The comment form is open.

AI didn’t create new duties for your firm. It just made the old ones impossible to ignore.

What the Heppner AI Ruling Actually Means for Your Firm

TL;DR. Your firm can confidently use commercial AI, including Claude Team or Enterprise and ChatGPT Business or Enterprise, as an everyday part of practice. The Heppner headline scared a lot of lawyers out of that, but the case was about a consumer tool used without a lawyer’s direction, under terms the court found inconsistent with confidentiality. Your commercial terms are materially different: the provider doesn’t train on your content and treats it as confidential. They won’t make every chat privileged, and for most of your work they don’t need to. With the right settings and matter controls, these tools belong in your firm’s daily work.

You probably saw the headline. “Federal judge rules AI conversations aren’t privileged.” Maybe a partner forwarded it with a one-line note: “Should we be worried about this?”

And for a second, you wondered. You’ve got associates running research through Claude, someone in litigation support summarizing depositions with ChatGPT, the whole firm quietly building these tools into the day. Now a federal judge is putting the words “not privileged” on the front page.

The bottom line: your firm can use these tools with confidence

Here’s the good news, stated plainly. If your firm runs Claude Team or Enterprise, or ChatGPT Business or Enterprise, under the applicable commercial terms and firm-approved settings, your lawyers can use these tools as a normal, everyday part of practice. Research, drafting, summarizing, review, working through a problem. Use them. The defendant in Heppner independently used a publicly available AI service, without counsel’s direction, under terms the court found inconsistent with a reasonable expectation of confidentiality. That is not the normal law-firm workflow addressed here.

Why is that defensible? Because the threshold question for everyday firm use is usually not whether every prompt becomes privileged. It is whether the firm has taken reasonable steps to protect information relating to the representation. Commercial terms provide the core protections missing in Heppner: no model training on customer content absent agreement, restricted use, and contractual confidentiality. Paired with appropriate vendor diligence, settings, matter-level controls, and supervision, that gives a firm a strong basis for using these tools consistently with its confidentiality obligations.

The honest boundary is narrower than the headlines suggest. A commercial contract gives you a strong confidentiality framework, not blanket privilege. Privilege can still matter when an AI workflow contains or facilitates lawyer-client communications. In litigation, qualifying prompts and exploratory analysis may also receive work-product protection, with counsel-created prompts presenting the strongest case. The rest of this piece is about getting those distinctions right.

What the case was actually about

The case is United States v. Heppner, decided by Judge Jed Rakoff in the Southern District of New York. A represented criminal defendant, Bradley Heppner, used a publicly available version of Claude on his own, without counsel’s direction. His lawyers later argued that some of his inputs incorporated information he had learned from them. After a grand jury subpoena, knowing he was a target, he generated 31 documents working through his defense. Possible arguments. Facts. Strategy. Then he shared them with his lawyers.

Federal agents seized the documents. Heppner’s lawyers said they were privileged. The judge said no.

He gave more than one reason, and that matters for what comes next. The chats weren’t communications with his lawyer, because Claude isn’t a lawyer. They weren’t confidential, because the version he used ran under a consumer privacy policy that let the company collect his inputs, use them for training, and disclose them in specified circumstances, including to regulators and in litigation. And although his lawyers said he meant to use the material in later conversations with them, they hadn’t directed the exercise, and Claude wasn’t their agent. Rakoff said the chats failed at least two, and maybe all three, requirements for privilege.

Then it got worse. The court added that, to the extent any information Heppner entered was privileged when counsel communicated it to him, he waived that privilege by sharing it with Claude and Anthropic.

So the real lesson, buried under the scary headline, is narrow. Drop your lawyer’s advice into a consumer chatbot and you may have placed that advice outside the privilege and exposed it to government or third-party inspection. Worth telling clients. Not a reason to pull AI out of your firm.

And here’s the part almost nobody quoted. In dicta, not in his holding, Rakoff floated a path: when counsel selects the system, directs its use, and uses it for a defined job in support of legal advice, the tool might be treated like another professional assisting counsel, the way a translator or an accountant sits inside the privilege. That wasn’t a ruling, and telling a client to “use Claude” wouldn’t be enough on its own. The safer pattern has lawyer selection, a defined legal purpose, confidential commercial terms, supervision, and use limited to that one matter.

Five questions hiding in one headline

Untangle something the coverage keeps mashing together. There are five different questions in these AI stories, not one.

Attorney-client privilege asks whether there was a confidential communication for legal advice between counsel, the client, and any qualifying agent. Work product asks whether material was prepared in anticipation of litigation by or for a party or its representative, and, where mental impressions are involved, whose strategy or analysis it reveals. The ethical duty of confidentiality asks whether the firm took reasonable precautions with all client information, privileged or not. A protective order can forbid an upload even when privilege and ethics rules would allow it. And discovery and preservation ask whether prompts, outputs, uploads, and activity logs are records that must be preserved, collected, or produced. A tool can pass one of these and fail another. Keep them separate and the rest of this gets easier to read.

Why commercial terms change one part of the math

Hold the Heppner facts next to what your firm uses.

The consumer privacy policy was one of the things that sank him. Collect, train, disclose. That’s the opposite of the commercial plans.

Under their standard commercial terms, absent the customer agreeing otherwise, the providers don’t use your workspace content to train their models. Anthropic’s commercial terms put Claude for Work, meaning Team and Enterprise, outside the consumer training policy and treat your content as your confidential information. OpenAI draws the same line for ChatGPT Business and Enterprise: no use of your content to develop or improve the service unless you explicitly agree, plus confidentiality obligations. Both commercial agreements incorporate a data processing addendum, the contract your privacy and security people already know how to read.

So what does that do to the Heppner analysis? It changes one part of it: confidentiality. No-training commitments, restricted use, contractual confidentiality, and deletion controls give you a much stronger fact pattern than Rakoff faced.

But here’s where I had to correct my own first instinct. Better terms don’t satisfy the other privilege requirements. You still need a communication with counsel, or a qualifying agent, for the purpose of getting legal advice. The contract gives you a materially stronger confidentiality argument. It does not, by itself, turn a chat into an attorney-client communication. And that data processing addendum, useful as it is, is evidence of careful vendor handling, not a privilege agreement.

One more thing. Approving the core platform isn’t approving everything bolted onto it. Both providers separate their own service from third-party apps, connectors, and actions, which carry their own terms and recipients.

The case that matters most for your firm’s controls

If you read one decision after Heppner, make it Morgan v. V2X.

Colorado, March. An employment case, a plaintiff representing himself, both sides using AI. The defendant wanted limits on what the plaintiff could feed into chatbots, and the magistrate judge did something useful: she wrote AI language into the protective order.

Under her order, you can put confidential discovery material into an AI tool only if the provider is contractually barred from storing your inputs or using them to train the model, from disclosing your inputs except where essential to running the service, and from using downstream providers unless they’re bound by equally protective terms. The provider also has to let you delete the information on request. And you have to keep written proof of all of it.

Read it carefully, though. The order didn’t say enterprise accounts are automatically safe. It said a qualifying enterprise-tier account could fall on the permissible side, that many ordinary consumer subscriptions may not, and that the product label alone proves nothing. The contract and the configuration do the work. That is still highly useful for a firm: a judge drew a contract-based line and recognized that an enterprise-tier account might qualify if its actual contract and configuration satisfy every requirement.

But look at the first item again. Training or storing. That word does a lot of work, and it’s where I’d slow you down. A no-training promise is not the same as no storage. A standard commercial workspace may retain chat history so you can use it, even while it refuses to train on it. That can be entirely appropriate, as long as the retention is deliberate and consistent with your policy and preservation duties. But Morgan’s language isn’t a national standard. It was written for one order, and read literally, its ban on “storing” could rule out ordinary enterprise chat tools that keep history without training on it. Don’t agree to that wording without reading it. And don’t tell a client your tool is “zero retention” unless the contract actually says so.

The case for your represented clients

Here’s the one your research-minded partners will care about most, and almost nobody is talking about it yet.

Most of the decisions so far involved someone representing themselves or a criminal defendant acting alone. The far more common setup is a represented civil litigant whose own AI conversations are at issue. In June, the Texas Business Court took that up. In Tate Group Automotive v. Legacy Automotive Capital, Judge Grant Dorfman reviewed the ChatGPT conversations of the plaintiff company’s principal, a non-lawyer, in a commercial dispute.

Tate is the clearest post-Heppner ruling involving a represented commercial party’s own AI conversations. Dorfman protected most of the principal’s chats under Texas’s work-product rule and held that using ChatGPT did not waive that protection. He expressly disagreed with the defendants’ reliance on Heppner as waiver authority and emphasized that Texas’s rule protects material prepared “by or for a party.” The minute entry did not decide attorney-client privilege. Be honest about its weight, too: this informal entry says on its face it isn’t a final ruling, evidence of where this is heading, not binding precedent.

It wasn’t a free pass either. The court made the plaintiff hand over the pages that weren’t real work product and, borrowing from Morgan, identify all discovery materials or products it had shared with ChatGPT, by Bates number where applicable. That’s the pattern worth remembering. The analysis can be protected. The fact of which produced documents you uploaded may not be.

The lawyer-use case everyone overlooks

There’s one more, and it predates the 2026 run, which is probably why it gets missed. It’s also the most on point if what you’re worried about is your own lawyers.

In Tremblay v. OpenAI, lawyers used ChatGPT in a pre-suit investigation. The court held that counsel’s prompts, and the testing that didn’t pan out, were opinion work product, because the questions a lawyer chooses reveal how that lawyer is thinking. That’s about as strong as protection gets. But the same case marks the limit: the prompts and outputs the plaintiffs leaned on in their complaint had to be produced. Lean on the AI testing in your allegations, or put the testing itself at issue, and you can lose the protection. So an associate’s exploratory prompts can be opinion work product. The screenshot you paste into a filing is a different thing. A later California decision, Concord Music Group v. Anthropic, took the same approach with attorney-created prompts and outputs.

The cautionary tale nobody at your firm should repeat

The last case is the one your clients will hear about at a conference, so be the one who explains it. And start with what it is: Fortis was not a privilege ruling. It was a warning about discoverability, intent, and preservation.

In Fortis Advisors v. Krafton, a Delaware Chancery case from March, Krafton’s CEO, Changhan Kim, was looking for a way around an earnout of up to $250 million. He turned to ChatGPT. When ChatGPT initially told him the earnout would be difficult to cancel, he kept pushing. At ChatGPT’s suggestion, Kim formed an internal task force called “Project X,” and Krafton subsequently followed most of the chatbot’s recommendations. The chats became trial exhibits, the court relied on them in finding Krafton’s later justifications were pretextual, and Kim admitted he’d deleted some of the relevant logs.

That’s the risk picture in one story. AI chats are records. They can be sought in discovery, subpoenaed, or pulled from devices and company systems. They can become the best evidence against you. And deleting them once litigation is anticipated can create a preservation problem worse than the original conversation. None of which argues against using AI. It argues for treating it as potentially discoverable ESI, not private scratch paper.

Where this honestly leaves you

I’m bullish on this, and I’ll tell you why in a second. But here are the caveats straight, because anyone who skips them is selling you something.

No appellate court has yet squarely resolved whether generative-AI prompts or outputs are protected by attorney-client privilege or work product. No reported decision squarely holds that a firm using ChatGPT Enterprise or Claude Enterprise gets privilege. Several of the early work-product wins involved people representing themselves, though that’s shifting now that Texas has extended the reasoning to a represented party. And privilege, work product, the ethical duty of confidentiality, and a protective order stay separate questions. A tool can pass one and fail another.

On work product, the direction is real. Warner, Morgan, Tate, and Assini, the last a June New York decision that quashed a broad subpoena to OpenAI for a self-represented litigant’s prompts, uploads, outputs, drafts, legal research, strategy, and account materials, point the same way: using an AI provider doesn’t by itself waive qualifying litigation-preparation material. The harder questions are whether it was genuinely prepared because of litigation, whose thinking it reflects, and whether sending it out made an adversary meaningfully more likely to get it.

So I wouldn’t promise a client that every chat is privileged. The sturdier, more useful claim: in litigation, the stronger argument isn’t privilege, it’s work product. Your prompts show which facts you think matter, which theories you tested, where you think the case is soft. That’s the mental-impression material the rules were built to protect. The cases since Heppner support that protection when the material was prepared by or for a party because litigation was anticipated. When the prompts reflect counsel’s own selection of facts, theories, and questions, the case for protected opinion work product is stronger still. Keep that separate from a business person brainstorming, which is how those Krafton logs became evidence.

One more limit. Don’t assume an expert’s prompts get a lawyer’s treatment. In Conservation Law Foundation v. Shell, a magistrate ordered production of the prompts a testifying expert used to cull Shell’s document production into a working set for the report. That order is stayed pending review, so it isn’t settled, but experts may sit in a different discovery category than your lawyers.

And the confidentiality from a commercial contract is real. It’s the same diligence you already run for cloud storage and contract attorneys, pointed at one more vendor. The ABA lands in the same place: under Formal Opinion 512, a lawyer using these tools still owns competence, confidentiality, supervision, client communication, verification, and the protection of client information. None of that says ban the technology. It says review the vendor and put reasonable safeguards in place, which is what a commercial contract and a real policy give you.

What to do Monday morning

You don’t need a six-month study. You need three decisions.

1. Approve the exact product, and document it. “Claude is approved” isn’t a policy; “Claude Enterprise on our contract” is. Name the plan and account type so nobody does sensitive work on a personal login, and remember that approving the platform doesn’t approve every app or connector attached to it. Keep the agreement, the data processing addendum, and your retention settings in one place, because Morgan says you may have to prove those protections one day, and don’t take the defaults on faith.

2. Check the matter, then put a lawyer on the sensitive work. A firm-approved platform doesn’t override a protective order, a client guideline, or a data-use restriction, so someone has to check those before discovery material, trade secrets, or medical records go into a tool. And when privilege or strategy is in play, have the attorney pick the tool, define the task, supervise the work, and review the output. That is the pattern Rakoff suggested might support treating the tool as counsel’s agent, although no court has yet turned that suggestion into a general safe harbor.

3. Treat the chats as evidence. Fold AI prompts and outputs into your litigation holds, and teach everyone a basic preservation rule: once a duty to preserve has attached, do not delete.

The dividing line was never free AI against paid AI. It's unmanaged use against an approved, contractually protected system, under settings and matter controls matched to the work. Get that right and the question stops being whether you're allowed to use these tools and becomes how well you use them, because for everyday work the duty isn't privilege, it's protecting the client's information, and your commercial terms are built for that. If you want a second set of eyes on where your firm draws that line, before a protective order or a partner draws it for you, I'm at steve@intelligencebyintent.com. Send this to the partner who forwarded you that headline.

Share Intelligence by Intent

Is Claude or ChatGPT as Safe as Microsoft Copilot? What the Commercial Terms Actually Say

TL;DR: On the business and enterprise tiers, the privacy commitments from Microsoft 365 Copilot, Claude, and ChatGPT line up more closely than most people expect. None of the three train their models on your data by default. All three will sign a Data Processing Addendum, encrypt your data, and offer a HIPAA agreement, though only on specific surfaces (more on which ones below). Retention works a little differently at each, but the shape is similar: your data is either controlled by your own admins or deleted on a short cycle. For the most sensitive work, a true zero-retention setup exists, but mostly on the API or by special arrangement, not as a switch you flip in the standard chat apps. The brand on the box matters far less than the contract you’re on and the tier your people are using. The real exposure is someone running client work on a consumer account.

Here’s the question I get, almost word for word, from the IT teams that support law firms and valuation shops:

“We already pay for Microsoft 365 and Copilot. Before we let anyone touch ChatGPT or Claude for client matters, how does their security and confidentiality actually compare to Microsoft?”

It’s a fair question. And it usually carries an unspoken assumption underneath it: that Microsoft, the incumbent the firm already trusts with its email and its documents, must be in a different class than the two AI startups.

So I went and read the terms. Not the marketing pages. The Product Terms, the Data Protection Addendum, the commercial privacy policies, the trust portals, the retention docs. Side by side. Business and enterprise tiers only, because consumer terms are a different animal and no firm should be running client work on a consumer account anyway.

Here’s what I found.

Where Copilot, Claude, and ChatGPT agree

Start with the part that surprises people. On the commercial tiers, the core confidentiality promises line up closely.

They don’t train on your data. This is the one everyone worries about, and on the commercial tiers all three land in the same place. Microsoft is explicit that prompts, responses, and anything Copilot pulls from your Microsoft Graph aren’t used to train the models. Anthropic says the same for its commercial products by default. OpenAI says it for Business, Enterprise, and the API. Three vendors, one position. Feedback is the one place the rules diverge: Microsoft can use it to improve the product but not to train the model, while a thumbs-up or thumbs-down to OpenAI or Anthropic can pull that conversation in. So govern the feedback setting on its own.

You own what you put in and what comes out. All three assign the rights in the output back to you. (Whether a purely machine-generated output can be copyrighted at all is a question for the courts, not the vendor, but that’s true no matter which tool you pick.)

The legal and security plumbing matches. Each one will sign a Data Processing Addendum and act as your processor, supports GDPR through Standard Contractual Clauses, encrypts data in transit and at rest, and has been through independent audits. So on the contract mechanics a security reviewer cares about, you’re not choosing between “serious” and “not serious.” They’re all serious.

One caveat on HIPAA, since healthcare-adjacent matters come up. All three will sign a Business Associate Agreement, but only on certain surfaces, and not the entry-level ones. Microsoft covers Copilot under your commercial tenant’s agreement, with web search queries excluded. Anthropic signs for its first-party API and for sales-assisted Claude Enterprise once an admin opts in, but not for Claude Team, the Console, or Cowork. OpenAI signs for sales-managed ChatGPT Enterprise and the API, but explicitly not for ChatGPT Business. So if protected health information is in play, the tier you pick matters as much as the vendor.

That’s the headline. Now the differences, because they’re real and they shape how you write your policy.

Where they differ: data retention

This is the part worth slowing down on, because “how long do they keep my data” is the question that decides a lot of vendor reviews.

Microsoft keeps your Copilot prompts and responses inside the Microsoft 365 service boundary. Think of it as living in the same place your email and SharePoint files already live, captured as activity history in a hidden folder in each user’s Exchange mailbox, governed by the same compliance and eDiscovery tooling. It’s encrypted, and you decide how long it sticks around with a Purview retention policy: 30 days, a year, seven years, whatever your records schedule already says. Your users can delete their own Copilot history. So with Microsoft, retention isn’t a number Microsoft sets. It’s a dial your admins hold in Purview, and like the dials at the other two, it only does anything once someone actually sets it.

Anthropic splits along how you connect. On the API, Anthropic deletes inputs and outputs within 30 days by default. On Claude for Work and Enterprise, the conversations stay in the product so people can pick up where they left off. Worth calling out for any reviewer: on Enterprise, the default is to keep that history indefinitely until an admin sets a custom retention period, where 30 days is the floor. So retention is yours to control, but you have to actually go set it. For the strictest work, Anthropic offers a Zero Data Retention arrangement, where your inputs and outputs aren’t stored at all beyond what’s needed to screen for abuse. But ZDR isn’t a toggle in the chat app. It applies to the API, products that run on your API key, and Claude Code on Enterprise, and it’s granted by agreement, not switched on per user.

OpenAI lands in the same place. ChatGPT Enterprise and Business put retention in your admins’ hands, and deleted conversations clear OpenAI’s systems within 30 days. The API defaults to 30 days too, with zero retention available on eligible endpoints by approved use case. Same story as Anthropic: zero retention lives on the API, not in the chat app.

Read those three paragraphs again and the pattern jumps out. Two models, really. Either your own administrators control retention (Copilot, Claude Enterprise, ChatGPT Enterprise), or it’s a short auto-delete cycle with a zero-retention escape hatch (the APIs). That’s it. The idea that one of these vendors is quietly hoarding your data while the others don’t isn’t supported by the terms.

And this isn’t a theoretical exercise anymore. The question “how long do you keep AI prompts and outputs” is now showing up in outside counsel guidelines, client security questionnaires, and RFPs. A few years ago it didn’t come up. Now it’s a line item. The retention choice you make isn’t just internal hygiene. It’s something you may have to put in writing and stand behind when a client or insurer asks.

Most IT teams I work with end up screenshotting this exact comparison:

The twist: Copilot runs the same engines

Here’s the part that reframes the whole question.

Microsoft 365 Copilot is built on OpenAI’s models, delivered through Azure rather than the public ChatGPT product. And Microsoft has since made Anthropic’s Claude models available inside Copilot too, in certain surfaces and regions. So when a firm asks me whether they should trust Microsoft over ChatGPT and Claude, the honest answer is that Copilot is running the same model families, wrapped in Microsoft’s contract and held inside Microsoft’s service boundary.

What you’re really choosing between, then, isn’t three different engines so much as three different contracts and three different places your data sits. Microsoft’s pitch is that the data never leaves your service boundary, and that you govern it with tools you already own. OpenAI and Anthropic’s pitch is that they hold the data on their own infrastructure but don’t train on it and delete it on a schedule you can tighten. For a firm that has already standardized on Microsoft, the boundary argument is genuinely worth something. Just know what you’re buying: a wrapper and a boundary, over models you could also reach directly.

One wrinkle worth flagging, because your IT team will ask. When Claude runs inside Copilot, Anthropic operates as a Microsoft subprocessor, Microsoft notes that those Anthropic models currently sit outside its EU Data Boundary commitment, and some preview models run under separate Anthropic terms. If EU data residency is a hard line for you, that’s a detail to check, not assume.

The bigger risk: the wrong tier

Now the thing I care about most, and the reason I tell every firm to read this section twice.

Everything above is true only if your people are on the business or enterprise contract. The consumer versions, ChatGPT’s personal plans, Claude’s Pro plan, the free Copilot you get when you sign in with a personal account, are governed by completely different terms. On the consumer side, the defaults flip. Anthropic now uses consumer chats to train its models unless the user opts out, and retention for opted-in consumer accounts runs to five years. Free and personal ChatGPT can be used for training unless the person turns it off. None of that comes with a Data Processing Addendum, which means no contractual confidentiality protection for your clients’ data at all.

So the dangerous scenario in a law firm isn’t “we picked the wrong vendor.” It’s “an associate pasted a deal document into a personal ChatGPT account at 11pm because it was faster than logging into the firm tool.” That’s the exposure. The vendor choice barely moves the needle next to the tier choice.

This is the line I’d put in bold in any AI policy: client work goes through the firm’s commercial account, on the commercial contract, every time.

The honest tradeoffs

They’re not identical, but the differences are about fit, not safety. Each one has an edge that matters to some firms and not others, and none of them makes the other two a weaker choice for your clients.

If you’re already a Microsoft shop, Copilot’s draw is that it inherits the governance layer you already run: the same data loss prevention rules, sensitivity labels, and eDiscovery console your records team already lives in. You’re extending a control plane you’ve configured rather than standing up a new one. That’s worth something if you’re deep in Microsoft, and not much if you aren’t.

Anthropic’s draw is the clearest language on employee access, staff can’t read your conversations without your consent or a safety flag, plus a separate FedRAMP-authorized product, Claude for Government, if the public sector ever enters the picture. OpenAI’s draw is customer-managed encryption keys and the widest set of data-residency regions, handy if you have specific in-country storage requirements. All three now hold ISO 42001, the AI-specific management standard, so that one’s table stakes across the board. And every copyright indemnity here carries the same fine print: keep the safety features on, don’t knowingly publish infringing output. Read it before you lean on it.

None of those is a safety advantage. They’re preferences. On the questions a managing partner actually asks, do they train on our data, will they sign a DPA, can we control retention, will they stand behind us on copyright, the answer for all three is the same: yes.

What to do Monday morning

1. Write the tier rule, then enforce it technically. Put in writing that all client work runs on the firm’s commercial or enterprise account, and back it up with single sign-on so people can’t easily use personal accounts for firm matters. The policy without the enforcement is just a wish.

2. Pick your retention posture and set it. Decide how long the firm wants to keep prompts and outputs, then actually set it: in Purview for Copilot, in the admin console for Claude or ChatGPT Enterprise, or ask about zero retention for your most sensitive practice groups.

3. Get the paper in the file. Make sure you have the signed DPA, the BAA if you handle health data, and the current audit reports for whichever tools you’ve approved. When a client’s outside counsel guidelines or an insurer asks how you govern AI, you want to answer in an afternoon, not a month.

The vendors are closer than the noise suggests. Your job isn't to find the one safe AI company. It's to make sure your people are on the right contract, with the right settings, each and every time. Get that right, and the brand on the box is almost a footnote.

Share Intelligence by Intent

Fable 5 was here for three days. Law firms should pay attention.

TL;DR: Fable 5 was available for only a few days before the U.S. government suspended it. The question is no longer whether models will get good enough for serious legal work. They already are. The question is whether firms can test, govern, and use them before the next jump catches them flat-footed.

The three-day window

Every once in a while, you use a new model and it feels different right away. That was my reaction to Claude Fable 5.

Anthropic released Fable 5 on June 9, describing it as a Mythos-class model made safe enough for general use. Its launch note said the model exceeded anything Anthropic had made broadly available before. Fable was built on the same underlying model as Mythos 5, with safeguards for higher-risk areas like cybersecurity, biology, chemistry, and model distillation.

Three days later, on June 12, Anthropic suspended access. The company said the U.S. government had issued an export-control directive blocking access by any foreign national, including foreign nationals inside the United States and even Anthropic’s own foreign-national employees. Because Anthropic could not enforce that kind of restriction cleanly in real time, it took Fable 5 and Mythos 5 offline for everyone.

The public story is national security. Fine. For law firms, the more useful story is simpler: you may get only a few days to understand a tool before policy, capacity, or governance changes the menu.

For a few days, lawyers got a look at what happens when the model stops feeling like a very smart associate with uneven judgment and starts feeling like something closer to a senior thinking partner that can stay with the work.

What felt different

My own experience with Fable 5 was that it was a significant leap over Opus 4.8, even when using Max Thinking.

The writing jumped. So did the reasoning and analysis. By “better,” I don’t mean prettier sentences. I mean it seemed to understand the job behind the words.

It could take a messy set of facts and find the line of argument inside them. It could see why one point mattered more than another. It could hold more of the problem in its head.

That matters because the best legal writing is not just correct. It is selective. It knows when the reader needs law, chronology, or the human logic that makes a fact pattern make sense.

Think about litigation. A trial team is not merely collecting evidence. It is deciding what the case is about.

Is this a contract case about words on a page, or is it really about a party trying to rewrite the deal after the fact? Is the bad email the centerpiece, or is it a distraction from a cleaner pattern in the documents?

Those are judgment questions. No model should answer them alone. But a model that can help lawyers pressure-test them is a very different tool from one that simply drafts a brief.

Where I would test it first

If Fable 5 comes back, I would not start with demos. I would start with work the firm already knows how to judge.

I would start in four places:

• Argument development: give the model the facts, law, bad facts, and other side’s best position, then ask it to build and attack the argument.

• Evidence assessment: have it map documents, testimony, chronology, and inconsistencies into a testable case theory.

• Storyline development: ask it to explain the case as a judge, jury, client, regulator, or opposing counsel might hear it.

• Draft revision: use it not just to rewrite, but to explain what is weak, what is missing, and what the draft is asking the reader to believe.

That last point is where many firms still underuse AI. They ask for output. They should ask for editorial judgment.

“Make this better” is a small prompt.

“Tell me why this argument is not yet persuasive, identify the three places where the logic depends on an unstated assumption, and rewrite only the section with the most risk” is a different kind of work.

Fable 5 seemed much better at that kind of work. It was less eager to decorate the page and more able to reason.

For a managing partner or COO, that means the business case is not just faster drafting. It is better use of lawyer attention. If a partner can get to the real issue in 20 minutes instead of two hours, the value is that the partner spends more time making judgment calls and less time clearing brush.

The shutdown matters too

The shutdown should make law firm leaders pause.

Not panic. Pause.

According to Anthropic, the government directive was based on national security concerns tied to a potential jailbreak. Anthropic has said the issue was narrow, non-universal, and involved finding a small number of previously known, minor software vulnerabilities. It also said other public models could perform similar work without the same bypass.

The Associated Press reported that more than 100 cybersecurity leaders and experts urged the administration to lift the directive, arguing that taking these models away from defenders could help adversaries. Business Insider, citing Politico, said Anthropic staff met officials to try to resolve the dispute.

The record is still moving. I would not build a whole column around guessing who wins that fight.

But I would take one thing back to the management committee: access is now part of the risk analysis.

First, there is model risk. The better the model gets, the more it can help with both good and bad uses.

Second, there is data risk. Anthropic’s Mythos-class retention note says prompts and outputs for these models are retained for 30 days for trust and safety purposes on every platform where the models are offered. Anthropic says the data is not used to train new Claude models and is deleted after 30 days except in rare cases. That may be fine for some work and barred for other matters.

Third, there is access risk. A model can be available on Tuesday and gone by Friday.

Most firms are not set up for that. Vendor selection can no longer be a once-a-year procurement exercise. It has to become an active operating question: which model, for which work, on which platform, and with which backup plan?

What I would do Monday morning

If Fable 5 returns, firms should be ready. Not with hype. With a sober test plan.

1. Pick five recent matters or work products the firm can safely use for testing.

2. Remove or mask client-sensitive information unless governance has approved the platform.

3. Run the same tasks through the firm’s current model and the new model.

4. Score the outputs for legal reasoning, factual care, writing quality, usefulness, and review time.

5. Decide which work categories are approved, prohibited, or sandbox-only.

This is not about crowning a winner on a public benchmark. It is about one practical question: does this model materially improve how our lawyers work, under conditions we can defend to clients, courts, insurers, and ourselves?

If the answer is yes, the firm should move quickly. But quickly does not mean casually.

For client matters, especially litigation, investigations, employment disputes, trade secrets, M&A, regulated industries, and anything with a protective order, the retention and access rules need to be read carefully. General enthusiasm is not a privilege analysis.

At the same time, firms should not use governance as a polite way to avoid learning.

There is plenty of work that can be tested safely: public filings, old anonymized briefs, training hypotheticals, internal knowledge projects, marketing drafts, and sanitized case-theory workshops. A firm that waits until every question is answered before it starts learning will be behind the firms that learned in a sandbox first.

The part that sticks with me

Fable 5 may or may not come back in the same form.

But the direction is clear.

The next generation of legal AI will not just summarize documents and draft memos. It will help lawyers think through arguments, challenge evidence, build narratives, and prepare for the reader.

That does not make lawyers less important.

It makes lawyer judgment more visible.

Because when the machine can produce a decent draft, the lawyer’s value moves up a level. What is the theory? What should we concede? What should we fight? What will the judge care about? What will the jury remember?

Those are not typing questions.

They are lawyering questions.

For three days, Fable 5 gave us a glimpse of a tool that could help with those questions in a much more serious way than the models most lawyers are using today.

So yes, the shutdown matters.

But the more important story is what those three days revealed.

The capability jump is real. The governance burden is real. And the firms that learn to hold both ideas at the same time will be better prepared for what comes next.

Fable 5 may come back in the same form, or it may not. But the direction does not depend on one model. The next generation of legal AI will not just summarize documents and draft memos, it will help lawyers think through arguments, challenge evidence, and prepare for the reader, which makes lawyer judgment more visible, not less. The firms that learn to hold two ideas at once, that the capability jump is real and the governance burden is real, will be the ones ready for whatever lands next. If you want to talk through what a sober test plan looks like for your firm, reach me at steve@intelligencebyintent.com. The model was here for three days. The decision it forces is here to stay.

Five Things Google Needs to Fix Before I Can Recommend Gemini to Law Firms

I want to recommend Gemini to my clients. I really do.

Gemini 3.5 Flash is part of my daily routine now. NotebookLM is one of the best tools I’ve ever handed a client. The new Omni video model is genuinely something to see, and the weird little Labs app called Dreambeans has somehow become my favorite thing Google has shipped this year. And 3.5 Pro is coming any week now, which I’m waiting on with more anticipation than I’d admit out loud.

So this isn’t a hit piece. It’s the opposite. It’s a list of fixable things, written by someone who roots for the product and keeps running into the same walls when a managing partner asks me, “Should we standardize on Google?”

Here’s the honest answer I have to give right now: not yet. And the reasons aren’t about the models. The models are great. The reasons are about how Google packages, prices, and protects those models across its different tiers. The sprawl has gotten ahead of the rollout, and the people paying the price are exactly the business customers Google should want most.

Let me walk through the five.

What this is, in plain English

Google sells Gemini three ways: a consumer plan, a Workspace plan that comes with your company’s Google apps, and a separate Enterprise edition most people don’t even know exists. You’d assume the business tiers are the most capable. They’re not. Features show up on the consumer account first, arrive late or never on Workspace, and land in some third configuration on Enterprise. Same brand, three different products, and the gaps hit law firms, private equity shops, and valuation firms right where it hurts.

That’s the whole story. Now the specifics.

Model selection is a mess across tiers

Start with the consumer app. As of today, June 14, 2026, you get 3.1 Flash Lite, 3.5 Flash, and 3.1 Pro. Better still, you can dial the thinking level up or down: standard or extended. If you’re an Ultra subscriber like me, you also get Deep Think.

One note before I go further. Everything I’m about to lay out about the three tiers comes from the accounts I actually hold and log into. Google moves these menus around constantly, so what I see today might shift by the time you read this. That’s sort of the point.

Now flip to a Workspace account. Your choices are 3.5 Flash, 3.5 Thinking, and 3.1 Pro. At I/O back in May, Google announced 3.5 Flash and, as far as I can find, never mentioned anything called “3.5 Thinking.” From everything I can piece together, 3.5 Thinking is just 3.5 Flash with extended thinking switched on. So why the new name? And here’s the part that actually costs business users something real: on Workspace, you can’t get 3.1 Pro with extended thinking. That combination, the strongest model with the deepest reasoning, is the one business customers want most, and it’s the one they can’t have.

I’ve started telling clients that if they need 3.1 Pro with high thinking, go use Google’s AI Studio with an API key. That works. But think about what I’m saying. I’m routing paying Workspace customers out of their own product to get the capability they’re already paying for somewhere else.

Then there’s Enterprise, the flagship aimed at the biggest customers. The models there? Auto, 3.1 Pro, 3.5 Flash, and 2.5 Pro. No thinking levels at all. Read that again. The enterprise tier ships an “Auto” mode, includes a year-old 2.5 Pro, and gives you no way to toggle thinking to pull the best work out of these models. This is the version Google is selling to the largest organizations, and it offers the least control of the three.

Three tiers. Three model menus. None of them consistent. If I can’t explain to a partner in one sentence which model they’re actually getting, that’s a problem.

Projects exist, but not where firms need them

I’ve loved NotebookLM since the day it launched. Google keeps adding to it, and it has quietly become the center of what Gemini calls “projects,” the same idea ChatGPT and Claude built their own versions of. I like the approach. But there’s a catch, and it’s a privacy catch.

On the consumer side, NotebookLM gives you something rare: your chats inside a notebook are private. Genuinely private. Anything you create or discuss in NotebookLM stays out of training, and that holds even on the free version. But here’s the trap. Go into Gemini, add that same notebook as a source, and start a chat. That conversation is now eligible for training, unless you’ve gone in and turned all activity off. The chat shows up in NotebookLM just like the private ones. Same place, same look. But because you started it in Gemini, it gets thrown into the training pile. Almost nobody knows this split exists.

Now, where are projects most valuable? For business customers. A law firm wants one project per client, or one project per client per matter. That’s the dream setup. And on Workspace, you don’t get the same dedicated Notebooks-in-Gemini project surface the consumer app has. NotebookLM itself is there as a separate app. But the project container, the one that lets you swap the underlying model and organize your work by matter, isn’t wired in the same way. For my legal, PE, and valuation clients, that’s not a small gap. That’s the feature they’d use the most, missing from the tier they actually pay for.

Enterprise does something different again. You can add NotebookLM as an agent and reach it through a combined interface. Sounds nice. But you can’t use it the way a consumer can, swapping the underlying model, running a chat, building a true project. So that’s three different answers to one simple question: can I organize my work by client? Consumer says yes, with a privacy footnote. Workspace says no. Enterprise says sort of.

Memory is missing from the tier with the most users

Memory is one of the most useful things any AI tool does. I turned personalization on for my consumer account the moment it was available and connected everything I could. I know the privacy-minded folks reading this just winced, and I get it. But I understand the tradeoff, Google already has most of this data on me anyway, and the payoff is real. When I ask Gemini something, it knows my context. When Dreambeans builds my morning stories, the insight is sharp because it actually knows me.

Now turn to Workspace. You can resume a past chat inside a single app now, which is a start. But the real memory, the kind that learns a profile of you and carries it across every conversation the way consumer does, isn’t there. No profile-level memory, no option to turn it on. Set against ChatGPT and Claude, both of which carry memory into their business products, this is a glaring hole. I know memory is hard to build and harder to secure properly in a business setting. But if the competition has solved it, Google can.

And I know Google can, because the Enterprise edition has it. There, memory can pull from past chats, connected apps, and saved data. So the capability exists inside the house. It just hasn’t been shipped to the Workspace app, which has to hold the largest base of business subscribers by a wide margin. Every tech firm I’ve worked with runs on Google apps. Every one of them is missing this.

Connectors are a grab bag that changes by the week

This one moves so fast that it might be different by the time you read it. Google keeps bolting connectors onto different tiers with no obvious logic.

At the consumer level you get the Workspace apps, plus Search, YouTube, YouTube Music, Google Photos, GitHub, OpenStax, Canva, Contacts, Instacart, OpenTable, and Verify AI. A genuinely random mix.

Workspace gives you the Google apps plus Chat, then YouTube Music but not YouTube, GitHub, Asana, HubSpot, Salesforce, Verify AI, and Mailchimp.

Enterprise is a different list entirely: Apollo GraphOS, Asana, Box, the Google apps, Clinical Trials, Confluence, Crypto, DocuSign, Excalidraw, GitHub, GoDaddy, Hugging Face, Jira, Linear, Microsoft Learn, Notion, OneDrive, Outlook, SharePoint, Sites, Slack, Teams, and Trivago.

Look closely and the seams show. Take Salesforce. On Workspace it’s there, but it’s barely there. The connector does one thing: it finds a Salesforce contact from an email address. That’s it. “Find my Salesforce contact named John.” For a CRM that runs the revenue operations of most firms I work with, a single contact lookup isn’t a connector, it’s a toy.

Now go to the Enterprise app, the one a firm subscribes to and a user actually works inside. Salesforce isn’t there at all. Not a thin version, not a lookup, nothing you can switch on. Yes, Salesforce can be wired into the Gemini Enterprise platform on the back end, as a data store an admin sets up in Google Cloud. But that’s a different thing from what a person sitting in the app can connect to on their own. The user-facing Enterprise product, the tier built for the biggest, most CRM-dependent companies, gives an end user no way to reach the most important CRM there is. I can’t construct a rationale for that. It reads like three teams shipped three lists and nobody compared notes.

I built an infographic that lays all of this out side by side, because it’s the kind of thing you have to see to believe.

The Antigravity privacy gap most clients never see

This last one is deep in the weeds, and it matters more than anything else on this list.

When you have a ChatGPT Business or Enterprise plan, or a Claude Team or Enterprise plan, and you use a coding agent like Codex or Claude Code, those sessions fall under the commercial terms you signed. Private. Not trained on. That’s the deal, and it’s clean.

Google works differently, and I still don’t fully understand why. Here’s what I found when I was writing a privacy paper and wanted to be precise. If you sign into Antigravity (the app or the CLI), with your workspace account, the consumer terms end up governing your sessions, which means your interactions can be used for training. The cleanest way to get real privacy protection is to pay for a separate developer license, the Code Assist license, or to use the API and pay per token. I’m confident almost none of Google’s customers assume this. I didn’t, and I do this for a living. I only caught it by reading the actual terms line by line.

Sit with that for a second. A firm buys Workspace believing it has bought business-grade privacy. A developer on that firm’s account opens Antigravity to write some code. And depending on how they signed in, that code and those prompts can flow into training, because the protection they thought they paid for doesn’t reach that tool. For a law firm handling privileged material, that’s not a footnote. That’s a problem you have to know about before anyone touches the tool.

What to do Monday morning

If you’re weighing Gemini for your firm, here’s where I’d start.

1. Map which Gemini tier each team is actually on, then list the exact models and thinking levels they can reach. Don’t assume the business tier is the strongest. Check it.

2. For anyone who needs the top model with deep reasoning, set up AI Studio with an API key as a stopgap, and treat it as a workaround, not the destination.

3. Before you let any developer use Antigravity, confirm how they sign in and which terms govern their sessions. If privacy matters, budget for a Code Assist license or API access.

4. Hold off on standardizing firm-wide until projects and memory reach Workspace. Run a small pilot, keep your client matters organized somewhere with real privacy in the meantime.

The bottom line

Google is building some of the most powerful models in the world. I don’t doubt that, and I’m not rooting against them. But I serve clients who need three things at once: the best models, real control over privacy, and connections to the tools they already run on. Until Google lines those up across the tier that businesses actually buy, I can’t tell a managing partner to standardize on Gemini.

3.5 Pro is almost here. I hope a few of these get fixed on the way out the door. I’d love to change this answer.

Note: I have endeavored to be as accurate as possible in building the connectivity table and statements around privacy. Any errors are fully on me. If someone sees something that is incorrect, please let me know and I will happily address in future updates!

Google's models aren't the problem. The packaging is. Until the best features reach the tier your firm actually pays for, with privacy you can prove and your client matters kept apart, standardizing on Gemini is a decision you make with your eyes open, not on faith. If you're weighing this for your own firm and want a second read before you commit, I'm at steve@intelligencebyintent.com. Save this one for the partner who's about to sign the contract.

Claude Fable 5 Is Here, and It’s Not Just Another Model Update

TL;DR: Anthropic released Claude Fable 5 today, the first model in a new tier that sits above Opus. It leads nearly every benchmark that matters, including legal, and after six hours of hands-on testing I think it’s the best model available right now. Try it before June 22 while it’s included in paid plans. But read the data retention section first, because the rules changed, and legal teams need to know that before anyone at the firm touches it.

I’ve been testing AI models for years, and most “new model” days follow a script. The company posts a chart, the numbers tick up a few points, and by Thursday nobody remembers the announcement.

Today wasn’t that.

Anthropic released Claude Fable 5 this morning, and I’ve spent the last six hours doing nothing but putting it through my actual workload. Writing. Coding. Spreadsheet work. The verdict, at least so far: this is the biggest single jump I’ve felt since I started doing this work.

What Fable 5 actually is

Here’s the short version. Fable 5 is not a new Opus. It’s the first model in a new class entirely, what Anthropic calls Mythos-class, a tier that sits above Opus the way Opus sits above Sonnet.

Anthropic built the underlying model, called Mythos, and found it was so capable in areas like cybersecurity that they initially restricted it to a small group of cyber defense partners working with the US government. Fable 5 is that same model, made safe for the rest of us. When you ask it something in a handful of high-risk areas (offensive security, certain biology and chemistry topics), it quietly hands the question to Opus 4.8 instead. Anthropic says that happens in fewer than 5% of sessions. In six hours of legal, writing, and coding work, I never hit it once.

So when you use Fable 5, you’re using the most capable model Anthropic has ever built, with guardrails on the dangerous parts. That’s the whole product.

The benchmarks, including the one your firm cares about

The numbers are not subtle. On SWE-Bench Pro, the standard test for real-world software engineering, Fable 5 scores 80.3% against Opus 4.8’s 69.2%. An eleven-point jump on a benchmark where models had been grinding out gains of one or two points at a time. Stripe reported that during early testing, Fable 5 completed a migration across a 50-million-line codebase in a single day. Their estimate for a human team doing the same work: over two months.

But here’s the result I want you to sit with. Harvey, the legal AI company, published its own evaluation today. On their Legal Agent Benchmark, which measures end-to-end completion of complex legal tasks under an all-pass standard, Fable 5 scored 13.3%. That sounds low until you see the field: Opus 4.8 sits at 10.4%, GPT-5.5 at 2.1%, and Gemini 3.1 Pro at 0.0%. On Harvey’s BigLaw Bench, Fable 5 hit 93.4%, the highest score any Anthropic model has posted. Harvey’s lawyer evaluators called out drafting and markup analysis specifically: catching off-market provisions, term sheet deviations, and inconsistencies in counterparty redlines, and tracing defined terms accurately across large document sets.

For complex legal work, there is currently no contest. One model leads, and it leads by a lot.

What six hours of testing told me

Benchmarks are one thing. Here’s what I actually saw today.

The writing is better than Opus 4.8, at least for my use cases, and Opus 4.8 was already the best writing model I’d used. It’s also fast. Noticeably, almost surprisingly fast, which is not what I expected from a model this large.

The coding has been rock solid. Every initial test I ran produced one-shot outputs well beyond what Opus 4.8 gave me on the same prompts. And inside the Excel add-in, the difference is almost comical. It’s far faster than Opus or Sonnet, it asks smarter clarifying questions before it starts working, and the output comes back polished rather than 80% done.

Early reviews match what I’m seeing. Andrej Karpathy called it a genuine step change rather than an incremental release, and the customer quotes Anthropic published, from Cursor, GitHub, Cognition, and others, all describe the same pattern: long, hard tasks that previous models couldn’t finish, now finished.

The part your general counsel needs to read

Now the tradeoff, and for my legal audience this is not a footnote.

Anthropic is requiring 30-day data retention for all traffic on Mythos-class models. All of it. First-party and third-party surfaces, regardless of how you access the model. They’ve stated clearly that the data won’t be used to train future models and won’t be used for anything beyond safety purposes, that human access is logged, and that deletion happens after 30 days in almost all cases. The stated reason is defending against novel jailbreaks and attacks that play out across many requests.

That’s a reasonable security posture. It’s also a meaningful departure from the zero-retention arrangements many firms negotiated, and it applies even if your enterprise agreement says otherwise for older models. Harvey flagged the same thing in its announcement and made Fable 5 opt-in for exactly this reason.

So before anyone at your firm runs client matters through Fable 5, your information governance people need to answer one question: can this engagement tolerate 30-day retention by the model provider? For a lot of work, the answer is yes. For some matters, it won’t be. Know which is which before Monday.

What to do this week (Monday is too far away!)

Three things. First, try it now: Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans through June 22, after which it moves to usage credits until capacity catches up. That’s a “free” window to test the best model on the market against your real work. Second, run your own bake-off. Take three tasks you did last week with your current model and rerun them in Fable 5. Compare. Third, get the retention question in front of whoever owns your AI governance policy before broad rollout, not after.

OpenAI and Google will answer. ChatGPT 5.6 and Gemini 3.5 Pro are coming, and this leapfrog game never stays settled for long.

But today, right now, the most capable AI model in the world is sitting in your Claude account. Part of your subscription, for the next two weeks.

I’d go find out what it can do.

Here’s where this leaves you. The most capable AI model ever released is sitting in your Claude account right now, part of your subscription until June 22, and it leads the legal field by a margin that won’t be ignored by the firm across the street. The only thing standing between you and a real answer is two weeks of testing and one retention question your governance team can settle in a single meeting. If you want help structuring that bake-off, or thinking through which matters can tolerate the new retention terms, reach out: steve@intelligencebyintent.com. The firms that win with AI aren’t the ones that move first. They’re the ones that move first with their eyes open.

Deepfakes Are Walking Into Court Before the Rules Are Ready

TL;DR: This isn’t a 2027 problem waiting on new federal rules to show up. It’s a today problem. The firms that still treat a video clip as presumptively real are the ones who’ll get burned, and not always by a fake. Sometimes by a real recording the other side simply calls fake. Provenance is becoming the new privilege. Build the habit before you need it.

I spent this weekend in Carlsbad at an LA County Bar Association Family Law off-site, on a panel about AI and the law. Good crowd, sharp questions, the kind of weekend where you leave with more notes than you came with.

The best conversation, though, happened off to the side.

Someone asked a question that didn’t have a tidy answer. What happens when a litigant hands the court a video that looks one hundred percent real and one hundred percent didn’t happen? Not a grainy, obviously doctored clip. A clean one. Lighting right, audio synced, the defendant clearly saying the thing he swears he never said.

I’d been chewing on this already. A couple of weeks earlier I was at Google I/O, watching them show off Gemini Omni, their new model that builds and edits video from a sentence, a photo, a few seconds of audio, whatever you feed it. It is genuinely impressive. It is also the kind of thing that should make every litigator sit up a little straighter.

Here’s what actually changed

Making convincing fake video used to be hard. It took a skilled team, real money, and time. Now it takes a prompt and a few minutes.

The security world’s estimates put deepfake files somewhere near half a million in 2023 and in the millions by last year. Treat the exact figures as ballpark. The direction is the part that matters, and the direction is straight up.

Volume is the obvious worry. It’s also the wrong one to fixate on, because something quieter is doing more of the damage.

For about a century, courts leaned on a simple assumption: a photo or a recording is presumptively what it claims to be. Authentication was mostly a box you checked on the way to the real fight. That assumption just died. And when it dies, it doesn’t only let fakes in. It hands everyone a reason to doubt the real thing.

The bigger threat isn’t the fake. It’s the doubt.

Here’s what I mean. Back in 2023, in a wrongful death case over a fatal Tesla crash, the plaintiffs wanted to use recordings of Elon Musk making big claims about Autopilot’s safety. Tesla’s lawyers argued the recordings might be deepfakes and shouldn’t count. The judge wasn’t having it. She pointed out the obvious problem. If being famous enough to be deepfaked means your own recorded words can’t be used against you, then anyone prominent can say whatever they want and wave it away later.

That case is the whole thing in miniature. The danger isn’t only that someone slips a fabricated video past a jury. It’s that every authentic video now arrives with a built-in defense. “That’s AI.” Researchers even have a name for it. The liar’s dividend. The more fakes that exist, the easier it gets to dodge real evidence by calling it fake.

Watermarks help, right up until they don’t

The reasonable hope is that the same technology creating the problem will also label it. Google stamps its Gemini Omni output with SynthID, an invisible watermark baked into the file. There’s a broader industry standard called C2PA, sometimes branded as Content Credentials, that attaches a tamper-evident record of where a file came from and how it was changed along the way.

Both are good ideas. Neither saves you.

The honest player watermarks. The person trying to frame someone, win a custody fight, or sink a deal does not. And plenty of capable tools, the ones that never make a keynote stage, won’t mark anything at all. So the absence of a watermark tells you almost nothing. You cannot build a case strategy around a label the other side was never going to apply.

The rules are coming. Slowly.

The federal rulemakers see it. There are two efforts worth knowing about, and they solve different problems.

The first, proposed Rule 707, would treat machine-generated output offered without a human expert the way we treat expert testimony, holding it to real reliability standards. Helpful. But it only kicks in when a party admits the evidence came from a machine. It does nothing about a video someone passes off as a genuine home recording.

The deepfake piece is a proposed addition to Rule 901, the authentication rule. The idea is a burden shift. If you challenge something as fabricated, you have to put forward enough to make fabrication plausible, not just say the word. Then the burden moves to the side offering the evidence to prove it’s real under a higher standard than the usual light touch. The committee has been careful here, and rightly so, because a world where anyone can stall a trial by shouting “deepfake” is its own kind of broken.

Now the catch. The comment window on the machine-evidence rule closed in February. The committee took up the deepfake language this spring. Even on the optimistic path, the earliest any of this becomes binding federal law is December 2027. Louisiana has already passed its own AI evidence verification law, so the states aren’t sitting still. But if you’re trying a case in 2026, you are working in the gap. The technology is here. The rules are not.

Before you turn every exhibit into a forensic fight

One caution, because the cure can be worse than the disease. If lawyers start crying deepfake over every inconvenient clip, dockets jam and the guilty get a free tool for muddying clean evidence. That’s the committee’s nightmare, and it should be yours too. Most of the time the boring stuff still wins anyway. Chain of custody. Metadata. A witness who can say under oath where a file came from. Those beat any detection software on the market, and they cost a lot less than the expert you’ll otherwise be hiring at the worst possible moment.

What to do Monday morning

You don’t need a new rule to start. You need new habits.

1. Treat provenance as a workflow, not an afterthought. When evidence comes in, capture the original file, its metadata, and a hash, and write down who touched it and when. The cheapest insurance against a deepfake fight is a clean record of where the thing came from.

2. Find your forensic person before you need them. Build the relationship with a media authentication expert now, while it’s a calm conversation and not a fire drill two weeks out from trial. You want a name in your phone, not a frantic search.

3. Train your team for the “that’s AI” move. Your intake staff and your litigators should both ask “where did this come from” by reflex, and your trial lawyers should expect opposing counsel to wave a hand at your best exhibit and call it a fake. Build the answer before the question lands.

None of this is about fearing the tools. The shift worth watching is quieter than any fake: the burden is sliding off the people who lie and onto the people telling the truth, and provenance is becoming what protects you, the way privilege always has. You can build that habit now, while it's still a choice and not a fire drill two weeks out from trial. If your firm is working through what this means for how you handle evidence,

I'd be glad to talk it through: steve@intelligencebyintent.com. Soon, holding the real video won't end the argument; it'll start one, and the firms that built the muscle early are the ones who'll win it.

When AI Does the Grunt Work: What Anthropic’s Self-Improvement Report Means for Law Firms

TL;DR: Earlier this week, Anthropic published its own internal numbers on how much of its engineering AI now does. More than 80% of the code it ships is written by Claude. The average engineer pushes 8x the code they did in 2024. The part worth your time isn’t the AI. It’s what broke once execution got cheap. Human review turned into the bottleneck, and judgment turned into the whole job. Your firm runs on the same shape, and Anthropic hit the wall first.

Anthropic is your firm, three years early

You’ve run the math on your own shop. A few partners at the top, a wider band of associates beneath them, and the whole thing holds together because the people at the bottom do the hours and the people at the top sell the judgment. That ratio is the business. Move it and everything moves with it: profit per partner, what you can bill, who you hire, how you train them.

Now open Anthropic’s new report on what they call recursive self-improvement, and walk right past the science-fiction headline. What’s underneath is plainer and more useful to you. Anthropic is a pyramid too. Smart people at the top, a mountain of execution underneath. And they just showed us, with their own numbers, what happens to that structure when the execution underneath it mostly disappears.

The grunt work, gone

Start with one engineer, this past April. He pointed Claude at a backlog of errors and let it run. It shipped more than 800 fixes and cut that whole class of error by a factor of a thousand. How long would a person have taken? His estimate was four years. Not because the work was hard. Because it was slow, scattered, and no human can hold that much unfamiliar detail in their head at once.

That’s not a stray win. Across the company, Claude went from writing a few percent of the code in early 2025 to more than 80% by May 2026. The average engineer now pushes roughly 8x the code per day they did in 2024. Anthropic is straight about the weakness in that number: lines of code rewards quantity over quality, so the real gain is smaller. Fine. The direction is the point.

And the direction is the lesson. The work you bill by the hour, first-pass research, the standard memo, document review, the cleanup nobody wants, is the work that gets cheap first. Not eventually. It already happened somewhere, and somebody wrote down the numbers.

Then everyone’s waiting on the same few people

Here’s the part most coverage skips, and the part that should hold a managing partner’s attention.

When the doing gets cheap, the doing stops being the constraint. Something else takes its place. At Anthropic it was review. They say it flatly: once Claude writes most of the code, getting a human to read and approve that code becomes the thing that holds up everything behind it. Your reviewers can only go so fast. Speed up everything in front of them and you haven’t fixed the line, you’ve just moved where it backs up.

You’ve watched a smaller version of this already. Technology-assisted review pulled first-pass document review away from junior associates years ago. The work didn’t vanish. It climbed. Someone senior now validates what the model flagged, defends how it was done, signs the certification. The doing got cheap. Standing behind it never did.

Anthropic is living the bigger version now. They run an automated reviewer over every code change before it merges, and they found it would have caught about a third of the bugs behind past production incidents, misses by engineers who are among the best in the world at this. Their newest model, out at the end of May, is sold partly on being far less likely to let its own errors through, so teams can lean on human review a little less. Sit with that. The product roadmap is now openly about relieving the review bottleneck, because the bottleneck is real and it’s expensive.

For you, that bottleneck already has a name: professional responsibility. You can’t put unreviewed work into a client’s matter. The duty to supervise doesn’t relax because the first draft came from a machine. So the wall Anthropic hit on speed, you hit on speed and ethics at the same time, and the second one doesn’t move.

A hundred people doing the work of a thousand

The report’s other claim is the one I’d put on the whiteboard. A 100-person company, they argue, increasingly does the work of a 1,000-person one, because each person sits on top of a stack of agents handling the execution.

Run that through a firm and the questions get uncomfortable fast. If a small team produces what used to take a big one, how big should your associate class actually be? What are you selling, once the hours collapse? And the one I keep coming back to: where will your associates learn judgment, if not by grinding through the execution that’s now automated?

That’s the quiet crisis in all of this. The grunt work was never only grunt work. It was the apprenticeship. Pull it out and you’ve solved a cost problem while handing yourself a much harder training problem, one that won’t show up on the P&L for a decade.

Where I’d push back on all this

A few honest caveats, because this report has a tell.

Anthropic is grading Anthropic. A company that sells AI, telling you how good AI is at its own work, on a metric it admits flatters the result. Take the trend seriously. Take the exact multiples with salt.

And software isn’t law. Code comes with tests that tell you in seconds whether it runs. A settlement posture or a regulatory read has no green light that flips on when you’re right. The execution that automates first in your shop is the most mechanical part, and the judgment that resists it is a far bigger slice of legal work than it is of writing code. That’s a real cushion. Don’t mistake it for a wall.

The trend could also bend. Anthropic says so themselves: the curve might flatten, the hard part of judgment might not yield to more compute. They don’t believe it. But they said it, which beats most vendors.

One last thing, since it’s in the headlines. Anthropic is not calling for a pause on AI. They’re asking for the ability to coordinate one later, the machinery that would let several labs in several countries stop together and check that the others actually stopped. They’re blunt that one company pausing alone does almost nothing but hand the lead to someone less careful. If someone tells you “even Anthropic wants to stop,” they’ve got it backward.

What to do Monday morning

Three moves. None of them needs a technology budget.

1. Find your review bottleneck before you buy a single tool. Name who reviews AI-assisted work, count the hours they really have, and be honest about whether they can keep up. Speeding up the front of the line while the back stays fixed just builds a faster pile-up.

2. Rebuild the apprenticeship on purpose. If associates used to earn their judgment by doing the work a machine now does, design a path to that judgment deliberately. It won’t happen by osmosis anymore.

3. Trace where your fees attach. Split your work by how much is execution and how much is judgment, then ask what’s left worth charging for when execution runs close to free, and where your name, and your liability, actually sit.

The bottom of the pyramid is hollowing out, and that part isn't really up for debate anymore. What's still open is whether you redesign the top on your own terms, before the cheap execution does it for you. If you want a second set of eyes on where your firm sits, reach me at steve@intelligencebyintent.com. The firms that move first won't be the ones with the biggest tech budget. They'll be the ones who saw the shape early.

Share Intelligence by Intent

What Law Schools Just Learned About AI, and Why It’s a Warning for Law Firms

TL;DR: A Stanford study had law professors blindly grade answers to student questions, and they picked the AI over their own colleagues about 75% of the time. Read it as a tutoring story and it’s a curiosity. Read it as a story about how lawyers actually get made and it turns into a warning, because the two things that have always built a lawyer, the grind work and the hours spent next to someone who knows more than you, are getting automated in the same stroke. Skip the part where you replace them and you’ll be running a firm full of associates who are quick, fluent, and can’t tell when they’re wrong.

How lawyers actually get made

Think back to how you became good at this.

Odds are it didn’t happen in a classroom. It happened at 2 a.m., in a document review, the night you finally saw the pattern nobody had bothered to explain. It happened when your first real memo came back from a partner so marked up you could barely find your own words underneath. It happened in a hallway, when someone with twenty years on you said one sentence that made a doctrine click in a way three years of school never had.

That’s the apprenticeship. Grind plus proximity. You did an enormous volume of junior work under supervision, and you absorbed judgment by standing close to people who already had it. Every lawyer reading this was built that way.

A new study out of Stanford is being passed around as a story about law students and AI tutors. It’s getting the easy headline. But read it as a measurement of that apprenticeship, and it says something most firms have not noticed: the ground the next generation trains on is washing out from under them, and it’s going from two directions at once.

What they actually did

The setup is worth a minute. Sixteen contracts professors, fourteen law schools. They wrote the questions a 1L actually asks after class, answered them in their own words, then turned around and judged the answers blind. Roughly 3,000 head-to-head pairs. Some from other professors, some from AI, nobody told which was which.

The AI won about 75% of the matchups. It held its own with the best human in the room, and the weaker humans got buried. The number that stopped me wasn’t the win rate, though. It was harm. When a judge marked an answer as the kind that sets a student back, the human answers drew that flag more than three times as often as the AI’s. 12% against 3.5%.

And keep in mind what the task was: pick the answer you’d actually rather hand a student, all things considered, not the one that scores best on a rubric. On settled doctrine, the stuff with a known shape, the one they kept reaching for was the machine’s.

The training ground is disappearing

Those two pathways, grind and proximity, are both being automated. And this study is the proof of the second one.

Start with the grind. The junior work that taught you, document review, cite-checking, first drafts, research memos, due diligence, is the exact work these tools now do in minutes. Firms are already shrinking first-year classes and arguing about the shape of the pyramid. Whatever you think about the economics, watch what it does to training. Partners romanticize this part, the war stories, the all-nighters, and I roll my eyes at most of it. But strip out the nostalgia and the point survives: the work that made lawyers was the work nobody wanted to do, and it’s the first thing to go.

Picture the change at the individual level. A first-year used to lose a weekend to a research memo, and the losing was the point. The dead ends, the case that looked perfect until the last paragraph, the slow build of a feel for where the law actually lives. Now that memo takes twenty minutes and a decent prompt. The work gets done. The learning that used to come with it does not.

Now the proximity. For decades, the answer to “I don’t understand mutual assent” or “remind me how the parol evidence rule works” came from a person, a professor in office hours, a senior associate down the hall. That’s the clarification layer. The Stanford study just showed that for those questions, a frontier model gives an answer experts rate higher than what a rushed human produces, and is less likely to mislead.

So both rungs of the ladder are being sawed off. The boring reps that built pattern recognition. The on-demand explanation that filled the gaps. A 25-year-old with a JD used to climb from one to the other. Take both away and put nothing in their place, and you don’t get a faster lawyer. You get a fluent one who never learned to judge.

Why the machine won

This is where I’d slow down, because the reason the AI won is the actual warning.

The professors lost at recall-and-explain. Asked to perform in three minutes, with no time to research, they did the same thing the model does: pulled up a known rule and explained it cleanly. It’s very good at that, and it doesn’t get tired, annoyed, or pulled into a partner meeting.

But recall-and-explain is the cheapest kind of legal knowledge to acquire and the least valuable to own, precisely because AI now owns it too. And it’s the thing junior training still rewards most. We hire for it. The bar exam is mostly a test of it. And years in, we still light up at the associate who can recite the standard cold.

Sit with what that means. We’re spending enormous effort turning young lawyers into excellent versions of the one skill that just got automated. The recall-and-explain associate is a depreciating asset. The machine reset the price of it to roughly zero, and plenty of firms are still paying full freight to build it.

The part that wasn’t on the test

Now ask what the study didn’t measure, because that’s the part clients still pay a premium for.

It didn’t test knowing a settlement number is wrong when the spreadsheet swears it’s fine. Or reading a general counsel well enough to know she wants the careful answer this quarter, not the clever one. Nothing in it touched the instinct to run a conflict check before anyone asks, or the feel of a deposition when a witness shifts in his chair and you know the case just turned.

That’s judgment. It gets taught by reps on real stakes and by proximity to people who have it, the two things AI is now eating. And the machine that explains the parol evidence rule beautifully will still invent a case citation with total confidence. I’ve written about that verification gap before, and it has not closed. So the human in the loop isn’t going anywhere. They are moving up, from the person who knows the rule to the person who knows when to break it and is willing to put a name on the call.

A gap is opening between those two roles. I’ve written about the verification gap; this is a different one. Call it the apprenticeship gap: the distance between automating the work that used to train lawyers and building anything to replace it.

The objections, and where they hold up

Some of the pushback is fair, so let me take it head on.

The humans were rushed. Completely true. They had three minutes and no research, and a partner who sits down and thinks would beat the model on most of these. But that cuts the wrong way for comfort. The rushed expert is exactly the version a junior usually gets, and the machine is the version that’s always there and never irritated. No one here showed that AI can practice law. What they showed is smaller and more uncomfortable: on settled doctrine, an always-available machine out-explains a distracted human, and the distracted human is the one most juniors actually reach. That is a real hole in how you train them.

We’ve seen tech panics before, and people who cry wolf about automation tend to be wrong. I make my living on this technology, so weigh the next sentence accordingly. The mechanism here is not spreadsheets replacing paper ledgers. It’s the training itself going away, which is a different kind of problem, and a quieter one, because nobody files a memo the day an apprenticeship stops working.

And yes, this is one model family, one casebook, contracts, short answers. It is not proof that AI can run a matter. It’s a clean look at one rung of how lawyers learn. Worth noting, too: the strongest models on the market now are a step past what the study even tested. Its top performer was Claude Opus 4.7. Today’s leaders, Claude Opus 4.8 and GPT-5.5, are newer still, and the gap to human instructors in this kind of task has only widened.

What to do Monday

None of this is “form a committee.” It’s three habits, and they’re harder than they sound.

The first is just honesty about your own training reps. Walk through what your juniors actually do to learn, and mark the ones the machine now does for them. If AI writes the first research memo, nobody learned anything by writing it, and pretending otherwise gives you an associate who believes they’ve done the work. Find the reps that still build judgment and guard them like they cost something, because they do.

The second is to move your senior people off explaining and onto judging. Partners answering “what’s the rule” is a waste now; AI has that cold. What the machine can’t do is sit on a live matter and say out loud, with a junior listening, here’s why I’m going to ignore the rule this time, for this client, this week. That sentence is the entire job. Spend the expensive hours there.

And the third, the one almost nobody does, is to say what you’re training people to become. Out loud, in plain words. You’re building judgment and the nerve to own a call. Then go make your reviews and your raises line up with that, which is the uncomfortable part, because it stays a lot easier to reward the associate who can recite the standard cold.

So here’s where I land. The machine didn’t beat the professors at being lawyers. It beat them at the one slice of being a lawyer we already knew how to automate. The firms still standing in five years will be the ones that quit paying people to be good at that slice and start spending real money, and real partner time, on the part that was never on the test. I keep having this conversation, and the leaders who get it have stopped asking how to make their associates faster. They’ve started asking what their associates are supposed to become.

If you enjoyed this, please share it with three friends. If you want to chat further on it, reach out to me at steve@intelligencebyintent.com. If there’s a topic you’d like me to cover, let me know.

Share Intelligence by Intent

How to Choose Where to Start With AI When It Can Do Almost Anything

TL;DR: The hard part isn’t the model. It’s deciding what to point it at first. Don’t start with the most impressive thing it can do. Start with the task you repeat constantly and can check in under a minute. Score your options on three things: how often you do it, how much it drains you, and how fast you can verify the result. Keep the high-stakes work for later. And when one sticks, turn it into a repeatable step so the whole team gets the gain, not just you.

A client asked me last month what the model is best at. Honest answer? I don’t fully know, and neither does anyone else.

Here’s the problem that creates. You open the box, the cursor blinks, and you can ask it to do almost anything. Write the client memo. Summarize the deposition. Build the spreadsheet. Plan the offsite. So you do one of two things. You freeze, because where do you even begin. Or you reach for the hardest, most important thing on your desk, hand it over, and watch it confidently invent a case that doesn’t exist.

Both reactions kill momentum. I’ve watched sharp partners try the model once on a brief, catch one bad citation, and write off the whole thing as hype. They started in exactly the wrong place.

Here’s the filter I actually use with firms.

Stop asking what’s most impressive

The instinct is to test AI on your most valuable work. It feels efficient. If it can draft the M&A memo, that’s where the money is.

But that’s backwards. Your most valuable work is usually your highest-stakes, hardest-to-check work, which makes it the last place you’d want to be testing whether you can trust the model yet. Your first mistakes have to land somewhere. Better they land where nobody’s watching.

The better question is quieter. What do you do over and over that you could check in twenty seconds?

Think about it this way. You want reps, low stakes, and a fast feedback loop. That points you at the boring stuff. The work that doesn’t feel impressive at all.

What makes a task a good first pick

When I sit down with a firm to pick a first use, I’m really weighing three things, and they don’t matter equally.

Start with how often the work comes up. You want something constant, a dozen times a week or more, because the saved minutes compound and you get fluent fast. Five minutes saved once is a rounding error. Five minutes saved forty times a week is most of a workday handed back.

Then there’s the drag. The work people avoid, delay, or grind through with their jaw set. Cleaning up a rough transcript. Turning six pages of email into something a partner will actually read. It’s rarely the strategic work, it’s the irritating work, and that’s the point. If a task doesn’t bug you today, handing it off won’t feel like anything.

The one that matters most, though, especially when your name is on the result, is whether you can check the answer fast. This is the test people skip and regret. If verifying the output takes as long as doing the work yourself, you’ve gained nothing. The tasks worth starting with are the ones where a wrong answer jumps out at you. You read the summary, you were on the call, you know in three seconds if it’s off. A brief full of citations is the opposite. There, checking the work is the work.

And keep an eye on what a mistake costs. Early on, you want work where being wrong is cheap. An internal summary that’s a little off costs you nothing and a minute to fix. A fabricated quote in a client deliverable costs you the client, maybe worse.

What good looks like in practice

A few tasks that fit.

Turning a messy call transcript into clean notes and action items. You were there. You can spot a wrong takeaway instantly, and you do this constantly.

Drafting the first version of a routine client email, the kind you’ve written a hundred times. You’re not shipping the model’s words. You’re skipping the blank page. Thirty seconds to read, easy to fix.

Pulling the three things that matter out of a long document so you can decide whether the whole thing is worth your time. Wrong? You’ll know the moment you open it. Right? You just saved twenty minutes.

All three look different and share the same bones. The raw material is already in front of you, the transcript, the email thread, the document. You can eyeball the output in seconds. And if it’s wrong, nobody gets hurt.

It shows up well outside the practice of law, too. A deal team buried in call notes. An executive trying to turn a dozen scattered updates into one Monday brief. Different work, same fit.

Now the one that fails. The appellate brief with case citations. High stakes, slow to verify, and the exact spot where these models still make things up. Get there eventually. Don’t start there.

How you’ll know it worked

Measuring this is simpler than people expect. Did it actually save time you can count? Rough math is fine. A task that took fifteen minutes and now takes four, including your review, is eleven minutes back. Write it down.

Quality has to hold, too. The output should be at least as good as what you shipped before, after your edit. If you’re spending more time fixing than you saved, the task isn’t ready, or the way you’re asking isn’t.

But the real test is duller than any metric. Are you still using it in week three? Most AI experiments die quietly. People try it twice, forget, slide back to the old way. If a use survives three weeks of your actual calendar, it’s real. If it doesn’t, drop it without a second thought and pick the next one.

Then make the win repeatable

Here’s the part most people skip. When a use sticks, don’t leave it in your head.

Turn it into something other people can run. A saved prompt. A one-page template. A line in the intake process that says draft the summary first, then review. That’s the difference between you saving eleven minutes and forty people saving eleven minutes. One person getting faster is a good habit. A repeatable step is how a firm actually changes.

What to do Monday morning

1. Write down every task you repeated last week. Circle the three you could check in under a minute.

2. Pick one. Run it through the model every time it comes up this week, and jot the minutes saved each time.

3. Friday, decide: keep it, change how you ask, or drop it and try the next one on the list.

That’s the whole thing. You’re not hunting for one perfect use. You’re running cheap experiments until a few of them stick.

The firms pulling ahead didn’t go hunting for the impressive use case. They started on the boring, repetitive work, racked up small wins, and kept going. Everyone else is still staring at the blinking cursor, waiting to figure out the perfect place to begin.

If you enjoyed this, please share it with three friends. If you want to chat further on it, reach out to me at steve@intelligencebyintent.com. If there’s a topic you’d like me to cover, let me know.

Share Intelligence by Intent

AI and Jobs: The Honest Answer Is Tasks, Not Jobs

TL;DR: There’s a lot of conversation these days about AI and jobs. I see a lot of the research, and I wanted to share where I’ve come down on it. One camp says AI will erase most jobs. The other says it will create more than it destroys. From where I sit, working with firms every week, both miss what’s actually happening. AI is very good at pieces of a job and bad at the whole job. For small and mid-sized firms, the threat isn’t mass layoffs. It’s slow margin erosion, a broken training path for young lawyers, and clients who expect more for less. Here’s what to change, and what to tell your kids.

Every managing partner I talk to has read the same two headlines. One says AI is about to wipe out half the jobs in the building. The other says relax, technology always makes more work than it kills. Both ran this week. Both sounded certain. And you’re supposed to make payroll decisions based on that.

I’ve been making the same argument in every keynote this year, so I’ll skip the windup. They’re both half right. The catch is that the half each side gets wrong is the half that actually matters when you run a firm.

The two forecasts everyone’s arguing about

The doomer story goes like this. AI gets better fast, it does what knowledge workers do, and soon enough we’re all on universal basic income arguing about how to spend our afternoons. The accelerationist story is sunnier. Yes, some jobs go away, but new ones show up, often jobs we can’t picture yet, and it happens faster every cycle. Net positive. Don’t worry about it.

The research that lands in my inbox is calmer than either camp wants it to be. Yale’s Budget Lab. Microsoft Research. A big study out of Denmark that tracked actual payroll records. None of them shows a broad jobs apocalypse. Unemployment hasn’t broken. One economist at Apollo, Torsten Slok, looked at weekly payroll data this spring and said flatly that there’s zero evidence of AI-driven job losses so far. If anything, he argued, the build-out is creating demand: people to install the tools, data centers to run them, chips and power to feed them. Jevons paradox in real time. Make a thing cheaper and people use more of it, not less.

So the doomers are overclaiming. We have a lot of AI adoption and very little evidence, so far, of broad AI-driven unemployment.

But the accelerationists are too comfortable, because underneath the calm averages something specific is already moving. Stanford is the cautionary exception. Its work on ADP payroll data found a 6 percent drop in employment for workers aged 22 to 25 in the most AI-exposed jobs, even while overall employment kept growing. The pressure is landing first on the youngest workers, and that matters a great deal if you sign the checks.

What I actually see when I sit with firms

When I work with a firm, I don’t watch AI replace lawyers. I watch it eat tasks.

It drafts the first version of a demand letter in ninety seconds. It turns a messy pile of documents into a clean chronology. It summarizes a deposition and pulls the key admissions. It cleans up billing narratives that used to cost a paralegal an hour. Real, useful work. The kind of work that used to fill a young associate’s week.

What it doesn’t do is the job. It doesn’t sit across from a frightened client in a custody fight and know when to push and when to wait. It doesn’t read a judge. It doesn’t decide whether this is the motion worth filing. It doesn’t carry the weight when the call goes wrong.

Most jobs are a bundle of tasks. AI is compressing some of those tasks toward zero and leaving the rest alone. Think about it this way. The job is a bag of forty things, and AI just got very good at twelve of them. The bag doesn’t vanish. But the shape changes. And that changes who you hire and how you bill.

That’s the in-between answer. Not robots take everything. And not relax, you’re fine. The work is getting rearranged, and the firms that come out ahead will be the ones doing it on purpose. Most are doing it by accident.

The squeeze nobody wants to talk about

This next part is the one that worries me.

The tasks AI is best at are the exact tasks we used to hand first- and second-year associates. Research, summaries, first drafts, document review. The grunt work. And that grunt work was never just grunt work. It was how young lawyers (or consultants, or investment bankers) learned to think.

The numbers are already moving. In 2025, only 38 percent of associate hires at U.S. firms came straight out of law school, down from 46 percent in each of the two prior years. For the first time in years, lateral hiring outpaced hiring directly from law school, bringing in people who already have judgment and can step into a live matter without a two-year runway. Read that again. Firms are quietly deciding they need fewer people learning the trade and more people who already know it.

The short-term logic is obvious. Why pay a first-year to spend twelve hours on a research memo when the tool does the first pass in twenty minutes and a senior associate cleans it up?

The trap shows up later. Cut junior hiring and training, and you save money this year while starving your partnership pipeline for the next decade. The expertise you bill at premium rates came from somewhere. It came from people doing the lower-level work, making mistakes, getting marked up, slowly building the instinct you now sell. Pull out that bottom rung and you don’t feel the damage for years. Then you feel all of it at once.

The easy takeaway is “we don’t need juniors anymore.” I think that’s a mistake, and the harder truth is that we need them learning faster, with better supervision, using AI as a study partner and not a shortcut around the thinking.

What small and mid-sized firms should actually do

Let me be direct, because this is the part clients pay me for.

You are not Kirkland & Ellis. Kirkland just said it’s spending $500 million over the next three to four years to build its own AI platform, starting with $100 million this year. That’s an arms race for firms with ten-billion-dollar revenue. It is not your model, and you don’t need it to compete. You need three things. A safe way to use good tools. A short list of workflows where AI actually pays off. And a pricing model that stops handing the savings to your clients for free.

Set the guardrails before you do anything clever

Decide which tools are approved and which are banned. Decide whether confidential client information can go into a tool, and which vendors are allowed to train on your data. (The answer to that last one is almost always no.) Put it in writing. The ABA’s first ethics opinion on generative AI, Formal Opinion 512, is blunt about what’s at stake: competence, confidentiality, client communication, supervision, candor to the court, and reasonable fees. Those duties don’t soften because the draft came from a machine.

And verification is not optional. AI hallucinations in court filings aren’t edge cases anymore. The exact counts vary depending on who’s counting, but the direction is clear, and it’s the wrong way. Even Sullivan & Cromwell, one of the most prestigious firms on the planet, apologized to a federal judge this spring after a filing went out with fabricated, AI-generated citations. If it can happen to them, it can happen to you. Make the rule one sentence and don’t let it bend. No lawyer cites, files, or sends AI-generated legal authority unless a human has checked it in a real research source.

Pick five workflows, not fifty

The biggest waste I see is letting every lawyer experiment forever. Endless dabbling, no traction. Pick a handful of workflows that are high-volume, repeatable, painful today, and easy to check. For most firms that means intake summaries and early case assessment, fact chronologies and timelines, first-pass deposition and discovery summaries, drafting support for routine letters and memos, and cleanup of billing narratives and client updates. Five things. Get good at those, write down how you do them, and stop there for now.

Notice what those five have in common. Each one produces a draft or a summary that a person then verifies. That’s the safe zone. The tool suggests, the lawyer decides.

Fix your pricing before AI eats your revenue

This is where firms get hurt, and it’s the part most of them haven’t touched.

If a task used to take six hours and now takes two, and you bill by the hour, you just cut your own revenue for the same client outcome. Clio’s small-firm research this year found something I keep repeating to clients. Roughly seven in ten small and solo firms now use AI, but fewer than a third have actually grown revenue with it. The bigger firms are pulling ahead on that number. Same tools, very different results. The winners changed how they charge. Everyone else kept billing hours on work that no longer takes hours.

So move the repeatable, AI-compressible work onto fixed fees, phased fees, or a monthly retainer. Keep hourly billing for the genuinely uncertain, bespoke, high-volatility matters where it still fits. Here’s the simple math. Say a severance review used to run four hours at $450, so $1,800 to the client. With AI, your lawyer spends ninety minutes on it. Price it as a flat $1,250 or $1,500. The client gets a lower number and certainty. You keep the margin the efficiency created instead of giving it back. That’s how AI turns into profit rather than a quiet leak.

Rebuild the apprenticeship on purpose

You’ve made AI part of the work. Now make it part of the training, or you get the productivity and lose the next generation.

The method I’d teach every junior is simple. Make them think first, before they touch the tool. What’s the issue, what facts matter, what law might apply, what’s the likely answer, what would change their mind. Then let them use AI to test that thinking, find the gaps, build the chronology, draft the outline, argue the other side. Then make them verify everything and write up where the tool was wrong or thin. That last step is the gold. Reward associates for catching the machine’s mistakes, because catching mistakes is the thing that becomes judgment.

The other move costs almost nothing and pays off for years. Get your best partners to record how they think. Capture the reasoning, not just the templates. How they decide whether a TRO is worth filing, how they read whether the other side wants to settle, what makes them trust a construction delay claim. Ten-minute explanations, captured and shared with the people coming up. A few firms are already building “AI twins” of senior partners for exactly this. You don’t need anything that elaborate. You need your partners’ thinking written down somewhere a young lawyer can find it.

The objections worth taking seriously

Maybe I’m underrating the pace. Anthropic’s Dario Amodei, for one, thinks the hit to entry-level white-collar work is closer and bigger than the calm averages suggest, and he could be right. A snapshot misses inflection points. The accelerationist case has real weight too. Cheaper production usually creates new demand, and a small firm that can suddenly offer a fixed-fee service it could never afford to staff before might do more work, not less. I’ve watched that happen. The cost of doing a thing drops, and appetite for the thing goes up.

There’s a quieter risk on the other end. If everyone moves to flat fees and the machine does the heavy lifting, do clients start asking why they’re paying lawyer rates at all? I don’t have a clean answer. My best guess is that judgment, trust, and accountability stay valuable precisely because a model can’t carry them. But I hold that loosely. Anyone who tells you they know exactly how this plays out is selling something.

A note to parents with kids in high school and college

I get asked this more than anything else. Usually after the keynote, usually quietly, by a parent who isn’t worried about their firm at all. They’re worried about their kid.

What do we tell them? The honest answer is that we can’t hand them a safe career the way our parents tried to hand us one. The titles are going to shift under their feet. But that’s not a reason to panic, and it’s not a reason to steer them away from learning hard things.

What I usually say is this. Don’t have your kid bet everything on a single narrow skill a tool might swallow. Have them get genuinely good at thinking, writing, and arguing, because those carry everywhere, and they’re exactly what’s left once AI handles the first draft. Have them learn to work with these tools now, fluently, the way our generation learned to work a search engine. The kid who can direct AI, check it, and catch what it gets wrong will be worth far more than the kid who either refuses to touch it or trusts it blindly.

And don’t let the tool do their thinking while their thinking is still forming. I feel strongly about this one. There’s early evidence that leaning on AI too hard dulls the very skills you’re trying to build. A college student who has AI write every paper isn’t saving time. They’re skipping the reps that build the muscle. Let them wrestle with the hard problem first. Then let them use the tool. It’s the same three passes I’d give a young associate. Think, then assist, then verify.

There’s one more thing I tell parents, and it has nothing to do with desks or law degrees. Point your kid toward work where being human is the actual point. A veterinarian with her hands on a scared animal. A therapist sitting with someone who’s coming apart. The trades, too: electricians, plumbers, people who build and fix the physical world. High-touch sales, where the deal closes because a client trusts a person and not a pitch deck. The machine isn’t coming for those any time soon, because the human connection isn’t decoration on the job. It is the job. That’s not a fallback. For a lot of kids it’s a better life than the one I used to picture for them.

The deeper thing I’d want my own kids to hear is this. The future rewards people who can do what a machine can’t hand off. Judgment. Taste. Knowing which problem is worth solving, and being able to read another person well enough to understand what they actually need. None of that is on a syllabus, and none of it is going obsolete. Help your kids build those, stay curious, and learn how to learn fast, and they will be fine in a world none of us can map yet.

I mean that. It’s the most useful thing I know.

What to do Monday morning

If you run a firm, start here. Not a year-long study. Three moves this week.

1. Write a one-page AI rule. Approved tools, no confidential client data in unapproved tools, and a hard verification standard for any legal authority. Put it in front of every lawyer and staff member by Friday.

2. Pick one workflow to pilot. Intake summaries or chronologies are the easiest wins. Measure the old time, the new time, and the rework, and let the numbers tell you whether it’s working.

3. Pull your three most repeatable matter types and ask one question of each. Should this still be billed by the hour? If the honest answer is no, draft a flat fee and test it on the next client.

The firms that struggle won’t lose their lawyers to AI. They’ll lose their margin, their speed, and their best young people to the firms that figured this out first. I’d rather you be the firm they’re losing to.

If you enjoyed this, please share it with three friends. If you want to chat further on it, reach out to me at steve@intelligencebyintent.com. If there’s a topic you’d like me to cover, let me know.

Share Intelligence by Intent

Claude Opus 4.8 Just Shipped. Here’s Why It Matters for Your Firm.

TL;DR: Anthropic released Claude Opus 4.8 on May 28. Same price as the last version. The gains that matter for lawyers aren’t the coding headlines. They’re three quieter things: the model can now grind through long, document-heavy work for hours without losing the thread, it makes things up less often and flags its own uncertainty, and it posted a real jump on a benchmark that measures actual professional work. Harvey, the legal AI platform, reported it hit the highest score they’ve ever recorded on their attorney-work benchmark. That last one is the story.

I’ve been watching these releases come fast. Opus 4.7 landed in April. Six weeks later, here’s 4.8.

That pace tells you something. But it’s not what you should care about. You don’t need to know there’s a new model. You need to know whether it changes anything for the work your firm actually does. So let me skip the benchmark parade and get to the three things that should land on your desk.

And I’m not writing this from the launch notes. It dropped today, and I’ve already been running it hard, building demos and example content for a CLE I’m delivering to the Beverly Hills Bar Association (BHBA) Tax group next week. So this isn’t a press-release read. It’s a first-day field report from real use. More on what that felt like in a minute.

What actually changed

Three things, in plain terms.

It can work longer without falling apart. Anthropic built this version to hold a plan across stages, track what it’s done and what’s left, and adjust when something breaks instead of stopping cold. Old models drafted you a paragraph. This one can actually work a matter.

It lies to you less. Anthropic’s own word for this is “honesty,” and they put a number on it. The model is roughly four times less likely than the prior version to let a flaw in its own work slip by without flagging it. Early testers said it’s quicker to say “I’m not sure about this part” instead of bluffing.

And it does professional work better. Not chat. Not trivia. The kind of valued, billable work your people do all day.

The honesty thing is the one that should get your attention

Here’s what I mean. The thing that has kept AI off the critical path at most firms isn’t capability. It’s trust. You can’t hand a model 4,000 documents and a deadline if it might confidently tell you it reviewed all of them when it skimmed half. That’s not a productivity tool. That’s a malpractice exposure with a friendly interface.

So the number worth circling is this one: four times less likely to let its own mistakes pass unremarked. And testers reported the model now tends to raise its hand and say “this input looks off” or “I’m not confident here,” which is exactly the behavior you want from a junior associate and exactly the behavior these tools have lacked.

And that’s the part I’d underline for a law firm. A model that hedges when it should hedge beats a slightly smarter one that bluffs. Every time. Because confidence without accuracy isn’t a small flaw in a research tool, it’s the whole problem. This release didn’t solve it. It moved in the right direction, which is more than I can say for most of the gains people get excited about, and for your world that’s the headline.

I’ll be honest about the limits, though. Anthropic itself called this “a modest but tangible improvement,” not a leap. The honesty gains are measured on the model’s own coding work, and how cleanly that carries over to reviewing a deposition transcript is something you’d want to test on your own matters before you trust it. Believe the direction. Verify the magnitude.

The long-running work is what makes it useful on a real matter

A model that can only handle one question at a time is a search box. A model that can plan a multi-step job, run it for hours, check its own work, and come back with something usable is a different category of thing.

Anthropic is leaning hard into this. The new version is built to carry context across long sessions and manage multi-day projects end to end. On their own long-running tests, testers said the work came back faster and the output was denser with the useful stuff and lighter on the filler. The detail that stuck with me: the model kept catching problems in the inputs and raising them, the kind of thing other models just leave sitting there for some human to trip over later.

For your firm, picture the document-heavy parts of a matter. Reviewing a production set. Pulling every reference to a single issue across thousands of pages. Building a first-draft chronology from a pile of records. These are exactly the jobs that used to break when the model lost the thread halfway through. This release is aimed straight at that failure mode.

I’ll give you a concrete one, because I ran it myself today. I had a client matter that needed reasoning across more than 3,000 documents to narrow them down to the roughly 600 that carried the most value for the analysis we were running. That’s not a search. That’s judgment applied at volume, the kind of first-pass triage you’d normally hand to an associate and a long afternoon. It crushed it. Smart, fast, and accurate. And here’s the part that matters most: when it started to drift partway through, it caught itself and corrected, rather than confidently barreling down the wrong path and handing me a clean-looking but wrong answer. The output was excellent. I ran it in Claude Code, but I could just as easily have done it in Claude Cowork, which is the version more of your people would actually touch.

Here’s where today comes in. Building the BHBA tax material, I threw genuinely messy prep at it: dense source content, multi-part drafting, demos I needed to look right in front of a room of tax attorneys who will notice if something’s off. Two things stood out. It’s fast. Noticeably faster than what I was using a month ago, to the point where the iteration loop stopped being a bottleneck. And the output held up. Not “good enough for a rough draft.” Good enough that the editing I did was shaping and judgment, not cleanup. For content going in front of a sophisticated bar audience, that’s the bar that matters, and this cleared it.

The professional-work number is the part I keep coming back to

There’s a benchmark called GDPval. It measures how well a model does economically valuable work across professional fields, including legal and finance. The version everyone’s citing today is run by an independent group, Artificial Analysis, not by Anthropic, which is part of why I trust it more than the usual self-reported scores.

Opus 4.8 scored 1890 on it. The prior version scored 1753. The leading competitor, GPT-5.5, scored 1769. So this version didn’t just improve on itself by a wide margin. It pulled clearly ahead of the strongest rival on the one benchmark built to mirror real professional output.

That’s a big move. Bigger than the coding gains everyone is writing about today. Because GDPval is the closest thing we have to a measure of “can this thing do the work I’d actually pay a person to do,” and the answer moved up sharply.

And there’s a legal-specific data point that’s even more direct. Harvey, the legal AI platform, reported on the launch that Opus 4.8 hit the highest score it’s ever recorded on its Legal Agent Benchmark, and that 4.8 was the first model to clear a bar they’d been measuring against. Their framing was that this kind of accuracy gain translates directly into how much real attorney work their clients can hand off with confidence. That’s a vendor talking its book, sure. But it’s a specific, falsifiable claim about legal work, not marketing fog.

The objections, stated fairly

This isn’t all upside, and you should hear the other side before you get excited.

It’s an incremental release, not a revolution. Day to day, you may not feel a dramatic difference on any single task. The gains show up over volume and over long jobs, not in a single quick question.

The honesty improvements are measured mostly on the model’s own coding work. Whether that same self-awareness holds up when it’s reviewing your discovery set is an open question you should test, not assume.

And none of this changes the governance work you already have to do. Which version you use, on what platform, under which terms, with what data protections, is still the question that determines whether any of this is safe to put near client material. A better model on the wrong terms is still the wrong tool. That part hasn’t gotten easier.

What to do Monday morning

Three moves, no more.

First, pick one document-heavy, low-stakes task this week and run it on the new version against your current process. A first-draft chronology, an issue pull across a record set, something where you can check the work. Measure whether it actually flags its own uncertainty the way the launch claims.

Second, confirm which version your firm is actually getting and on what terms. The model name changed. Your platform, your data protections, and your enterprise-versus-consumer terms decide whether this matters or whether you’re just reading about it.

Third, tell your skeptics the real reason to look again. It isn’t that the AI got smarter. It got more honest about what it doesn’t know, and for a profession that runs on accuracy and exposure, that’s the change that earns a second look.

The models keep getting better. The firms that win won’t be the ones with the newest model. They’ll be the ones who figured out where to trust it, where not to, and built the habit of checking. This release just made that line easier to draw.

The models will keep getting better, and they'll keep shipping faster than you can read the launch notes. But the firms that win won't be the ones running the newest version. They'll be the ones who figured out where to trust it, where not to, and built the habit of checking every time. Opus 4.8 didn't make that line disappear. It just made it easier to draw, because a model that raises its hand when it's unsure is finally behaving like the junior associate you'd actually keep. If you want to talk through where that line sits for your firm, or how to test it on your own matters before you trust it, reach me at steve@intelligencebyintent.com. The smartest tool in the room was never the goal. The honest one is.

Share Intelligence by Intent

AI Doesn’t Have a Bar Number

TL;DR: The bar moved while everyone was arguing about whether to use AI. The real exposure now sits in the gap between using AI and using it well, and the attorneys closing that gap right now are setting the standard the rest of the profession gets measured against. I’m running a workshop for family law attorneys next weekend on what closing that gap looks like, and the four rules I’m giving them are the same four rules every practice group needs.

Where this actually leaves us

I’m running a workshop for family law attorneys next weekend. The slide I keep rewriting is the one that explains why any of us will be in the room at all.

Here’s where we actually are. Mata v. Avianca was not the last sanction. Fabricated citations keep showing up in pleadings. The discipline is public, the names are searchable, and the malpractice carriers are paying attention. And avoidance isn’t safe anymore. Opposing counsel are using these tools, clients increasingly assume you are too, and refusing to engage doesn’t insulate you from malpractice exposure. Using AI badly is what gets attorneys sanctioned. Not using it at all is what gets firms left behind. The bar associations are moving with the technology, not against it. Technology competence is part of the duty of competence in more states every year. The standard is no longer whether you use AI. It’s whether you use it well.

That’s what the workshop is about. Four rules.

Privacy first.

Commercial accounts only. If the material is privileged, it stays on the commercial side of the line: the side that doesn’t train on your inputs and gives you retention controls. Consumer accounts (free or personal ChatGPT, personal Claude, consumer Gemini) train on what you paste by default. They keep it for years. No BAA. Not HIPAA-eligible. If you wouldn’t email a deposition exhibit to a Gmail account you share with your kid’s soccer coach, don’t paste it into one of these tools.

Match the tool to the actual work.

Most of what attorneys do with AI still happens in a chat window. Quick question, paste a paragraph, get a draft. That’s fine for one-off tasks. But the bigger leverage is happening elsewhere. Tools like Cowork can work across entire folders: renaming files, converting formats, applying Bates numbers, building indexes. A chat window can’t do that. If you’ve never actually watched one of these tools tear through a real folder of real files, you’re judging the technology by the wrong interface.

Work in Projects, not one-off chats.

Treat a Project like a matter file with a paralegal who has already read every page. Upload the pleadings, the financial production, the deposition transcripts, the prenup. Set the standing instructions once. Every conversation inside that Project draws from the same baseline. The next time the matter heats up, you don’t burn ninety minutes on a Sunday morning reconstructing the asset picture. The new associate doesn’t need to be re-briefed on the family tree. The Project remembers. The team builds on top of it.

Prompt with intent.

Context, objective, role, examples. Ten seconds of structure on the front end saves an hour of cleanup on the back end. Treat the AI like a smart associate who will do good work if you brief it the way you’d brief any smart associate. Vague in, vague out.

What this looks like in one matter

Here’s the example I’m betting will land hardest with the room.

New case lands. Big production. A hundred-plus financial documents come in as a mix of PDFs, scanned exhibits, photographed pages, and Word files with whatever names the producing party felt like using. Welcome to discovery. Standard practice is to hand it to an associate or a paralegal who spends the better part of a day renaming files, converting them to PDFs, putting Bates numbers on every page, and building an index in Word.

I pointed Cowork at the folder. One set of instructions: standardize the names into a common structure, convert everything to PDF, stamp Bates numbers in sequence, build the index. Then I walked away.

Ten minutes later the folder was done. Every file renamed, converted, Bates-stamped in order, and accompanied by a clean Word index.

Run the math. Four to six hours of associate or paralegal time, the kind you actually bill, compressed into the time it took me to refill my coffee. Multiply that across a year of matters and across a firm. The number gets serious.

And the work product is better, not worse. The naming is consistent. The Bates numbering doesn’t skip. The index doesn’t have a typo because nobody got tired on page seventy-three.

The rule that has no exceptions

AI has no malpractice insurance. No reputation to protect. No career built case by case. You do.

So the one rule I won’t bend on, and the one I’m going to make sure every attorney in the room hears twice, is this: human review happens after the final AI modification, not before. Not after the first draft. After the final draft. Every change the AI made, including the ones it made quietly in a paragraph you didn’t ask it to touch, gets a human read before anything goes out.

That rule sounds obvious until you watch how people actually use these tools. They review the first AI output, ask for a revision, glance at the revision, and ship it. The error that gets you sanctioned is almost always in a revision the attorney barely read. Because they’d already signed off on the version before it.

Final read on the final version. No exceptions.

What to do Monday morning

If you run a practice group or a firm, three moves matter more than the rest right now.

1. Audit your accounts. Right now. Find out which of your attorneys are still dropping client material into consumer-tier accounts. Move them to commercial-tier today, even if you haven’t finalized your firmwide policy. The privacy bleed is happening while you write the policy.

2. Pick one matter and turn it into a Project. Not a pilot program. One matter. One sub-practice lead, one matter, one Project. See what happens in a month. The conversation about firmwide adoption gets easier when you’ve already seen what good looks like inside your own walls.

3. Make the final-review rule a written standard. Human review after the final AI modification, not before. Put it in the policy. Put it in the training. Make it the one sentence every attorney can recite.

One last thing

These tools are as dumb right now as they’ll ever be for the rest of your career. The standard is being set right now by the attorneys learning to use them well. Everyone else will be measured against that standard whether they signed up for the test or not.

The attorneys figuring this out right now aren't the ones with the most tools. They're the ones who decided to learn what good looks like before the standard hardened around them. If you're working through any of this inside your firm, the privacy audit, the first Project, the final-review rule, reach me at steve@intelligencebyintent.com. The standard is being written this year, and you want your name on it.

Share Intelligence by Intent

AI Won’t Fix Your Revenue Engine. It Will Just Automate Your Dysfunction.

Why pointing autonomous agents at a broken go-to-market process is the most expensive mistake you can make this year.

TL;DR: Pointing AI agents at a broken go-to-market process doesn’t fix it. It speeds up whatever’s already wrong. Before you buy another AI tool, your executive team has to agree on what a Qualified Lead is, what a Booking is, and who owns the forecast when the model says one thing and the CRO says another. Skip that work and you’re paying six figures to scale your confusion at machine speed.

The boardroom scene you already know

Picture the moment. Your CRO has missed two quarters. CAC payback has crept out past twenty months from a historical fourteen. The board chair, halfway through your operating review, raises an eyebrow and asks the question every board is asking right now: “What’s our AI strategy?” She points to a competitor who’s been bragging publicly about an autonomous SDR that books meetings while their reps sleep.

You panic-buy. A six-figure AI forecasting platform. An autonomous outbound agent. A pipeline-intelligence layer. The line items make their way onto the next slide, the board nods, and you head into Q3 hoping the technology will do what the team couldn’t.

It won’t.

In October 2025, Gartner published a survey of 413 marketing technology leaders running AI agents in pilot or production. Forty-five percent reported that the vendor-supplied agents weren’t delivering the business results the vendor had promised. About half admitted their data and technical stacks weren’t ready to run the agents at all. Half also said they didn’t have the technical talent to operate them. Gartner’s separate prediction puts more than forty percent of agentic AI projects on track to be cancelled by the end of 2027.

The agents aren’t the problem. The companies buying them are.

Horsepower, not engineering

Here’s the line I keep coming back to:

AI is horsepower. RevOps is the chassis, steering, brakes, and instrumentation. More horsepower on a broken chassis doesn’t create speed. It creates instability.

A 700-horsepower engine bolted to a chassis with worn suspension and an unreliable brake system isn’t a faster car. It’s a more spectacular wreck. Same idea here.

Most companies are treating AI as a substitute for the operating work they avoided when revenue was easy. The work of agreeing on definitions. The work of designing the handoffs. The work of figuring out which number is right when Sales and Finance walk into the board meeting with two different ones. All of it was deferrable when growth was 35 percent. None of it is anymore.

You’ve got a capital reset behind you, the Rule of 40 acting as a screening filter at every funding conversation, and AI vendors selling speed into operations that were never built to handle it. The order of operations matters more than it ever has.

The new tech stack trap is faster than the old one

Buying tools instead of building processes is not a new failure pattern. Salesforce in 2010 didn’t magically create a sales process. It digitized whatever process you already had. If you had a discipline, the CRM made it visible. If you didn’t, the CRM enshrined the mess in pretty dashboards.

The AI version of this trap is worse. And it’s worse because of one word. Autonomy.

A poorly configured CRM sits there waiting to be corrected. It doesn’t do anything on its own. An ungoverned AI agent doesn’t sit. It drafts emails. It scores accounts. It books meetings. It updates pipeline. It applies discounts. It runs at machine speed against data nobody in the company actually agreed on. By the time anyone notices, the agent has done forty things, in three systems, that you now have to untangle.

In 2017, MD Anderson Cancer Center cancelled its contract with IBM’s Watson for Oncology after five years and $62 million, without ever deploying the system for patient care. Watson worked as a pattern-matching engine. What failed was the assumption that the institution could supply the clean, structured, integrated data it needed. The same assumption is failing inside B2B revenue organizations right now, on smaller invoices, every quarter.

You can’t afford one $62 million mistake. You can afford a sequence of $200K ones that compound.

Automating the Silo Tax

The friction between Marketing, Sales, Customer Success, and Finance has a specific cost. I call it the Silo Tax. IDC puts the floor at 10 percent of revenue annually for B2B companies whose go-to-market functions aren’t aligned around consistent processes and data. In mid-market practice I use 10 to 25 percent as a working diagnostic range. For a $100M ARR company, that’s $10M to $25M leaking every year through definitions nobody locked down.

Here’s what AI does to that number.

Marketing defines a Qualified Lead one way. Sales has a different definition, and treats Marketing’s leads as inventory rather than intent. Finance won’t accept either, because in their world Committed Pipeline requires a countersigned order. None of these are wrong inside their own world. The world is wrong.

You deploy the AI stack on top of all of it.

Marketing’s lead-scoring agent promotes an account that just hit its scoring threshold. Sales’ outbound agent immediately fires aggressive messaging at the same account, because the account appeared on its own queue with a different signal an hour later. Meanwhile, the AI forecast model produces a confident, two-decimal-place number based on Stage 3 CRM data that three reps invented over coffee on a Friday.

No agent is wrong inside its own rules. Each one is acting on an internally consistent definition. The four functions had four versions of the same word, and the agents made the consequences happen so fast that no human in the meeting could intervene before the account churned and the forecast was already in the board pack.

That’s the Silo Tax with the volume turned all the way up.

The Bottleneck Inversion

There’s a specific operational pattern AI creates that I want to name, because once you see it you can’t unsee it. I call it the Bottleneck Inversion.

For most of the history of B2B sales, the binding constraint on the deal cycle was drafting capacity. It took hours for a rep to build a custom proposal. It took a day for Legal to mark up a contract. It took a week for Deal Desk to price a non-standard configuration. The slowness of producing artifacts kept the whole motion contained. Reps and managers operated inside a natural friction that protected margin by default, even when the rules weren’t formally enforced.

AI removes that friction. Drafting capacity is now effectively free and instant. A rep with an AI assistant can produce a polished, customer-ready, fully-formatted proposal at any discount and any liability carve-out the prompt allows, in roughly the time it takes their coffee to cool.

Which means approval capacity is now the bottleneck. Not drafting. Approvals.

If you don’t have a published discount authority matrix, a Green/Yellow/Red contract classification, and response-time SLAs of 24, 48, and 72 hours that actually get measured, your reps are about to send out beautifully written, legally exposed, deeply discounted proposals before your CFO finishes their morning latte. The constraint flipped. The control surface you didn’t need (because drafting was slow) is now the control surface you can’t operate without (because drafting is instant). Pricing governance matters more in the AI era, not less, because the cost of producing an ungoverned proposal has fallen to roughly zero and the volume has gone up.

Same pattern shows up in forecasting, lead routing, and renewal motions. Whatever your historical bottleneck was, AI just relocated it to whichever governance function used to be invisible.

Treat AI like model risk, not magic

Here’s the mental shift the revenue leaders I trust have already made. Stop thinking about AI in your revenue stack as productivity software. Start thinking about it the way a Wall Street trading desk thinks about a new algorithm. Model risk.

Model risk is a familiar discipline. Credit functions have run on it for forty years. Risk officers know how to validate a model, backtest it, monitor drift, define when the output is suspect, and document override authority when the model fails. None of this is new. None of it is mysterious. RevOps adopts the same posture when an AI agent joins the revenue function.

That changes what a serious AI program looks like. The work isn’t vendor selection. The work is validation, backtesting, drift monitoring, and explicit failure modes: who owns retraining cadence, what conditions trigger a model review, what the override path is when the AI prediction contradicts the operators and the operators turn out to be right. The CFO and the RevOps leader own those answers jointly, or no one does.

Before any of that happens, you need what I’m calling the Agent Contract.

The Agent Contract is the data dictionary, written to be read by both humans and machines. For every revenue metric (Bookings, ARR, MQL, SQL, Opportunity Stage, Churn, NRR) the Contract specifies four things: the mathematical formula, the authoritative system of record, the human owner, and the review cadence. The C-suite signs it. Engineering wires it into the CRM metadata, the warehouse, and the governed prompt layer the agent actually reads. Every agent in the stack treats it as ground truth.

If an autonomous workflow cannot identify the formula, system of record, owner, and review cadence for a metric, it should not be allowed to touch that metric. That’s the rule.

What to do Monday morning

Block ninety minutes on Monday. Pull up your calendar, your CRM, and a blank page. Then work three things, in order.

1. Pick the three definitions your functions argue about most. Probably Bookings, MQL, and Churn. Lock the formula, the system of record, the human owner, and the review cadence for each one, in writing, before any agent goes near them.

2. Audit one pending AI pilot against this question: can it answer those four items for every metric it touches? If the answer is no, the pilot isn’t ready for production. It’s ready for design work.

3. Put a standing AI-deployment item on your weekly revenue meeting agenda. Make a signed Agent Contract the price of admission for any new workflow.

That’s it for week one. Definitions first. Governance second. Agents third. Reverse that order and you’ll spend the next four quarters cleaning up the consequences of moving fast on a foundation that wasn’t ready.

The vendors are right that the technology is real. They’re wrong that it’s a substitute for the operating system underneath. Build the engine AI deserves to be plugged into.

Then turn the key.

None of this is about being anti-AI. It's about respecting what the technology actually does, which is run, at speed, against whatever you give it. The companies that win the next four quarters will be the ones that did the unglamorous work first: definitions locked, governance signed, agents validated like models instead of trusted like magic. If you're staring at a stalled AI pilot, a frustrated board, or a CRO who can't quite explain last quarter's number, I'm happy to talk through it. Reach me at steve@intelligencebyintent.com. The engine is real. Build something worth plugging it into.

Share Intelligence by Intent

When ChatGPT and Claude Say “Memory,” They Mean Three Different Things

The first time I really understood this, I was sitting with a partner who had just been told by his IT director that ChatGPT memory was “off.” He showed me the toggle. He showed me the setting. He felt fine.

Then he asked ChatGPT to summarize a contract.

It opened by referencing his other client.

He didn’t paste anything. He didn’t upload anything. He didn’t even mention the other matter. Something was still on.

That’s the whole point. A setting labeled “memory” does not necessarily cover every place a product can pull in prior context. The toggle he flipped was one of several. The others kept doing their work.

That moment is what this post is about.

TL;DR

When ChatGPT or Claude talk about “memory,” they’re describing at least three different things stacked on top of each other. The privacy concern isn’t really what the model “knows.” It’s what the product silently retrieves from one matter and drops into another. For law firms, two of the sharpest practical risks today are non-Enterprise ChatGPT projects with default memory on, and Claude used as standalone chats instead of project-scoped workspaces. Fix those two things and you’ve cut most of the practical exposure.

Memory is not one thing

Here’s what’s happening under the hood, in plain English.

The first kind is the active thread. When you continue that same chat, the model can use the prior messages in it. Start a new chat, and that thread is not automatically included, unless a memory feature, a project, a chat search tool, or another retrieval surface pulls it back in.

The second kind is the context window. That’s how much the model can hold in its head at once. Numbers vary by plan and model, but think of it as working memory for one task.

The third kind is the one that causes problems. It’s persistent storage. Saved facts about you. Summaries of prior chats. Indexed documents in a project. Files in a library. Connected apps that, once authorized, can be referenced when the product decides the data is relevant.

That third layer is what most people mean when they say “memory,” and that’s where ChatGPT and Claude diverge architecturally.

ChatGPT’s design: convenience that crosses matters

OpenAI gives you two persistent stores. Saved memories are facts the system writes down about you. Reference chat history is looser reuse of helpful details from past chats. These are separate stores. Deleting a chat does not delete what the system already wrote down about you from that chat. To actually remove a fact, you have to delete the saved memory entry, delete the originating chat, and sometimes also remove files from your library and disconnect any app that fed in data.

That’s not one click. It’s a runbook.

Now layer on projects. In ChatGPT projects, you get a choice when you create one: default memory or project-only memory. The behavior of “default” depends on your plan. On Enterprise and Edu, projects stay contained inside the project boundary. On non-Enterprise accounts, including Plus, Pro, and Business, default-memory projects can reference chats from outside the project unless the other project is set to project-only.

Read that again. On a non-Enterprise account, a default-memory project can pull in conversations you had outside the project.

Think about what that means in a firm. Same user, multiple matters, default memory on, no project-only setting. The model can reference context from one client’s project while you’re drafting in another. There might be a small “remembering” indicator. Most users won’t read it. The answer just lands.

A side note on shared ChatGPT projects. Every member can see the project’s chats and files, so the visibility risk goes up. The memory risk actually goes down, because shared projects automatically switch to project-only memory. The exposure is real, but it’s a different shape of exposure than the cross-project memory bleed above. Don’t confuse the two.

Claude’s design: stricter walls by default

Anthropic made different choices.

In Claude, each project is its own workspace. Project knowledge is the main intentional shared context across chats. Claude also runs a separate per-project memory summary when memory is enabled. So the right way to think about it: the whole project is the matter boundary, not any single chat inside it.

The important point is what does not happen. Claude’s memory is separated by project. If you’re using projects matter by matter, Claude should not pull from another matter’s project into the current one. Within the same project, project memory and project knowledge can still shape later chats. That’s working as intended if you’ve kept that project to one matter.

Outside projects is where the bigger risk lives. Claude runs standalone memory too. It synthesizes summaries from your standalone chats, updates them every 24 hours, and uses those summaries to inform new standalone chats. If you do client work in standalone chats, those summaries can travel across matters. So the rule is simple: client work belongs in projects, not standalone chats.

Sharing also works differently in Claude. Sharing a project gives teammates access to the project’s knowledge base and instructions, but individual chats stay private unless you explicitly share them. That’s the opposite of ChatGPT’s default sharing model, where project chats are visible to all members.

One detail compliance officers should know. Disabling memory at the organization level in Claude Enterprise wipes everyone’s memory synthesis. The toggle has teeth. But Team plans don’t have org-level memory controls at all. If centralized governance matters to you, that’s an Enterprise question, not a Team question.

The risk most firms aren’t thinking about: memory poisoning

This one is newer and uglier. It applies to agentic tools, including Claude Code and managed agent setups, where the AI can write to its own memory store as part of doing work.

If an agent reads input from an untrusted source (a webpage it scrapes, a document someone emailed in, a tool output it doesn’t control) and that input contains instructions telling the agent to write something into memory, the agent might do it. The next session reads that poisoned note as if it were a real, trusted fact and acts on it.

Anthropic warns about this directly in its documentation for read-write memory stores. It’s not theoretical. Cisco’s research team published a proof of concept against Claude Code memory earlier this year, and Anthropic has been working through mitigations.

For a firm, the practical translation: any AI tool that can both write to memory and read untrusted content needs tighter scope. Read-only references for trusted material. Read-write stores only for tightly controlled matter-local data. No shared writable memory across matters.

A point about prompts

There’s a temptation to think this can be solved with instructions. “Tell the AI not to use other matters.” “Tell the agent not to write to memory.”

It can’t.

Anthropic says this explicitly about Claude Code memory files: they are context, not enforced configuration. The same principle should be applied across vendors. Prompts can guide behavior, but they don’t bind it. Storage controls, workspace boundaries, connector permissions, and admin settings are the real control surfaces. Treat prompts as helpful supplements, not as security walls.

What to do Monday morning

Three things, in order.

1. Move active client work off consumer accounts. Commercial and business tiers generally improve the privacy posture, including no training by default. Enterprise tiers are where firms get stronger retention controls, compliance and audit exports, SSO, SCIM, and the contractual terms a firm actually needs. The price difference is small compared to the exposure.

2. One matter, one workspace. In ChatGPT, that means project-only memory, one project per matter, and avoid shared projects unless the entire matter is genuinely team-shared. In Claude, that means projects rather than standalone chats, each matter in its own project, and only matter-approved documents in the knowledge base.

3. Write down your deletion runbook. When a matter closes, what gets deleted, in what order, by whom? In ChatGPT that’s chats, saved memories, library files, and connected apps. In Claude that’s chats, memory reset if needed, and project knowledge or the whole project. If you can’t answer this in writing today, that’s the first gap to close.

The line that matters

Memory is convenience. Memory is also exposure. The vendors aren’t hiding any of this, but they describe it across a dozen help pages and three different products, so it’s easy to miss.

The firms that get this right won’t be the ones with the longest AI policy. They’ll be the ones whose lawyers know, without thinking about it, which workspace this matter lives in.

Memory isn't a feature. It's surface area. The firms that handle this well won't be the ones with the thickest AI policy on the shelf. They'll be the ones whose lawyers know, without thinking, which workspace this matter lives in and what gets deleted when it closes. If you want a second set of eyes on what your firm is actually exposed to, I'm at steve@intelligencebyintent.com. The toggle was never the answer. The workspace is.

Share Intelligence by Intent

Notes From a Week I Couldn’t Keep Up With

TL;DR: This was the first week in a long time I felt behind on AI. Google I/O was the gravity well. While I was inside it, Andrej Karpathy joined Anthropic, the enterprise picture kept tilting Claude’s way, and the AGI-in-a-few-years drumbeat got louder from the people who would actually know. Below: what I saw, what I think it means, and the advice I’m giving the people I care about.

The Week That Got Away From Me

I’m going to write this one differently. No single theme, no clean argument, no tidy takeaway. Just notes from a week that moved faster than I could.

I was at Google I/O. Two days of demos, keynotes, hallway conversations, the works. And while I was inside that bubble, the rest of the industry kept going without me. I came home and looked at my queue and realized I wasn’t just behind on stories. I was behind on stories that actually mattered.

Here’s why I’m sharing this anyway. You feel it too. Every executive I work with has the same complaint right now. The pace is too high. The signal-to-noise ratio is getting worse, not better. And the cost of falling behind compounds every week.

So let me try to clear the table.

Karpathy to Anthropic Is the Signal of the Week

The top story, by some distance, was the report that Andrej Karpathy is joining Anthropic. If you don’t know the name: founding member of OpenAI, ran AI at Tesla, came back to OpenAI, then left to start his own thing. He is one of maybe ten people in the world whose move tells you where the puck is going.

He’s going to Anthropic.

That tells you something. It tells you the same thing the enterprise data has been telling me for six months. Every company I work with has made Claude their primary model. A handful keep ChatGPT or Gemini as a secondary for specific tasks. But Claude is winning the rooms that actually pay for software. Legal, financial services, consulting, life sciences. The serious buyers.

Anthropic also appears to be profitable, which shocked a lot of people who follow this closely. The frontier-lab assumption for the last three years was simple: nobody makes money on these models, you burn cash and hope the next round closes. Anthropic broke that pattern earlier than anyone expected.

And then there’s the compute story. Anthropic now has access to both Colossus I and Colossus II. I cannot overstate how big that is. The single biggest constraint on Anthropic’s growth has been getting enough chips to serve demand. That constraint just loosened by a lot.

Add it up. Profitable. Compute-rich. Hiring the best researchers. Owning the enterprise narrative. That’s what running away looks like.

Google I/O Was Amazing in Person and Mixed on Reflection

Let me be honest. I/O in person is incredible. The energy is real. The product surface area is huge. The leadership talks well. And I’m grateful I was there.

But I’ve spent the week reading what the industry is saying about what actually got announced, and the verdict is mixed at best. I covered the launches earlier in the week, so I won’t rehash them here. Two things stand out on reflection.

First, the product sprawl. Even as someone who pays close attention, I lost track of which Google product does what. Multiple overlapping consumer assistants. Multiple developer surfaces. Multiple agent stacks. A video model whose name I had to look up twice this week. If I can’t keep them straight, your in-house counsel sure can’t. Sprawl is not strategy. It’s a tax on every buyer trying to evaluate you.

Second, the pre-announcements. There were a lot of “coming soon” demos this year. Some are genuinely close. Some are months out. A few felt years away. I’m not ready to call it vaporware. But it reminded me of Microsoft in the early 2000s, when “announcing” became a substitute for “shipping.” If that pattern continues at Google, it will cost them credibility with the buyers they need most.

That said, I came out of two sessions feeling genuinely good. Matthew Berman’s interview with Sundar Pichai was sharp. Berman asked real questions, not the soft puffballs you usually get from a CEO at a developer conference. If you want one piece of source material from this week to actually watch, that’s the one.

The other was Demis Hassabis on stage. Two things he said are still rattling around in my head.

One: the point of AI is to improve human life. Cure disease. Solve problems we’ve been stuck on for decades. That’s the mission. Not chatbots. Not coding assistants. Those are stepping stones.

Two: AGI is a few years away.

I’ve heard that phrasing, or something close to it, from multiple frontier-lab leaders in the last sixty days. Demis on stage. Dario in different language. Sam in his own way. The people who actually run these labs have all moved their public estimates inside this decade. I take that seriously. And I increasingly believe Anthropic has cracked something on recursive self-improvement, which is the mechanism that takes us from “very capable assistant” to something different in kind.

You can argue the timeline. I’d push back hard on anyone who argues the direction.

A Long Conversation With Chubby

I spent almost 90 minutes this week with Chubby covering more ground than I can fit in one paragraph. Google’s strengths and weaknesses. Opus 4.7 with Adaptive thinking versus 4.6 with Extended thinking. US AI strategy against China. Anthropic’s economics and pricing. Codex against Claude Code against Google’s new AntiGravity. The shape of the developer-tools fight over the next twelve months.

I’m going to write that conversation up as its own piece next week, because it deserves real space. The short version is this. The developer-tools layer is where the next big platform shift happens, and the order of finish there is not yet settled. If you’re a firm leader making bets on which AI vendor your team standardizes on, watch the coding tools. They are the leading indicator for everything else.

What I’m Telling My Kids

I’ll allow myself one personal aside, because it kept coming up in family conversations all week and I think it matters for you too.

My oldest is graduating from the University of Oregon this spring. Go Ducks! My middle is now a senior in college. My youngest is finishing up his junior year of high school. Three kids, three different angles into the same question. What does AI mean for the work I was planning to do?

What I tell them, in plain language, is this. The jobs that look most defensible right now are the ones where you’re the person who decides what AI should do, who checks the output for things that would embarrass you, and who owns the relationship with the human on the other end. The jobs that look least defensible are the ones doing the production work in the middle. That’s the part getting eaten first.

So I tell my college kids: build judgment. Get reps making real decisions with real consequences. Internships matter more than ever, not less. And I tell my high schooler: learn to write clearly, learn how to argue, and learn how to use these tools fluently. The combination is rare and it’s going to stay rare.

I also tell all three of them to plan for continual learning as a permanent feature of work, not a phase. Whatever you learn in the next four years has a half-life that’s getting shorter. The people who do well are going to be the people who pick up something new every six months and don’t act like that’s a hardship.

And one more thing that gets undersold in most AI conversations. Look hard at careers that actually require a human across the table from another human, and look just as hard at the skilled trades. An electrician with twenty years of judgment is harder to replace than a paralegal with two. A plumber on a Saturday emergency call is irreplaceable. Therapy, hospice, high-end sales, complex negotiation, hands-on craft. These are not consolation prizes. In the world that’s coming they may turn out to be the most stable bets on the board.

This is the same conversation I think you should be having with your team. Different vocabulary, same substance.

One More Thing

I told you up front this was a stream of consciousness, and it is. I could write a full piece on any one of these threads, and I probably will. If a section pulled at you, hit reply and tell me which one. I’d rather go deep on what you actually want to read than guess.

The pace isn’t slowing down. None of us is going to keep up with all of it. The job is to stay close enough to the parts that matter.

That’s the work this year.

One more thing before I close the laptop. It’s Memorial Day weekend here in the US, and after a week like this one I’m putting the phone down, firing up the grill, and getting in the pool with my kids while they’re still around to get in it with. The news will be waiting on Tuesday. It always is. If you’re stateside, I hope you get a long slow weekend with the people who matter, a good burger, and at least one nap you didn’t plan for. And if you’re reading this from somewhere the calendar didn’t pause, I hope your regular weekend is a little less regular than usual.

Take care of yourselves. See you next week.

Google’s I/O 2026: From “Gemini Everywhere” to Agents That Actually Do the Work

TL;DR

A year ago, I told you Google had quietly become the company to beat in AI. After watching this week’s I/O, that argument feels almost too soft. Google didn’t show up with a faster model. They showed up making the case that the chatbot phase is over, and the next phase is agents doing your work for you, inside the apps you already use. Gemini 3.5 Flash. A 24/7 personal agent called Spark. Antigravity 2.0. A Search box rebuilt for the first time in a quarter century. Workspace going voice-first. Glasses this fall. And pricing finally pulled into line with Claude and ChatGPT at $100 and $200. From my seat, Google is turning its full stack into an agent runtime for consumers and the enterprise at the same time, and pricing it to win on volume. If you’re still planning your 2026 around chat interfaces, you’re planning for the wrong year.

The moment the keynote actually landed

I sat through most of the keynote waiting to feel it. New models, new benchmarks, new demos. We’ve all seen this movie. Then they brought Spark on stage, and the whole event clicked into focus.

Here’s what I mean. For most of 2024 and 2025, every AI keynote followed the same playbook. New model, benchmark chart, a demo where the thing writes a poem about somebody’s dog. Everybody clapped, closed their laptops, and went back to typing into chat windows like nothing had really changed.

This one was different. Google didn’t lead with intelligence. They led with action, which is honestly what most of us in this work have been waiting for. Spark working on its own VM with your laptop closed. Antigravity running parallel agents in the background. A Search box that takes a Chrome tab as input and books your dog groomer on the way out. Workspace where you just talk to your inbox.

The chatbot era didn’t end with a bang. Google decided to skip it.

What they’re really saying

Let me simplify this the way an operator needs it.

Gemini isn’t the product anymore. What Gemini does for you while you’re doing something else is the product. Every announcement on stage was a version of that one idea. The model is just the engine. The actual work, the part that matters for your firm, is everything you wire it into.

The model: frontier intelligence at a Flash price

Start here, because the economics drive everything downstream.

Gemini 3.5 Flash launched today, generally available everywhere. Google’s pitch is “frontier intelligence at Flash latency and Flash cost.” Their internal benchmarks have it beating Gemini 3.1 Pro (the flagship from three months ago) on most coding and agentic tasks, while running roughly four times faster than the comparable frontier models from other labs. Pricing is $1.50 per million input tokens and $9.00 per million output. Roughly half of Claude Sonnet. A fraction of GPT-5.5.

The number that should make you look up is GDPval-AA. It’s an independent benchmark maintained by the evaluation firm Artificial Analysis, built to measure “economically valuable knowledge work,” which is the closest thing the industry has to “AI doing real professional work.” Flash now sits at 1,656 Elo. Their last model was at 1,314. That’s a 342-point jump in three months on a benchmark Google doesn’t own. Which puts a Flash-tier model within throwing distance of the best models on the market at the price of a cheap one.

Think about what that does inside a firm. Cost stops being the thing that gates rollout. When the cheap fast model can hold its own on real work, you stop rationing AI to a handful of power users and start putting it into every workflow. For law firms specifically, that means associate-level research, intake summaries, document review, conflict checks, and a hundred other things that have been stuck in “we should pilot that” for two years become “let’s just turn it on this quarter.”

Pichai said the quiet part out loud during the keynote. If the biggest companies shifted 80% of their workloads to 3.5 Flash, they’d save over a billion dollars a year. That’s a Google scenario, not an audited number, but the direction is right. Google is forcing the market price of frontier-grade intelligence down. They can afford to because they own the chips.

Gemini 3.5 Pro is in internal testing and ships next month. Gemini Omni Flash, a new multimodal family that generates and edits video conversationally, started rolling out today across the Gemini app, Google Flow, and YouTube Shorts.

Spark: this is what I was waiting for

Of everything they announced, Spark is the one to actually study.

Spark is a personal AI agent that runs on its own Google Cloud VM, twenty-four hours a day, with native access to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps. Third-party integrations through MCP start with Canva, OpenTable, and Instacart. Beta opens next week for US Ultra subscribers. A Workspace version is “soon in preview” for business customers, whatever that ends up meaning in practice.

Here’s what makes it different. Spark doesn’t need your laptop open. You give it a task and it goes off and works on it, in the background, on Google’s hardware. It pings you for confirmation before doing anything high-stakes, like sending an email or paying for something or posting publicly. And it surfaces through a new Android status-bar element called Halo so you can see what your agent is doing without opening anything.

I’ve been using Claude as my primary tool for two years, and Cowork has been the most credible agentic experience I’ve had inside that stack. Spark is now the second product I’d put in that category. Both are real. Both are good. And the choice between them, for most firms, is going to come down to where you already live. If you’re a Workspace shop, Spark is going to feel like it grew out of your tenancy. If you’ve built around Anthropic or OpenAI, you have a better model quality argument and a more mature governance story, but you’ll have to wire up the integrations yourself.

The bigger point. Google isn’t asking you to pick a new chat app. They’re asking you to let an agent watch your inbox at night. That’s a much bigger ask, and it’s the right one for them to make if they want to win the next phase.

Antigravity is the developer side of the same idea

Antigravity 2.0 is Google’s agentic development platform. Desktop app, CLI, SDK, full integration with the Gemini Enterprise Agent Platform. You can run multiple agents in parallel, schedule background tasks, and pick up isolated sessions where you left off.

One thing worth flagging. Google is officially transitioning Gemini CLI users to Antigravity CLI. On June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for AI Pro, Ultra, and free individual users. Enterprise license holders on Gemini Code Assist Standard or Enterprise keep their access unchanged. If your firm has technical teams on individual or AI Pro seats running Gemini CLI in production, that migration is a calendar item. If you’re on the Enterprise license, you’re fine, but Antigravity is where Google is putting all the new development effort going forward.

For firms with real data residency rules, here’s the part that matters. Antigravity inherits Google Cloud’s privacy controls and runs inside your secure cloud boundary by default. That used to be a months-long conversation with IT and security. It just got a lot shorter.

Search and Workspace stop being apps and start being interfaces

Google rebuilt the Search box for the first time in a quarter century. The new one takes text, images, files, videos, and Chrome tabs as input. AI Mode runs on 3.5 Flash globally. They added information agents that watch the web for you 24/7, agentic booking for local services in the US this summer, and “Mini Apps,” which are custom dashboards Search generates on the fly using Antigravity.

The scale numbers matter here. AI Mode is now over a billion monthly users. AI Overviews is at 2.5 billion. The Gemini app is at 900 million MAU, up from 400 million a year ago. Across all surfaces, Google is processing 3.2 quadrillion tokens a month, seven times last year.

Workspace got the equally important update. Gmail Live lets you ask your inbox questions out loud. Docs Live turns a spoken brain dump into a structured draft, pulling from Drive and Chat with permission. Keep does the same for notes. AI Inbox just expanded from Ultra-only to Plus and Pro. Spark for Workspace is coming.

If you run a professional services firm, here’s the thing about voice-first work. The friction of writing prompts is what kept most partners and senior associates from actually using these tools day to day. Talking to your inbox is different. That’s a behavior change a 55-year-old equity partner will actually adopt, and the moment that happens, AI inside your firm stops being a thing the associates do and starts being a thing the firm does.

What sets Google apart now, and it’s not the model

A year ago, I argued the moat was vertical integration. Chips, cloud, models, distribution, all under one roof. That argument got stronger this week, but it also changed shape.

The new moat is that Google owns more of the surfaces where agents need to actually do things than anybody else on the planet. Search. Gmail. Docs. Calendar. Drive. Maps. Chrome. Android. YouTube. Workspace identity. Google Pay. And now glasses, with Samsung and Warby Parker and Gentle Monster, shipping this fall.

Agents don’t need apps. They need surfaces. An agent that can read your inbox, check your calendar, search the web, pay for the thing, and route the result to your phone is worth more than ten specialized agents that each do one of those badly. Google can hand all of that to its agents under a single identity layer. That’s a structural advantage nobody else can catch up to in eighteen months.

The competition has pieces of this, not the whole thing. OpenAI has the better consumer brand and a credible argument for the best frontier model. Anthropic, in my view, still has the better workhorse for serious reasoning and coding, and the most mature governance posture. Meta owns social. Microsoft has Windows and the enterprise install base.

None of them has the full surface area Google is now running agents on top of. From where I sit, that’s the real change since I wrote about Google’s stack in December.

The other thing that changed is pricing. Gemini Ultra now has a $100 tier and a $200 tier. So does Claude Max. So does ChatGPT Pro. Google had been the outlier at $250 with nothing below it, and they just stopped being the outlier. The strategic read is that Google wants the same seat-level conversation inside firms that Claude and ChatGPT have been having all year, and pricing won’t be the thing that makes a CFO blink anymore.

Risks I’d actually pay attention to

I’d be lying if I said this is all upside.

Quality drift matters more once the model is acting. A wrong answer in a chat is embarrassing. A wrong action in your inbox is something somebody has to clean up, and depending on what got sent or paid or posted, the cleanup may not be free. Google’s Spark gates high-stakes actions behind explicit confirmation, which is the right default. It’s also the kind of friction users start clicking through after the third day.

Data governance gets harder, not easier. When an agent is moving data across Gmail, Calendar, Drive, and third-party apps under one identity, your DLP and access controls have to catch up. Most firms’ AI policies are written for humans clicking buttons, not for agents executing on behalf of those humans. If you run a law firm, this is the conversation to have with your CIO and your general counsel this month, not in Q4.

Lock-in is real. The same thing that makes Spark feel effortless inside Workspace is what will make it painful to leave. Go all-in on Spark for personal productivity and Antigravity for enterprise agents and you’re committing to Google’s identity and data layer in a deeper way than you ever did with Search.

And the gated rollouts are still gated. Spark is US-only, Ultra-only, beta, English. Workspace Spark has no date. Glasses ship this fall with no price. The story is strong. The Monday-morning availability is uneven.

What to do this week

1. Add Gemini 3.5 Flash to your model evaluation portfolio. Run it against whatever you’re using today, on your actual work, not on a benchmark. The cost story only matters if the quality holds for your tasks.

2. Re-read your AI governance policy with one question: does it cover agents that act on a user’s behalf, not just users who chat with a model? If not, that’s the next revision.

3. If you sell to consumers, audit how your business shows up in the new Search box and the Universal Cart. If your product can’t be added to a Google-managed cart from a Gemini conversation, your distribution math just got worse.

The bottom line

Google didn’t win 2025 by shipping a better demo. They won by shipping defaults.

In 2026, they’re trying to pull the same trick at a higher level. The default for asking a question is becoming an agent. So is the default for booking a thing, and for getting through your inbox, and probably for half a dozen other things I haven’t thought of yet. Google is making Gemini the engine those defaults run on.

I said in December that demos don’t compound and defaults do. Agents are the next default. And heading into the back half of 2026, I’m planning like Google is the one most likely to own them.

Google didn’t win 2025 with a better demo. They won by shipping defaults, and the default they’re going after now is the agent. If you’re a firm leader trying to think through what this means for your operating model, your governance posture, or your 2026 AI budget, I’d be glad to talk it through. Reach me at steve@intelligencebyintent.com. The firms that get out ahead of this won’t be the ones with the smartest model. They’ll be the ones who rewired the work first.

Ken Griffin on AI: Humans Were the Horses

TL;DR: Ken Griffin sat down at Stanford in April and gave the most honest CEO answer on AI I’ve heard yet. Work Citadel used to give Masters and PhDs in finance for weeks is now being done by generic AI agents in hours. He went home one Friday depressed by it. Then he layered the K-12 data on top, and the picture got worse. This is a quick take on why I think his segment matters, and what it should force every leader, parent, and citizen to sit with.

I rarely get genuinely unsettled by a CEO interview.

But I watched one this weekend that I can’t shake. Ken Griffin, founder and CEO of Citadel, sat down with Amit Seru at the Stanford Graduate School of Business. The whole conversation is worth your time. About 19 minutes in, Amit asks him about AI. What Griffin says next is the most honest answer I’ve heard a sitting top-tier CEO give on this topic.

Here’s the video:

The Ferguson Line

Griffin opens by recounting a conversation he’d had the day before with the historian Niall Ferguson. Ferguson walked through the usual story of technology waves. Horse and buggy gets replaced by the car. You know how the narrative goes.

Then Ferguson lands the punch. “The issue with AI is that in the world of AI, humans were the horses.”

Griffin’s response on stage was “that’s a really depressing way to start the day.”

No hedging. No pivot to a feel-good line about co-pilots.

What He’s Seeing Inside Citadel

Then he told the audience what’s actually happening inside his firm.

Work that Citadel used to assign to people with master’s degrees and PhDs in finance, work that took weeks or months, is now being done by AI agents in hours or days. His phrase: “automated by a generic AI.”

These are not the mid-tier white-collar jobs people have been writing about for two years. These are the high-skill research roles. The ones that pay extraordinarily well. The ones we tell our kids to aim for.

Griffin said he went home one Friday actually depressed by it. That’s his word. Depressed.

When the guy running one of the most profitable hedge funds on earth tells a room at Stanford he went home depressed because of what AI did inside his own four walls, that’s not hype. That’s a data point worth taking seriously.

The Race

Griffin didn’t stop there, and neither should we.

He framed the next decade as a race. Jobs will get destroyed at some clip. New jobs will get created, hopefully fast enough to keep up. Griffin is an optimist on the creation side, and part of his case is that the moats that used to protect incumbents are getting filled in by the same tools doing the destroying.

His example was perfect. A friend handed a pet insurance business to his 25-year-old son. The kid built a workflow that scrapes social media for puppy photos, identifies the breed with image recognition, and fires off a custom note. “Congratulations on your new golden retriever. Buy Spot pet insurance.”

They sold the business a few weeks ago.

A billion dollars. With a B. Run by a 25-year-old. Using tools any of us can buy.

One caveat. I went looking for reporting to corroborate the billion-dollar number and couldn’t find any. Spot Pet Insurance was acquired by Independence Pet Holdings in 2024, but financial terms were never publicly disclosed. Griffin told the story on stage at Stanford. I’m passing it along as he told it, with the honest note that I can’t independently verify the price.

And Then He Brought Up Eighth Grade Math

Then Griffin pivoted to K-12 education, and the warning got sharper.

Roughly a quarter of American high school graduates are proficient in math. About a third are proficient in reading. In Illinois a couple of years ago, there were 53 public schools without a single student at grade level in math.

Read that again. Fifty-three schools. Zero students. At grade level. In math.

Now layer AI on top of that. We’re about to ask a workforce that already struggles with basic math and reading to compete with the best minds in China, India, and Europe, in a market where the old moats are gone and the pace of change is faster than any of us have seen.

This is the part of Griffin’s segment that should keep parents and policymakers up at night. The race he described between destruction and creation isn’t even the right framing for a large slice of the population. They can’t read the starting gun.

So What Do We Actually Do With This

I’m not going to pretend Monday morning has a tidy three-step answer. It doesn’t.

But here’s what Griffin’s segment forces all of us to sit with.

The highest-skilled work in your organization is no longer safe. Not the partner-level work. Not the senior analyst pool. Not the research function. None of it. I’ve watched managing partners at law firms this year realize the same thing Griffin realized in his hedge fund: research that used to take an associate two weeks is now coming back in an afternoon. And almost none of them have redesigned how junior talent actually learns judgment. They will eventually. The question is whether it happens on their terms or someone else’s.

If you’ve got kids, the floor on what they need just moved. And it didn’t move toward coding. It moved toward learning. The ability to keep learning, switch fields, and stay curious. That’s the real skill now. Griffin tells his new hires they haven’t finished learning. They’ve just started.

The small-versus-big dynamic also flipped. Twenty-five-year-olds with AI agents are real threats to incumbents who got comfortable behind their old moats. Pet insurance, a sleepy category, just produced a billion-dollar exit [again - I could not corroborate this but he mentioned it so I’m leaving it in]. Pick your sleepy category. Someone’s coming.

And then there’s the K-12 number. I’ll be blunt. We can’t have a serious national conversation about AI policy if half our kids can’t read or do math at grade level. That’s not a partisan point. It’s arithmetic. We need to fix this, and the business community has every reason to be loud about it.

The Last Line

Griffin closed his answer with a question, not a victory lap. He said he doesn’t know where we’ll be in 20 years on AI. He just knows the people in that Stanford room get to help write what comes next.

I keep coming back to Ferguson’s line.

Humans were the horses.

It’s an uncomfortable framing. But the discomfort is the point. The horse didn’t get a vote on the car. We do.

Griffin closed his Stanford answer with a question, not a victory lap. That’s the right posture for the rest of us. The people running firms, raising kids, and writing policy are all going to be asked the same thing in five years: what did you actually do with what you saw coming? If you’re a firm leader trying to redesign your pipeline, your training, or your client work for what’s actually here, I’d be glad to compare notes. steve@intelligencebyintent.com. The horse didn’t get a vote. We do.