Buying Claude Is The Easy Part. Setting It Up Is Where Small Firms Get Stuck.

Last week I sat across from a managing partner who had already made the decision. The firm was going with Claude. Eight attorneys, a couple of paralegals, a real plan to use AI on real matters.

Then I asked the obvious question. Who’s actually setting it up?

Long pause.

That pause is the post.

The pattern I keep seeing

In the past two years I’ve met with hundreds of law firms. Solo shops, mid-sized, AM100. The bigger firms have IT teams. They have procurement, security review, and someone whose actual job is to roll software out across hundreds of users. They might move slowly, but they have the muscle.

Small firms don’t.

Most of them outsource IT to a vendor. Sometimes that vendor has locked the Microsoft 365 tenant down so tight that installing a Word add-in becomes a three-week ticket. Sometimes there’s no real IT relationship at all, and the most technical person in the office is a partner who once set up the printer.

And there’s a quiet assumption baked into the industry that small firms will just figure this out. That if you can pick a tool, you can deploy it. I don’t think that’s true anymore.

What people actually don’t know

Pick any of these questions and ask a managing partner at a six-attorney firm.

Where do you buy a Claude Team plan, and how is that different from someone walking into the office with their own Pro subscription? What’s the difference between a Standard seat and a Premium seat, and which do paralegals get? Should we turn on web search? Connectors? Cowork? Claude in Chrome? Code execution? Memory? MCP?

Most partners will give you an honest answer, which is some version of “I have no idea, and I don’t know who to ask.”

This is a real gap. The frontier labs have built genuinely good products for firms. They’ve shipped Team and Business and Enterprise plans. But they haven’t shipped the thing a 12-person firm needs most, which is a clear path from “we bought it” to “we’re using it well.” The product page assumes a buyer who already knows what a connector is. Most small firms don’t.

My first attempt at filling the gap

Several of my clients recently bought Claude Team. Rather than walk each of them through the same setup twice, I wrote it down.

The result is a step-by-step guide for small firms setting up Claude Team. It covers what to do, in what order, and why. Privacy and admin visibility. Web search and the other capability toggles. Seat types. Roles. The Office add-ins. Starter projects for a small firm. A verification protocol you can hand to your associates before they put anything in front of a court. Where things live in the settings, because nobody has time to hunt.

You can read it here: Setting Up Your Claude Team Account: A Guide for Small Law Firms.

Is it perfect? Far from it. Products change, settings get renamed, new features arrive every month. The guide will need updates. But it’s a starting point that didn’t exist a week ago, and the firms that have used it have gone from “we bought it” to “everyone is logged in, projects are set up, and we’re using it on real work” in about a week.

What’s next

The Claude Team guide is the first of four I’m planning:

1. ChatGPT Business setup for small law firms

2. Claude Enterprise setup for larger firms

3. ChatGPT Enterprise setup for larger firms

Plus I expect to be back to the Claude Team version inside a quarter to update it. Cowork, the Office add-ins, and the connector list are all moving fast enough that any setup guide is a snapshot, not a finished product.

A last thought

If you’re a managing partner reading this, the part I want you to hear is the smallest one. Getting Claude installed is not your AI strategy. It’s the first hour of work in a much longer effort to figure out what your firm actually does with AI. Training. Governance. Billing model decisions. Client disclosure. The harder questions about which work gets pulled forward and what that means for your associates. None of that gets resolved by a setup guide.

But you can’t have any of those conversations until the tool is in your people’s hands.

So start there. The rest is the actual work.

Buying Claude doesn’t make a firm an AI firm. The work starts the moment someone has to decide which seats go to which people, which capabilities get turned on, and what your associates are allowed to put in front of a court. If your firm is sitting somewhere in that gap right now, I’d be glad to help you think through it. Reach me at steve@intelligencebyintent.com. The setup is the first hour. What you do with it is the next decade.

Share

The Apprenticeship Was Always Accidental

A managing partner I worked with showed me a memo last month. First-year associate’s name on it. Three pages. Tight reasoning, clean cites, no padding. He told me it was the best first-year memo he’d seen in a decade.

Then he paused.

“And I have no idea if she could have written it without AI.”

That’s the conversation happening in every law firm, every consulting shop, and every investment bank right now. Quietly. In partner offices. Not in town halls. Not in marketing materials. Not on LinkedIn.

Here’s what nobody wants to say out loud. Professional services firms never really paid for training their juniors. They paid juniors to do the work. The training happened on the side. AI is breaking the labor model. And the training model is going with it.

TL;DR

Law firms, consulting firms, and investment banks built their economics on a pyramid of junior labor marked up to clients. The apprenticeship those juniors got was real, but it was a byproduct. It was paid for by the billable work, not by the firm. AI is compressing the junior work faster than firms have redesigned the training. Now firms have to decide whether they will fund judgment development directly, or quietly let it die while booking the margin.

How the model actually worked

Walk through the math. A second-year associate bills at $600 an hour. The firm pays her maybe $250,000 a year, plus benefits, plus overhead. She bills 2,000 hours. The firm makes a markup on every hour she sits at her desk.

She is not just labor. She is margin.

But she is also learning. She reads 400 depositions and starts to feel where witnesses are evading. She drafts 30 motions and learns what a strong argument feels like versus a polished but weak one. She sits through 50 partner calls and absorbs how senior people talk to clients. None of that was on her job description. None of it was billed as training. The firm did not write a check for it. It happened because she was at her desk, doing the work.

That is the accidental apprenticeship. The grind paid for itself, and the grind also produced the next generation of partners. The firm got both for one price.

It was a quiet miracle of vertical integration. And nobody noticed it was a miracle until AI started pulling it apart.

What AI actually compresses

The work AI is good at is exactly the work that fed the accidental apprenticeship.

First-pass document review. Initial case research. Issue spotting. Chronology building. Deposition summaries. Diligence list generation. Comparable company analysis. Market sizing. First-draft slides. Memo scaffolding. Cite checking. Discovery classification. Pitch book updates.

That’s not a marginal slice of junior work. That’s the substrate of it.

When a firm hands a junior an AI tool and tells her to be efficient, she becomes more productive. She also stops doing the slow reps that used to build her judgment. The work product looks better. The lawyer underneath is thinner.

You can run this experiment in your head. Picture the third-year associate in 2030 who has used AI for everything since law school. Plausible memos. Defensible models. Crisp summaries. Now ask: when something is genuinely hard, when there’s no precedent, when the facts are ambiguous, when a partner says “what do you actually think?”, what does she draw on?

Pattern recognition comes from patterns. AI gives her answers. The accidental apprenticeship gave her the patterns underneath the answers. Those are not the same thing.

The question firms are avoiding

Here’s the question I’d ask any managing partner right now.

Who pays for apprenticeship now that the work no longer pays for it?

There are four real answers. None of them are comfortable.

The firm can pay. That means treating judgment development as infrastructure. Real training time. Real senior teaching hours. Real simulation work. None of it billable. All of it on the partnership’s tab. Margins go down before they come back up.

Clients can pay, but only if firms reframe the value story. Clients will not fund “training.” They might fund verification, audit trails, lower error rates, better risk management on AI-assisted work. That is a real pitch. It just has to be made explicitly, with proof, not as a vague quality claim.

Juniors can pay, in the form of fewer slots, a higher bar, and faster expectations. The old bargain was brutal hours in exchange for credentialing and a path to partnership. The new bargain might be fewer people, more deliberate training, and faster judgment requirements. That is not necessarily worse for the juniors who get in. It is much worse for the ones who would have been hired under the old model.

Partners can pay through smaller pyramids and changed compensation. This is the part firms will resist hardest. The pyramid is how partners get paid. If the bottom of the pyramid shrinks, partner draws compress, or the comp model has to change.

Most firms, if they answer honestly, will need all four. Spread across all four payers, the redesign is manageable. Concentrate it on any one and the politics break.

The harvest temptation

Here is what will tempt most firms instead.

Give every junior AI. Cut the bottom of the pyramid. Celebrate the productivity. Book the margin. Hope nobody notices for five years that the partnership pipeline is hollowing out.

This is the rational short-term move for any individual firm. It is also the move that liquidates the future of the firm.

I’ve watched a version of this happen in other industries. Newsrooms cut their training pipelines in the 2000s and discovered a decade later they had no mid-career editors. Manufacturing offshored its skilled trades and discovered it could not staff a domestic factory line. The pattern is always the same. You can harvest the labor savings for a while. Then you wake up and find you cannot replace yourself.

The collective action problem makes this worse. The firm that does the redesign well will be more expensive in the short run than the firm that harvests. Clients say they care about quality. Procurement often selects on price. If the responsible firms get punished by the market for being responsible, the irresponsible move spreads.

That is the deepest risk in this whole story. Not that AI replaces juniors. That AI lets every firm skip the discomfort that used to produce expertise, and the market does not reward anyone for not skipping it.

What to do Monday morning

If you run a firm, or a practice group, or a training program, here is where I’d start this week.

1. Map your junior work into three buckets: pointless drudgery AI should compress, developmental volume AI can assist but not replace, and judgment friction you should actually increase. Get specific. If you can’t name the developmental volume in your practice, you don’t have a training model. You have a labor model.

2. Pull the real cost of judgment development out as its own line item, separate from the cost of junior labor. If you can’t see it as a budget, it isn’t a budget. And if it isn’t a budget, it will not survive the AI transition.

3. Talk to three clients about what verification and audit trails on AI-assisted work would be worth to them. Not a survey. A real conversation. You need to know whether the client-pays-for-quality story has any actual buyers in your market.

4. Tell your senior partners that the next review meeting will not be about line edits. It will be about the junior’s reasoning. What they verified. What they changed. What they still aren’t sure about. Then teach them how to run that meeting.

5. Decide what your pyramid looks like in 2030 and work backwards. If you are going to hire fewer juniors with a higher bar, you need to start now. The selection problem takes years to solve, not quarters.

The line worth remembering

Professional services never paid for apprenticeship directly. They smuggled it inside the economics of junior labor. Now that AI is disrupting the labor, firms have to decide whether developing judgment is an asset worth funding, or just a cost they used to hide.

The firms that figure this out fast will produce better professionals faster than the old model ever did.

The ones that don’t will look fine for a while.

Then they won’t.

The firms that figure this out won't announce it. They'll be quietly redesigning what juniors learn, what partners teach, and what clients are willing to pay for. The ones that don't will look fine for a while. Then they won't. If you're running a practice group, a training program, or a firm, and you want to think out loud about what this looks like in your shop, reach out: steve@intelligencebyintent.com. The apprenticeship was always accidental. Whatever replaces it has to be deliberate.

Subscribe now

Most Law Firms Live in Two Modes of AI. There Are Five.

TL;DR: There are five modes of AI work: Ask, Draft, Co-work, Delegate, Orchestrate. Most law firms operate in the first two. Microsoft’s 2026 Work Trend Index, published May 5, found that the most advanced AI users are producing 22 percentage points more output than typical users, and that two-thirds of that gap comes from organizational design rather than individual talent. The capacity is sitting unclaimed because most firms haven’t built the systems to capture it.

I’ve been giving a version of the same talk for the last six months. Different rooms, different cities, different firms. The slide that lands hardest is the one with five modes of AI work on it. Ask, Draft, Co-work, Delegate, Orchestrate.

I show it. I walk through it. Then I ask the room which mode their attorneys actually operate in.

The answer is almost always the same.

The first two.

On May 5, Microsoft published its 2026 Work Trend Index. Trillions of M365 productivity signals. 20,000 AI users surveyed across ten countries. The closest thing the industry has to a definitive read on how people actually work with AI right now.

Two findings stop me cold every time I revisit the report.

First. 58% of all AI users say they’re producing work they couldn’t have done a year ago. Among the most advanced users (Microsoft calls them Frontier Professionals, defined by their use of multi-step agent workflows, multi-agent systems, and routine workflow redesign), that number jumps to 80%. Twenty-two points. Same tools. Same training programs available to anyone who asks. The difference is which modes those people actually use, and how often.

Second. Microsoft tested a long list of factors against AI impact. Individual mindset, AI familiarity, job level, generation, market, industry. Then organizational stuff: culture, manager support, talent practices. Want to know which side won?

Organizational factors accounted for 67% of the variation. Individual mindset and behavior, 32%.

Two times the impact. From the system around the person, not the person.

That second number should keep firm leaders up at night. The question isn’t “are my people good with AI.” It’s “have we built a place where good AI work is possible.” Most firms have not. Most firms haven’t even tried. Most firms are still treating AI like Westlaw with a chat interface.

Let me show you what I mean.

The Five Modes

1. Ask. You open a tab, ask a question, get an answer, close the tab. Statute of limitations on breach of contract. Holding in a case. The drive-by use of AI. Most attorneys live here.

2. Draft. You feed AI a starting set of facts and let it produce the first version. A demand letter. A first cut at an indemnification clause. You edit. You take ownership. The model does the heavy lifting on the page; you do the thinking on the substance. A lot of attorneys have started living here too.

3. Co-work. You’re in the document. AI is working alongside you, pointing things out as you go. Section 7’s indemnification language is unusual, here’s the cite. Section 8’s limitation of liability is broader than what’s standard in this industry. AI sitting next to you, watching, suggesting, citing. You’re still driving. You just have a much smarter passenger.

4. Delegate. You hand AI a project and walk away. Review these 326 contracts and flag the unusual ones. Run diligence on this target. Build the closing checklist and pull documents into the data room. AI runs end to end. You review the output, not the process. This is the mode that requires real infrastructure: validated outputs, human checkpoints, audit trails.

5. Orchestrate. You’re supervising a system of agents handling different pieces of the practice in parallel. Intake. Review. Research. Diligence. Billing. Each agent autonomous, all coordinated. The lawyer’s job becomes architecture and quality control.

Where Most Firms Actually Live

Modes 1 and 2. That’s the whole footprint.

This isn’t a guess. It’s the consistent finding from every law firm leader survey I’ve seen for the last year, and it’s consistent with what Microsoft just published. Most users sit in low-intensity patterns. Most organizations haven’t built the systems to support higher-intensity use. Microsoft’s data tells you what that costs: 22 percentage points of additional output, sitting unrealized because the system around the user wasn’t designed to support it.

Now think about what staying in modes 1 and 2 means at a firm level. Faster legal research. Faster document drafts. The 10 to 15% efficiency lift that almost everyone gets from baseline AI use.

Helpful. Your competitors get the same lift. Net difference: zero.

The capacity gain is in modes 4 and 5. Work that used to take three associates a week now takes one and a smart workflow. Routine review pipelines that used to run on weeks-long cycles run overnight. Honestly, it stops looking like lawyering as the profession recognizes it. The work is still legal work. The cost structure for delivering it is something firms in modes 1 and 2 can’t price against.

Mode 5 Is Not Theoretical Anymore

A year ago, Mode 5 was a slide in a deck. It’s not anymore.

Harvey put 500-plus pre-built legal agents into early access on May 5, the same day Microsoft published its report. Slaughter and May, the magic circle firm, announced a firmwide Harvey rollout on April 29 covering M&A, due diligence, regulatory research, and document analysis. Manifest OS, which closed a $60 million Series A at a $750 million valuation in late April, runs an AI-native immigration practice posting visa approval rates 15% above the national average and client response times three times faster than traditional firms.

These are different shapes of business. Smaller staffing pyramids. Different fee structures. Different output at different price points. The fact that they exist right now, posting measurable outcomes, is the part most firm leaders haven’t reckoned with yet.

Why Firms Stay Stuck

Three reasons. They wrap around each other so tightly that pulling at one without dealing with the others tends not to work.

The first is trust. Most attorneys don’t trust an AI’s first draft enough to take it as a starting point, much less hand off an entire project. Fair enough. They’ll get there or they won’t. Time and exposure to good outputs handle this one over the next twelve months.

The second is billing. This is the uncomfortable one. If your firm bills by the hour and Mode 4 lets your associates produce in one-quarter of the hours, you have a revenue problem disguised as a productivity win. The math is brutal. You either bill less for the same work, or you find a way to capture the value of finishing faster, which means alternative fee arrangements, fixed pricing, value-based scopes, anything that decouples revenue from time spent. The profession has been talking about this since at least 2010 and most firms still haven’t solved it. The market is now going to solve it for them, one client conversation at a time. The firms that haven’t sorted this out have a structural incentive to keep their associates in modes 1 and 2, where AI helps but doesn’t compress the hours billed. They won’t say that out loud. They don’t have to. The behavior tells you.

The third is training. Almost every AI training program I see in law firms teaches people how to prompt and which buttons to click. Almost none teach people how to specify a piece of work in a way an AI can execute, how to evaluate AI output against a quality standard, or how to architect a workflow that uses AI well. That’s the gap between Mode 2 and Mode 4. It’s a training problem, not a tools problem. A firm could close it in a quarter if it decided to.

What to Do Monday Morning

Three things, in order.

One. Pick five attorneys at different levels and ask them, in detail, how they used AI in the last week. Listen for the mode. If everything you hear is in modes 1 and 2, you have your answer about where your firm actually operates.

Two. Pick one practice area and one workflow inside it that’s a candidate for Mode 4. Contract review. Due diligence. Intake. Anything routine and high-volume. Build the case for what it would take to move that workflow from Mode 2 to Mode 4. What evaluation criteria. What human checkpoints. What infrastructure. Don’t try to boil the ocean. Pick one workflow.

Three. Have the billing conversation. Honestly. With your finance partner, with your practice group leaders. If you successfully move work into Mode 4 and your associates produce in one-quarter of the time, what is your fee structure on that work? If you don’t have an answer, you have a structural disincentive to actually move up the modes. Solve that or stay where you are.

The Punchline

Microsoft’s data shows the gap exists right now. Not theoretically, in 2027 or 2030. Right now, in the data they collected between February and April of this year. Firms operating in modes 1 and 2 are getting one return on their AI spend. Firms in modes 4 and 5 are getting a meaningfully different one.

You can keep telling yourself you’re doing AI because your associates use it for research. Or you can look at where your firm actually operates and decide whether you’re satisfied with where the data says that path ends.

The five-mode map is a diagnostic, not a destination. The firms moving from mode two to mode four aren't working with smarter people or fancier tools. They've built different systems, asked different questions, and had the billing conversation their competitors are still avoiding. If you want to figure out where your own firm sits on this map, or what it would take to move it, that's a conversation I'm always up for. Reach out: steve@intelligencebyintent.com. The firms that own this decade are deciding right now what they're willing to build.

The Verification Gap

TL;DR: A Georgia prosecutor just lost the right to practice before her state’s Supreme Court because she didn’t verify the AI-generated authorities in her brief. About 30 of them were either invented, attributed to the wrong cases, or quoted text that wasn’t there. The day before, the California State Bar wrapped public comment on a rule package that would make “independently review and verify” any AI output an enforceable duty under the Rules of Professional Conduct. Verification is moving from best practice to baseline competence, and most attorneys still don’t have a system for it.

A six-month suspension for not checking the cites

On May 5, 2026, the Supreme Court of Georgia suspended Clayton County Assistant District Attorney Deborah Leslie from practicing before the court for six months. To get reinstated, she has to complete 12 hours of continuing legal education focused on AI in legal practice.

The story is uncomfortable to read. Leslie was handling the appeal of Hannah Payne, who is serving a life sentence for the 2019 fatal shooting of Kenneth Herring. During oral argument, the justices flagged nine citations in the state’s brief that either didn’t exist or didn’t say what Leslie claimed they said. The court asked her to explain.

She did. And the explanation made it worse. She acknowledged that she had used AI software to draft her reply briefs and the trial court’s order denying Payne’s motion for a new trial. After further review, the count rose. Twelve more cases she had cited at the trial level were also AI-generated and unverified. Then she withdrew nine more from her appellate brief that misstated the holdings or didn’t correspond to real Georgia or federal precedent.

That’s roughly 30 problematic authorities across multiple filings on a murder appeal. And here’s the part worth slowing down for: the AI didn’t just make things up. It also cited real cases for propositions they didn’t support, and quoted language that wasn’t in the actual opinions. So “verification” isn’t just “does this case exist?” It’s also “does it say what I’m claiming it says?” and “is this quote actually in the opinion?” The Georgia opinion lays all three failure modes side by side, and the court treats all three as the same problem.

Justice Benjamin Land’s eight-page opinion is the kind of language that ends careers: “These filings, as well as the trial court’s order, contain multiple case citations which either do not exist, or which exist but do not support the propositions of law for which they are cited. While we have no rule against the responsible use of artificial intelligence software by attorneys, citing cases that do not exist or do not support the proposition for which they are cited is a violation of this Court’s rules and falls far beneath the conduct we expect from Georgia lawyers.”

This is the first time I’m aware of that an AI-citation problem has cost an attorney the right to practice in a particular court. Not a fine. Not a public scolding. A suspension.

This didn’t come out of nowhere

Three years ago, Mata v. Avianca was the punchline. Two New York lawyers and their firm filed a brief with ChatGPT-generated cases, got hit with $15,000 in combined sanctions, and the legal world spent six months making jokes about the lawyer who didn’t know ChatGPT could lie.

The arc since then has not been funny.

I’ve been keeping a running list of these cases since Mata happened. The list is never the same length the next time I look at it. The pattern in 2023 was “lawyer didn’t realize AI could fabricate.” The pattern now is “lawyer knew AI could fabricate and didn’t check anyway.” The second version is much harder to defend in front of a judge.

In the first quarter of 2026 alone, U.S. courts imposed at least $145,000 in sanctions for AI-fabricated citations, according to research compiled by ComplexDiscovery. Most of that came from two clusters: $109,700 in Oregon, where the state appellate court built a per-citation fee schedule (think $500 per fake case, $1,000 per fake quote), and a $30,000 fine from the Sixth Circuit on two attorneys in a consolidated appeal.

The Sixth Circuit case matters because federal appellate courts had mostly been quieter on this. Not anymore. The court chose the elevated penalty because, in its words, smaller fines hadn’t been working.

Then in April, Sullivan & Cromwell, OpenAI’s outside counsel and self-styled expert on the safe and ethical deployment of AI, sent an emergency letter to a federal bankruptcy judge in Manhattan asking him not to sanction them. Their motion in a Chapter 15 case had roughly 40 errors, including citations to cases that don’t exist. The firm has policies. Required training. Office Manual language saying “trust nothing and verify everything.” The reviewer just didn’t catch it.

If S&C can’t catch this, your firm needs to assume you can’t either.

Ten days later, a federal magistrate judge in North Carolina publicly reprimanded Rudy Renfer, a former Assistant U.S. Attorney with a 30-year career and 17 years at the Eastern District of North Carolina USAO, for filing a brief built on AI-generated fabrications. The judge wrote that Renfer “intentionally submitted a brief containing false materials to the court” and that “in this court, his name will be synonymous with a failure to uphold the basic duties of competence and candor expected of every attorney.” Renfer is now under investigation by the Department of Justice’s Office of Professional Responsibility. He had already left the U.S. Attorney’s office in March.

California’s about to make it enforceable

While courts have been escalating penalties case by case, California is moving the question into the rules themselves.

The State Bar’s Standing Committee on Professional Responsibility and Conduct (COPRAC) approved proposed amendments to six Rules of Professional Conduct on March 13, opened a 45-day public comment period, and closed comments on May 4. The rulemaking was triggered by an August 2025 letter from the California Supreme Court’s clerk to the state bar asking for AI-specific rules.

The Rule 1.1 amendment is the one that should get every managing partner’s attention. The new comment language says that when using technology, including AI, a lawyer “must independently review, verify, and exercise professional judgment regarding any output generated by the technology that is used in connection with representing a client.”

Read that one again. There’s no exception for routine work, no carve-out for low-stakes matters, no safe harbor for “I trusted the vendor.” I’ve sat with managing partners who assumed the rule would leave wiggle room for triage on what to verify and what to skim. There isn’t any.

Rule 1.6 is the other quiet bombshell. It expands the meaning of “reveal” to include “exposing confidential information to technological systems, including artificial intelligence tools” where there’s material risk the information could be accessed, retained, or used in ways inconsistent with confidentiality duties. Lawyers who paste client data into consumer AI tools with unfavorable retention terms now have a confidentiality problem on top of a competence problem.

Rule 3.3 codifies the verification duty for citations submitted to a tribunal. Rules 5.1 and 5.3 push AI governance up to managing partners and out to paralegals and other staff. None of this is novel ethics. It’s just that California is willing to write the obvious into rules that carry disciplinary force.

And courts aren’t waiting for California to formalize this. In Lacey v. State Farm last year, a federal Special Master in the Central District of California hit two AmLaw firms with about $31,000 in combined sanctions over AI-fabricated citations in their filings. The senior lawyer on the team formally accepted responsibility in writing for failing to supervise the AI-assisted research. The supervisory duty is already being enforced. California’s proposed Rule 5.1 would just write it into the disciplinary rules.

The proposals still need COPRAC review of public comments, then State Bar Board of Trustees approval, then sign-off from the California Supreme Court. They’re not yet in effect. But California has a habit of leading on professional responsibility, and if these go through, every other state bar gets a template.

What this changes for your firm

The math has changed. AI cuts research and drafting time by real numbers, and that isn’t going away. But the cost of skipping verification is no longer just an embarrassing post on legal Twitter. It’s a $15,000 fine. It’s six months when you can’t appear in your state’s highest court. It’s a letter to a federal judge that a partner’s name will be Googled with for the rest of her career.

Treat AI output the way you’d treat a brief from a smart but unsupervised first-year associate. Useful. Faster than starting from scratch. Absolutely not ready to file.

What to do Monday morning

Three things that move the needle without overhauling your practice.

1. Write the verification step into the workflow. Not a memo, not a poster in the conference room. A required checkbox in your matter management or document review process that says “every cited authority in this filing has been pulled, read, and verified by a human.” If you can’t track it, you can’t enforce it.

2. Update your AI policy to address consumer versus commercial terms. Most lawyers can’t tell you whether the AI tool they used last week trains on their inputs. I ask managing partners this all the time and the answer is usually a long pause. Your policy should name which tools are approved for which kinds of data, with the contract terms written down somewhere a partner can actually find them. The best contract beats the best model.

3. Train the supervisors, not just the associates. Lacey already proves the point. Partners need to know what AI tools their teams are using, what verification looks like, and how to catch a hallucination in a cite check before it ships.

The line is moving

Three years ago, citing a fake AI case got you a fine and a bad week. This month, it cost a Georgia prosecutor her standing before her state’s highest court. If California’s rule package goes through, doing the same thing in California becomes a Rules of Professional Conduct violation, with discipline that can run all the way to suspension.

Verification used to be a virtue. It’s about to be a rule. The lawyers who’ve built a system for it will be fine. The ones who haven’t are running on borrowed time, and the clock just sped up.

Three years ago, this kind of mistake got you a fine and a bad week. Now it costs a six-month suspension before your state's highest court. By summer, in California at least, it might cost a license. The firms that build a verification system this quarter will spend the next three years compounding the time savings AI gives them. The ones that don't will spend that time explaining themselves to judges, clients, and disciplinary boards. If you want to talk through what a workable verification process looks like for your practice before the rules force the conversation, reach out at steve@intelligencebyintent.com.

You’re Paying for the Best AI Model. You’re Not Using It.

TL;DR: AI companies have buried the best models behind confusing pickers, and most apps default new users to the cheapest, fastest version. Worse, when paying users do pick the premium model, several apps silently reset them on the next chat. The result is that senior professionals at firms paying $20-$125 per seat per month are getting entry-level answers without realizing it. The fix is one rule per platform. Claude users want Opus 4.7 with Adaptive Thinking on. ChatGPT users want Thinking at the deepest level, or Pro. Gemini users want Pro. Grok and CoPilot users, same logic: skip Auto, find the deepest reasoning option, set it. Then check the picker on every new chat, because the apps don’t always remember.

I was presenting to a law firm this week, walking through how the different Claude models work and when you’d use each one. About ninety seconds in, the lead attorney held up a hand.

“Stop. Just stop. We don’t need to know. We don’t care. We will never change the model. Just tell us the one we should use.”

She wasn’t being rude. She was being honest. The firm runs Claude Teams. They handle work where quality matters more than token economics. They weren’t in the mood for a recipe. They wanted a rule.

I told her: Opus 4.7 with Adaptive Thinking on. That’s the one. Set it. Don’t touch it again.

And the thing is, four other clients had told me versions of the same story in the past two weeks. A managing partner pressure-testing a hiring decision. A COO trying to draft a board update. A practice group leader summarizing a deposition. A general counsel reviewing a vendor contract. Different platforms, different uses. Each one, in a slightly bewildered tone, said something like: “There are four options in the menu and I have no idea which one to pick, so I just use whatever comes up first.”

That’s the problem. And it’s quietly costing them the thing they’re paying for.

The labs built complexity into the front door

Open Claude. The picker shows Opus 4.7, Sonnet 4.6, Haiku 4.5, plus an Adaptive Thinking toggle. ChatGPT: Instant, Thinking (with multiple levels of thinking inside it), Pro. Gemini: Fast, Thinking, Pro. Grok: Auto, Fast, Expert, Grok 4.3 Beta. CoPilot is its own animal. Auto, Quick Response, Think Deeper, Opus, and a GPT option that opens up five more sub-models underneath it.

If you live inside AI every day, you know what these mean. If you open Claude twice a day to pressure-test a memo, you don’t. You see a dropdown and pick whatever’s at the top. And in most apps, the top option is a cheaper, faster, lower-capability model designed to keep cloud costs down across millions of free users.

I get why the labs do this, by the way. They’re balancing rate limits, latency, infrastructure cost across hundreds of millions of free accounts, and what their best models can crank out per second. There’s also the inference economics question, which I’ve spent more time than I probably should thinking about: every reasoning-model query costs the labs meaningfully more compute than an instant answer, and at scale that math has to come out of somebody’s pocket. From their seat, defaulting most users to the lighter model is a survival tactic. From the seat of the executive who paid full freight for the premium subscription, it’s something else entirely.

The math is worse than you’d think. You’re probably paying $20-$125 per user per month for one of these subscriptions. Fifty seats is $15,000 a year, minimum. Most of those seats are getting answers from a model a generation behind what they’re paying for.

It’s like buying a sports car and never shifting out of first.

And the apps actively work against you

Here’s the part that genuinely drives me nuts.

Even when a user has done the right thing, picked the premium model, set their preference, gotten comfortable with the answers, the apps will quietly reset them. Open a new chat in several of these clients and the model selector flips back to Auto, or Instant, or Fast. The default the user picked is gone. They’re back on the cheap model, and they often don’t notice for a while.

I had three clients get caught by this in the past week. They opened new chats to continue work they’d been doing successfully the day before, got back answers that were thin or partly wrong, and emailed me with some version of “WTF, I thought we were paying for the good model.” They were. The app had silently flipped them back to the entry-level default.

The labs will tell you this is about managing cost across hundreds of millions of users. Fine. From their seat, that’s true. From the customer’s seat, the message is: we sold you the premium tier, then handed you the entry-level product on the next page. Nobody designs a paid subscription that way on purpose. Or if they do, they shouldn’t.

So here’s what I tell clients. Claude: Opus 4.7 with Adaptive Thinking on. ChatGPT: pick Thinking and run it at the deepest level your plan offers, or jump to Pro if you have access. Gemini: pick Pro (you’ll be on 3.1 Pro). Grok and CoPilot users, same logic. Skip Auto. Find the deepest reasoning option in the menu. Leave it set.

Then check the picker every time you open a new chat. Yes, that’s a tax. Yes, it’s annoying. It still beats the alternative.

“But what about quick lookups?”

I had this exact exchange with an associate at a client firm last month. “Isn’t the cheap model fine for emails and quick stuff?” Honest answer: if a question is genuinely small enough that the cheap model is “fine,” Google can answer it faster than any LLM. The reason you opened Claude is because the question wasn’t small. You wanted synthesis. You wanted real writing. You wanted something thought through.

For all of that, you want the best model in the box. Every time.

The standard advice is to match the model to the task. I don’t buy it. Asking busy executives to triage their questions before they ask them is asking the wrong people to do the wrong work. They want to type a question and get a great answer. The picker should never be their problem.

The thinking models actually work the problem before they answer. They follow multi-step instructions. They catch their own mistakes. They hallucinate less. The gap between a thinking model and an instant one isn’t subtle on real questions. It’s the difference between an answer you can use and one you have to redo.

What to do Monday morning

Send your firm one note. Tell people which model to use on each platform you’ve licensed, tell them to set it as the default, tell them to check the picker at the start of every new chat because the app won’t always remember, and tell them why in one line: you’re paying for the top model, so use it.

That’s the whole memo. It pays for itself the first time someone asks a substantive question and gets a substantive answer instead of a confident-sounding skim.

The labs need to fix both halves of this. Premium subscriptions should default to premium models, and a user’s chosen model should stick when they open a new chat. The fact that neither is true is on them. The fact that you’re letting it stay that way at your firm is on you.

You're paying for the best AI in the box. Make sure your firm is using it. Send the memo, set the rule, and check the picker on every new chat until the labs fix what they should have fixed at launch. If you want help rolling a one-page rule out across your seats without turning it into a meeting about a meeting, reach out: steve@intelligencebyintent.com. The model you picked should be the model you get.

The Five Modes of AI Work: Why Most Law Firms Are Stuck at Two

TL;DR: AI in law firms is moving from a thing you ask to a thing you delegate to. There are five modes of use, from simple Q&A to running fleets of agents. Most firm professionals are stuck at modes 1 and 2. The early adopters are reaching mode 3. Modes 4 and 5 are where the operating model actually changes, and the trajectory is hitting different roles at very different speeds. Assistants and finance teams have a 12 to 24 month window. Paralegals are looking at hard repositioning inside two years. Lawyers are insulated by judgment but transformed by expectation. The firms that win won’t be the ones with the biggest tech budgets. They’ll be the ones whose people are furthest up the ladder.

A conversation that’s becoming familiar

I had a call last week with a managing partner who told me, “We’re all in on AI.” I asked what that looked like in practice. He said his associates were using Claude and ChatGPT for first drafts. His paralegals were summarizing depositions. His marketing team was running a couple of automations.

I didn’t say it out loud, but here’s what I was thinking: that’s the easy part. That’s not the part that changes the firm.

The chatbot era of AI was about substitution. Use AI instead of Google. Use AI instead of a blank page. It made each individual worker faster at doing the same job. That’s real value, and I don’t want to dismiss it. But it doesn’t change the shape of how a firm operates.

What’s coming next is different. AI is moving from a thing you ask to a thing you delegate to. Not faster Google. More like a junior colleague who can be told “handle this.” That’s the shift. And it’s already happening, unevenly, across the roles that make up a typical firm.

The five modes of AI work

Here’s the frame I use. Across the research I’ve seen and the work I’ve done with firms, the same trajectory keeps showing up under different names. This is the cleanest version.

1. Ask. Information retrieval. “What’s the statute of limitations in California for breach of contract?”

2. Draft. Content generation. “Write me a first-draft demand letter for this wage-and-hour claim.”

3. Co-work. You give AI a goal and it executes against it, browsing, drafting, pulling files, organizing, until there’s an outcome to look at. You can watch it run and redirect mid-flight, but you’re not driving every step. “Research the leading New York cases on tortious interference over the last five years. Pull the key holdings into a one-page memo and drop it in the matter folder.”

4. Delegate. You hand off a discrete project end to end. AI runs the work, you audit and refine. “Run the full first-pass review of these 2,400 documents against the privilege log criteria we discussed. Flag anything borderline and give me a summary.”

5. Orchestrate. You design and supervise systems where AI agents run processes from start to finish. “Monitor every active matter for deadline risk and budget variance. Run weekly status updates, escalate anything that needs human judgment, and keep the dashboards current.”

Most firm professionals are doing modes 1 and 2 today. The early adopters are reaching mode 3. Mode 4 is technically possible right now, but rare in practice, because firms haven’t built the trust, the governance, or the workflows for it. Mode 5 exists in pockets. E-discovery vendors. Some billing automation. But it’s years from being typical for most firms.

The thing to understand is that each step up the ladder isn’t just “more capable AI.” It’s a different relationship between the worker and the work. At mode 2, you’re a writer using a tool. At mode 4, you’re a director reviewing somebody else’s draft, except the somebody else is the AI. A writer produces. A director judges, chooses, rejects, sends it back. Those aren’t the same job, and the skills that make somebody great at one don’t automatically carry over to the other. Neither does the staffing math, the training program, or honestly what a good day at work feels like. Most firms haven’t sat with that yet.

How the modes land in different roles

Here’s where the strategic implications start to bite. The trajectory isn’t moving at the same speed across every role. That asymmetry is the part most firms get wrong when they plan training and staffing.

Legal assistants and secretaries

This group is both the most exposed and the best positioned to benefit. Their core work today, scheduling, formatting, intake, document binders, routine correspondence, is exactly the bounded, system-driven work that mode 4 handles well.

But “exposed” doesn’t mean “redundant.” The path forward isn’t fewer assistants. It’s assistants who run AI systems that handle the volume work, freeing them up for relationships, exceptions, and the higher-judgment tasks they were never given enough time for. Twelve to eighteen months from now, the assistants who adapt will be running small fleets of automations. The ones who don’t will be doing the same work as today, just compressed into less of the day. That’s a worse position to be in, not a better one.

Finance and billing

Same story, slower clock. Time entry narratives, invoice generation, A/R follow-ups, profitability analysis, rate card enforcement, client guideline compliance. All of it is moving from mode 2 (AI helps the analyst) to mode 4 (AI runs the workflow, the analyst supervises).

The blocker isn’t capability. It’s trust, integration with practice management systems, and partner buy-in. Two-year horizon for finance teams to shift from running invoices to designing the workflows that run them.

The real promise here isn’t cost reduction. It’s that finance stops being a backward-looking reporting function and starts being a real-time business advisor. Instead of “here’s last month’s realization,” finance can say, “this matter is trending eighteen percent over budget, here are the likely causes, here are three interventions.” That changes the role’s value to the firm.

Paralegals

This is the hardest conversation in the firm. First-pass document review, summarization, deposition prep, citation checking, chronology building. Everything that fills a junior paralegal’s day is already mode 3 work and is moving toward mode 4.

The good news: paralegals who become AI orchestrators, designing review protocols, validating output, running case workflows, become more valuable, not less. The bad news: routine paralegal hours billable to clients are going to compress, and they’re going to compress fast.

Two-year horizon is when this gets uncomfortable for firms that haven’t repositioned the role. Repositioning doesn’t mean turning paralegals into lawyers. It means evolving them into matter intelligence specialists whose value is structuring the work, validating sources, maintaining evidence integrity, and catching what the model misses.

Lawyers

Most insulated. Most changed.

Insulated because judgment, advocacy, client trust, and strategy don’t delegate well to AI yet, and won’t anytime soon. The professional rules require a human in the loop, and recent sanctions over AI-generated filings have made it clear that “I trusted the tool” isn’t a defense.

Changed because the work product lawyers are expected to produce will increasingly start as AI output they shape, not blank pages they fill. Within three years, a lawyer who can’t direct an AI workflow that produces a first-draft brief, research memo, or deal summary is going to be working at a meaningful disadvantage to one who can.

And here’s what I want firm leaders to hear most clearly. The skill isn’t “using AI.” The skill is knowing what good output looks like and being able to spot what’s wrong with the AI’s draft. That skill takes years to build. Which is exactly why partners can’t afford to skip it themselves and just push it down to associates.

The asymmetry, on one screen

Role Horizon and new value Legal assistants 12 to 18 months. From task execution to running automation fleets and managing exceptions. Finance and billing 18 to 36 months. From backward-looking reporting to real-time business advice. Paralegals 24 months to repositioning pressure. From document processing to matter intelligence. Lawyers 36 months to expectation shift. From blank-page producer to AI workflow director.

None of those numbers are precise, and I’d argue with anybody who claims theirs are. But the spread between them is real. Plan training, hiring, and pricing as if all four roles are on the same curve and you’ll mis-staff every part of the firm at once.

The training trap

The biggest mistake I see firms make is treating AI training as a one-time event. A vendor comes in, runs a two-hour CLE on prompt writing, hands out a tip sheet, and leaves. I’ve sat in plenty of these. They don’t work, and the reason isn’t really the vendor.

The half-life of any specific tool skill is now under a year, maybe less. UIs change. Models change. New products launch and old products absorb their features within a quarter. If you train your firm on the version of Claude or ChatGPT that exists today, half of what you taught is going to be wrong by the time the partners actually get around to using it.

What doesn’t change is the underlying skill stack. Workflow design, output evaluation, verification discipline, governance literacy. That’s the short version. The longer version is that workers need to know which tasks are safe to automate (and which aren’t), how to spot AI being confidently wrong, how to verify outputs against actual sources, and how to stay on the right side of confidentiality and the rules of professional conduct. None of that goes away when the next model ships.

The firms doing this well don’t run quarterly seminars. They run internal communities of practice, do monthly use case reviews, and carve out protected time for people to actually experiment. They treat AI capability as something the firm builds, not something it buys. The firms doing it badly are buying licenses and hoping people figure it out, which works about as well as you’d think.

How fast, and where this is going

I want to be honest about something. The time horizons people quote, mine included, are guesses. Anyone telling you with confidence what the firm of 2030 looks like is selling something.

What’s not a guess is the trajectory. The modes are real. The role-by-role asymmetry is real. The firms that adapt fastest won’t necessarily be the ones with the biggest tech budgets. They’ll be the ones whose people are furthest up the ladder, with the trust and governance to operate there.

My rough estimate, and you should hold it loosely: five years before the operating model of a mid-sized firm looks meaningfully different from today. Three years before staffing math starts to bend. Twelve months before the firms that have done the work pull visibly ahead of the ones that haven’t. Some firms will move faster than that. AmLaw 20 firms with mature governance, and specialized boutiques where the partners themselves are already using these tools daily, are running a year or two ahead of the curve. The variance across the industry is going to widen before it narrows.

What to do Monday morning

If you run a firm or a practice group, here’s where I’d start.

1. Run an honest audit of where your people actually are on the modes ladder. Survey by role: what mode are most of your assistants, paralegals, finance staff, and lawyers operating in today? You’ll find more variance than you expect. The gap between your most adapted person and your least is itself a strategic risk, especially if the most adapted ones are lateral hires you can’t afford to lose.

2. Pick one workflow per practice group to push to mode 4. Just one. Something bounded, low-risk, and high-volume, like a pre-bill review process or a closing checklist or a litigation status packet. The temptation will be to pick five things at once. Don’t. Run the one as a real pilot, with real metrics, and treat the lessons it surfaces as more valuable than the productivity gains.

3. Replace AI training with an AI practice program. Less classroom. More monthly use case sessions, paired work, and protected time for people to experiment without billing pressure. The vendors will sell you a curriculum. You don’t need a curriculum. You need a culture that treats getting better at AI as part of the job.

That’s the whole list. I know it sounds underwhelming after all the talk about modes and ladders and asymmetric timelines. But most firms aren’t suffering from a strategy gap right now. They’re suffering from a starting gap. Don’t let a vendor talk you into a transformation roadmap before you’ve done the audit.

The line that should stick

In the chatbot era, the question was: what should I ask?

In the era we’re entering, the question is: what should I delegate, what should I verify, and what should I never let a machine decide alone?

The firms that figure that out first will look very different from the ones that don’t.

The chatbot era taught firms how to ask better questions. The era we're entering asks something harder of firm leaders: knowing what to delegate, what to verify, and what to never let a machine decide alone. The firms that figure that out first will look meaningfully different from the ones that don't, and the gap is opening faster than most partners think. If you're working through what this means for your training, your staffing, or your pricing model, I'd genuinely like to hear what you're seeing. Reach me at steve@intelligencebyintent.com.

The China AI Panic Doesn’t Reach Your Firm (Yet)

TL;DR: The US-versus-China open-source AI debate is real. It’s also not your problem this year. Here’s why the premium you’re paying the frontier labs is the cheapest line item in your AI budget.

A managing partner forwarded me a YouTube video this week with two words: “Should we be worried?”

The video was good. It was made by a friend of mine, Matthew Berman, and it walks through a real argument that’s been picking up steam: US open-source AI is losing to China, and if it keeps losing, the country has a problem. Subsidized Chinese labs. No working business model on our side. A developer base that may end up building on Qwen and DeepSeek instead of anything American. He’s not wrong about the macro story.

The next day, by coincidence, I was sitting with a mid-sized private equity firm having exactly this conversation. They’d seen the same headlines. The Chinese open-source models are real. They’re cheap. They’re good. Should they be looking at them?

I told them what I told the managing partner, and what I’m telling you now. None of it changes what your firm should do this year.

That’s the whole article. If you want to stop reading here, the headline is: take the debate seriously at the country level, ignore it in your buying decision, keep going with your closed-source vendor of choice. The premium you’re paying the frontier labs is buying you something Chinese open weights can’t sell you, and for firms that touch sensitive data, that something is reputation insurance.

If you’re still here, here’s the longer version.

What the China AI Argument Actually Says

The short version. Open-source AI in the US doesn’t have a business model that works. When you spend hundreds of millions training a model and then give it away, your competitors host it on cheaper margins than you can. China gets around this because the state subsidizes whoever’s behind. We don’t. So our open-source labs are pulling back (Meta has, mostly) and the open-source models with real momentum right now are coming from Chinese labs.

The longer-run concern Matt’s pointing at is that developer tools, academic research, and a lot of long-tail applications end up built on Chinese open weights. That’s not nothing. Standards get set by whoever ships, and five years from now the AI underneath your tech stack might look very different.

I do take that seriously. It’s just not the question your firm is actually trying to answer this year.

The Three Gaps That Close the Door

Your firm can’t run these models. I don’t mean shouldn’t. I mean can’t, in the literal sense. Almost every law firm and PE firm I work with outsources their IT, which means there’s no internal ML team to even hand the project to. There’s no GPU procurement budget. There aren’t engineers on staff who could fine-tune a model and stand up an inference stack if you asked. Running an open-weights model in production is not downloading a file from Hugging Face and pointing it at your data. It’s serving infrastructure, eval pipelines, monitoring, security review, and a couple of people on payroll who actually know what any of those words mean. The PE firm I sat with that morning has four people in IT total, and most of them are running help desk tickets. The mid-market law firm I trained last month still has partners who can’t get Outlook to sync reliably. The notion that either of them is going to stand up Qwen on a private cloud is a joke. A small handful of firms have the technical bench for it. Most don’t. And the ones that do mostly aren’t bothering, because the next two gaps would still close the door.

Then there’s the trust problem, which doesn’t have a clean answer either. Even if you could run a Chinese-origin model on your own metal with the safety layer stripped off, your GC has to be ready to answer one question: where was this thing trained, and on what. Nobody really knows. I’ve sat in maybe a dozen rooms now where someone walks the partners through “but it’s open source, we control everything,” and I watch the GC’s face. The GC is doing math the rest of the room isn’t. When a client calls and asks where their privileged material got processed, “we pulled the model out of a Hangzhou lab and ran it locally” is not an answer you want to be the first firm to test in a malpractice claim.

And the model itself is maybe 5% of what you’re actually buying when you sign with one of the frontier labs. This is the part of Matt’s argument that I think misses the point hardest. When your firm pays Anthropic or OpenAI or Google, you’re not paying for a model. You’re paying for an enterprise contract, a BAA, IP indemnification, SOC 2 reports, an admin console, audit logs, support, a roadmap, and a vendor relationship someone in procurement can actually manage. You’re also paying for the harness around the model, which by now is most of the value: Claude Code, the agentic stacks, Word and Excel integrations, deep research, document connectors, the workflow tooling your associates already use every day without thinking about it. None of that exists for DeepSeek or Qwen in a form a normal firm can buy, manage, and defend. Even if every other gap got solved tomorrow, your firm would still be writing the check to a US lab, because that’s where the actual product lives. The model alone isn’t a product.

The Premium Is Buying You Something Specific

Here’s the thing nobody actually says out loud when the price comparison comes up. For a firm whose business is privileged client communications, deal documents, sealed pleadings, fund LP data, the cost of saving 60% on inference isn’t 60%. It’s 60% minus whatever a single bad headline costs you, and bad headlines in this business are not cheap.

Picture how it lands. “Law firm processed client M&A docs through Chinese AI model.” It doesn’t really matter at that point whether the model was running on your own private servers, or whether you stripped out every line of safety code. The headline writes itself. The client call comes the next morning. You’d burn through more money on the crisis communications retainer in a week than the inference savings would’ve covered in a year.

So the premium your firm is paying the frontier labs isn’t really an AI premium at all. It’s more like reputation insurance. You’re buying a US contract, US data residency, US legal recourse if something goes wrong, and a vendor name your clients have actually heard of. For firms whose entire product is trust, calling that overhead misses what you’re actually paying for.

I’ve made some version of this argument to maybe two dozen firms in the last six months. Every one of them nods. Then they go back to renewing with the frontier lab they were already using, because once you say it out loud the math is obvious.

What Matt Gets Right About the Long Game

I want to be honest about the parts of this argument that do hold up, because they do.

If a Western company eventually offers an enterprise-grade managed service running on Chinese open weights, the math gets a lot more interesting. Pricing drops. The trust gap narrows because your contract is with a US vendor instead of a Beijing lab. That’s the first signal I’d actually be watching.

The second is the developer base, and this is the piece of Matt’s argument I think holds up best. If the next wave of AI tools (the ones that quietly show up in your tech stack three years from now without anyone deciding to put them there) ends up built on Qwen or DeepSeek under the hood, you’re using Chinese AI whether you chose to or not. That’s not a panic, but it is a planning question, and it’s a real one.

There’s also the chip story, which I’m less sure about. If Chinese labs keep tuning their models to run efficiently on Chinese silicon, and that silicon keeps closing the gap, the cost economics of running anything cheaply could eventually shift in a direction the US doesn’t control. That’s a 24 to 36 month conversation, not a Q2 2026 one.

So I’m not telling you none of this matters. I’m telling you it doesn’t matter to the buying decision in front of you right now.

What to Do Monday Morning

Honestly, none of what I’m about to suggest is exotic. If you’ve been doing the work, most of this is already on your roadmap.

1. Pick your platform and go deeper. Most firms are still using AI like a slightly smarter search bar. The work this quarter (whether you’ve committed to Claude, ChatGPT, or Gemini) is integration: connecting it to your document management, pushing the agentic tools into associate workflows, training partners on the harness rather than just the chat box.

2. Stop changing your AI strategy every time a “China is winning” article hits the partner mailing list. The next one will be along in about a week. You can have an answer ready, which is roughly: infrastructure, trust, and tooling gaps mean Chinese open-source models aren’t in our buying conversation, we’re paying attention, moving on.

3. Watch a few signals that would actually shift the picture. A Western cloud offering managed Qwen with enterprise terms. A US frontier lab releasing open weights that beat the Chinese alternatives. The first credible IP or security incident involving a firm running Chinese open-source in production. None of those have happened yet. When one does, the conversation might be worth reopening.

That’s the list. The harder thing isn’t knowing what to do, it’s getting your firm to actually do any of it.

The Bottom Line

The open-source AI debate is a real debate. It just doesn’t reach into your firm’s buying decisions this year.

The places worth spending your worry are closer to home, and you already know what they are. Governance that hasn’t kept up with how people are actually using the tools. Partners who still haven’t touched any of this. Associates running personal ChatGPT accounts because IT moved too slowly to give them something approved. That’s where the next 12 months are won or lost in firms like yours, and Beijing has nothing to do with any of it.

The open-source AI debate will keep showing up in your inbox, and most of it won't change anything about how you should be running your firm this year. The work in front of you is closer to home: governance that hasn't kept up with how people are actually using these tools, partners who still haven't touched any of it, and the associate running a personal ChatGPT account because IT moved too slowly to give her something approved. The next 12 months get won or lost on those three problems, not on what's coming out of Hangzhou. If you want to talk through how that work is actually getting done at firms like yours, I'm at steve@intelligencebyintent.com. Bookmark the signals worth watching, ignore the rest, and keep going.

Twelve Things I Want from Google at I/O

TL;DR: Rumors of new Gemini models are loud, and I/O is the natural stage. Here’s a practitioner’s wishlist: a real production frontier model, a deeper harness with adjustable thinking, 2M context that actually recalls, parity between Workspace and consumer, real Office integration, an agentic Mac app that can actually do things, sane privacy defaults, glasses that follow your subscription, and a voice mode that actually knows you.

The rumors are getting loud. New models, fresh demos, all timed for the I/O stage three weeks from now.

I use Gemini 3.1 Pro daily in various capacities (almost always in AI Studio with a paid API key and thinking set to ‘high’). So I’ve got a pretty specific list of what I’d want Google to ship. This isn’t a prediction post. It’s a wishlist from one practitioner who uses these tools every day and trains attorneys on them every week.

Here’s what I’m hoping shows up.

A real frontier model, not another preview

First thing I want is a Gemini 3.5, or whatever they call it, that’s actually shipped as a production model. Not a preview. Not a research demo. A general availability model that leads most benchmarks and is ready for production on day one.

Google has a habit of teasing capability before delivering it. The gap between “look what we can do” and “you can use this today” needs to close.

A harness that doesn’t feel watered down

The model is half the equation. The other half is the harness around it. I want deeper tool calling, real agentic behavior, and a feature set that wasn’t engineered for the lowest common denominator consumer.

When I’m building agents, I want the same depth Anthropic and OpenAI give developers. Not a stripped-down version of it.

And I want to dial effort. ChatGPT lets you toggle thinking levels. Gemini does too at the API level, which is how I run it, but bring that control to the consumer app where most people actually live.

Two million tokens with actual recall

Gemini 3.1 Pro tops out at 1M context today. I want to see 2M back at the frontier model tier, with recall that holds up at the back end. I want to drop in a 600-page contract along with exhibits, deposition transcripts, declarations, judgements, and have the model remember what was on page 12 when I ask about page 484 and if there are any inconsistencies with the rest of the content.

Big context with weak retrieval is worse than smaller context with strong retrieval. Get the recall right.

Parity between consumer and Workspace

This is where Google has been the most frustrating.

Memory ships in the consumer app first. NotebookLM is a separate experience. Features land in one place and trickle to the other. I’d like parity. If consumer gets memory and projects, my Workspace account should get them at the same time. Same for NotebookLM integration. Same for plugins, skills, and agent building. Don’t push Workspace customers up to Enterprise just to get capability that’s already free in the consumer app.

While we’re here, fix privacy. If I’m logged into a Workspace account and I run the Gemini CLI, that session should inherit Workspace privacy. Not default to consumer terms. The current setup is a footgun for any law firm or regulated business.

Updated multimedia, on the same release

Imagen, Veo, Lyria. Bring them up to the new underlying model on the same launch day. One creative stack for images, video, and music at the level of the text model, shipped together. Not staggered over six months.

A real Mac app that takes action

Google did ship a native Gemini Mac app two weeks ago. Good first step. Option + Space, screen sharing, the basics.

But it’s observational. It can see your screen and answer questions about it. It can’t drive your Mac. Codex’s April 16 update added Computer Use on macOS, where the agent can click, type, and operate apps on your behalf while you keep working on something else. That’s the bar.

I want the next version of the Gemini Mac app to clear it. Search, Gemini chat, real computer use, and Workspace context in one place. Sorry, Windows users. I’m writing what I’d use.

Microsoft Office integration that’s real

I love Google Docs. But I work with law firms, and law firms run on Microsoft. Word, Outlook, Excel, PowerPoint, SharePoint, Teams. That’s the daily flow for most of my customers.

If Gemini wants to be a serious enterprise contender outside Google’s own customer base, it needs first-class Office integration. Not a third-party plugin. Not a Zapier bridge. Real integration.

A roadmap that explains how the pieces fit

Google has Gemini, NotebookLM, AI Studio, Vertex (which is now also called the Gemini Enterprise Agent Platform), Agentspace (which got folded into Gemini Enterprise at Cloud Next this month), Antigravity, Workspace, Search Labs, AI Mode, Project Mariner, and probably three more I’m forgetting. I want a single page that shows how they fit together, what each one is for, and which one I should pick when.

Right now it feels like a portfolio that hasn’t been edited. Edit it.

Glasses that travel with my subscription

If they ship the glasses form factor at I/O, I want them to know which account I’m logged into and bring that context with them. Personal, Workspace, Enterprise. Whatever I pay for, that’s what the glasses see.

A pair of glasses that only works with consumer Gemini and ignores my Workspace data is a toy. A pair that pulls from my actual work context is a tool.

Voice mode that actually knows me

The current Gemini voice is fine for trivia. It’s not connected to my account memory or my history. I want voice mode that knows what’s in my Drive, what I’ve asked before, and what I’m working on.

If I can text Gemini and have it remember me, I should be able to talk to Gemini and have it remember me too.

What to do Monday morning

If you’ve standardized your firm on Google, the next three weeks will tell you a lot about whether Google is serious about the business side of AI. Three things to do right now:

1. List the workflows where you’d actually use a 2M context window. Have those test cases ready for the day a stronger model lands.

2. Audit your Workspace privacy settings and document what data Gemini is allowed to touch today. Whatever Google ships, you’ll want a clean baseline.

3. If you’ve been waiting on Office integration before standardizing on Gemini, watch the keynote with that lens. The answer is coming or it isn’t.

The short version

Google has the model talent, the data, the distribution, and the cloud. What they’ve been missing is execution discipline and a clear product story.

I/O is the chance to show both. I’m rooting for them.

I'm rooting for Google. They have every ingredient to lead this market, and a serious I/O could change the conversation for any firm trying to decide where to plant its flag. The firms that watch the keynote with a checklist will move faster than the ones that watch it with hope. If you want to compare notes after the announcements, or talk through how to read the signal for your own practice, reach out at steve@intelligencebyintent.com. The next three weeks are a read on Google's seriousness, and your standardization decision should rest on what they ship, not on what they tease.

Codex Catches Cowork. Two Real Agentic Tools, Not One.

You know the feeling. Partner drops a folder on your desk. Five hundred PDFs, give or take. Due Friday. Or it’s 4:47 on a Thursday and a new matter intake just hit the inbox and you’re looking at a conflict check that’s going to eat what’s left of the evening.

This is the kind of work AI is supposed to help with. But the chat window keeps choking on it.

That’s been changing for me over the last few months, first with Claude Cowork and now with ChatGPT Codex. I’ve been running both on the heavy work. Document-heavy review, multi-step analysis, anything that needs to touch hundreds of files or open a browser and sit there clicking for an hour. Cowork has been my workhorse for this kind of thing since the start of the year. The April Codex rebuild puts the two roughly on the same footing for agentic work, with a couple of architectural differences that matter in narrower lanes than the headlines suggest.

TL;DR

OpenAI rebuilt Codex on April 16. A week later they dropped GPT-5.5 into it. Codex now controls your Mac, runs its own browser, works across whole folders of files, and can chain together long tasks that would crash a normal chat window. If your firm is already running Claude Cowork, Codex isn’t a replacement. It’s a second serious agentic tool worth knowing about. The piece Codex has that Cowork doesn’t is a hosted browser inside the app itself, though that matters less for typical law-firm work than the marketing implies. Caveats: consumer ChatGPT is the wrong door for client work, computer use is Mac-only right now, pointing a browser agent at the big legal research vendors runs straight into their terms of service, and some of the new features are gated to specific plan tiers in ways you should check before promising your partners anything. Worth a serious pilot if you’re past the kicking-tires phase on either platform.

What actually changed

Two things landed almost on top of each other.

The April 16 update. OpenAI called it “Codex for (almost) everything,” which is marketing copy but turns out to be pretty accurate. Codex picked up computer use on Mac, where it can see, click, and type into your apps with its own cursor while you keep working in other windows. It got an in-app browser built on the Atlas tech, image generation, more than 90 plugins for things like Gmail and Calendar, and automations that can wake up across days or weeks to keep working on something.

GPT-5.5 dropped a week later. The pitch is that it’s better at multi-step tasks. Planning, using tools, checking its own work, carrying things through to completion without needing you to nudge it back on track every two prompts. Early users describe it as a research partner more than an answer engine. That matches what I’ve seen, mostly.

If you’ve been running Cowork, most of the desktop-agent picture should sound familiar. File access, plugins, multi-step automation, the ability to run for hours on something. Cowork has been doing all of that for a while. Both tools can also drive an external browser through computer use, and Cowork has the Claude in Chrome extension on top of that. So when I say Codex caught up, that’s what I mean. They’re in the same lane now.

The piece Codex has that Cowork doesn’t is its own hosted browser inside the app, where the agent can render pages and you can comment directly on a rendered element to tell it what to do. Useful for specific work, mostly frontend dev and review of pages that don’t need a login. Worth flagging: this in-app browser doesn’t handle authentication, cookies, browser extensions, or your existing tabs. Calling it a “browser” is technically right but the capability is narrower than the word implies. Took me longer than it should have to internalize that.

The bigger shift, taking both tools together, is harder to summarize neatly. AI agents are starting to do work, not just answer questions. A chat window is fine for “what does this clause mean.” Codex and Cowork are built for the next thing over: “go through this folder, tell me which contracts have non-standard indemnification, summarize the variations, and put it in a spreadsheet.”

Where this actually helps a law firm

Four real examples. None are demos. They’re things that take real billable hours and real associate sanity. Most work in either Codex or Cowork. I’ll flag the spots where the differences matter.

Reviewing a folder of 800 documents at once

You drop 800 PDFs into a project folder. You ask the agent to read every one, classify them by document type, pull out key dates and parties, flag anything that mentions a specific issue, and put the results in a spreadsheet with file paths so you can jump back to the source.

In a chat window this hits a wall. Context limits, file upload caps, you end up doing it in twenty rounds. In an agentic tool the agent reads the folder directly, walks through the documents in order, and writes results to a file as it goes. A 500-document review that used to take an associate a week of nights can run while you’re at lunch. You still review what comes back. Always. But the first pass is done by the time you finish your sandwich.

Probably where most firms will land first. The time savings are easy to see and the risk is easy to contain. Either tool handles this well.

Bates numbering without the night shift

The grind work. Stamping every page in a production set, tracking the ranges, building the index, organizing privileged versus produced versus withheld. Either tool can run a script across the folder, apply the numbers in sequence, generate the production log, and put a clean output folder together for transmission.

A 2,000-document set runs in about an hour. Clean log, no skipped numbers. A paralegal still verifies, but the verification is faster than the original work by a wide margin. Firms have been quietly paying overtime to get this kind of thing done for years.

Cite-checking a brief

The one that gets partners’ attention. Either tool can run a cite-check workflow by driving your real Chrome browser through computer use. The agent searches for the cases cited in a brief, checks whether each one exists, was quoted accurately, and is still good law, then builds you a memo with flags for the problems.

Codex’s in-app browser doesn’t help here, by the way, because it can’t handle logged-in pages. So on this specific workflow, Codex and Cowork are essentially even. I had originally assumed Codex would have the edge here. It doesn’t. The bigger question isn’t which agent does the work. It’s whether your subscription terms allow either of them to do it at all.

A real caveat before anyone runs this on Westlaw or Lexis: do not point a browser agent at a paid legal research subscription just because you have login credentials. The published terms I reviewed prohibit automated, robotic, scripted, or software-mediated access without written vendor approval, and they separately restrict scraping, bulk downloading, storage, third-party access, and AI-related use of the materials. The technical capability is there. The contract permission usually is not. Use vendor-approved AI features, licensed APIs, or get written approval from the vendor and your firm before automating anything. For first-pass research, use public or lower-cost legal databases only where their terms actually permit that use.

That paragraph is doing a lot of work. Read it twice.

The intake automation nobody’s demoing

This is the sleeper. Either tool can be set up as an automation that runs every morning. It checks a shared inbox for new matter intake forms, pulls names, parties, and adverse parties out, runs a first-pass search across your firm’s existing client list and document store, drafts a conflict memo, and puts it in your queue with a “review needed” flag.

Twenty to forty minutes per matter. Multiplied by every new matter your firm opens. A 200-matter-per-month firm is looking at a hundred or so hours a month of associate time that doesn’t have to be spent on intake friction. The associate still signs off. That part doesn’t change.

How this changes the work for associates

Some of this is uncomfortable to talk about. The first-year work that has historically been how associates learn to think like lawyers, document review, cite-checking, intake screening, that’s exactly the work agents are getting good at. This isn’t a Codex problem or a Cowork problem. It’s an agentic-AI-in-general problem.

I don’t think the answer is to stop using the tools. The answer is to get intentional about how junior lawyers learn judgment now. If the agent does the first pass, the associate has to do the harder, more interesting second pass. Deciding what’s actually important. Spotting what the agent missed. Explaining the result to the partner. That’s a real skill. It just isn’t the same skill we used to teach by volume.

If you’re a managing partner, this is a training problem you have to solve on purpose. It will not solve itself.

Risks worth raising with your tech committee

The big one is the contract you’re on. Consumer ChatGPT and the Business or Enterprise terms are not the same agreement. Consumer ChatGPT can train on your inputs unless you turn that off. Business and Enterprise have data protections by default. Same logic on the Anthropic side: consumer Claude and the Team or Enterprise Claude contracts are different documents. Whichever platform your firm uses, you need to be on the right contract for client work. Forget the price tier. The contract language is what matters.

There’s a related wrinkle worth surfacing. Some of the new features ship first on the consumer plans and reach the enterprise tiers later. Cowork’s computer use is currently a research preview on Pro and Max only, with Team and Enterprise not yet in. Codex memory and parts of the agentic stack are landing on Enterprise on a later schedule. So the plan that gives you the right contract may not currently be the plan with full feature access. That gap should resolve over time. Right now it means you sometimes can’t pilot on the same plan you’d want to deploy on, which is a less satisfying piece of advice than I’d like to give you.

Capability is uneven across platforms in other ways too. Computer use in Codex is Mac-only right now and rolling out to the EU and UK on a separate timeline. The Codex desktop app runs on Windows, but full computer use isn’t there yet on that side. Cowork has its own platform-specific quirks. Plan around them.

And these agents still make mistakes. They miss things, they occasionally hallucinate confidently, they sometimes generate output that looks right and isn’t. Whether you’re running Codex, Cowork, or both, the agent’s result on a document review is the start of the work, not the end. Build human review into your process and bill it that way.

What to do Monday morning

1. Check what version of ChatGPT and Claude your firm is on. If anyone’s paying out-of-pocket on consumer accounts and using them for client work, that’s the first thing to fix.

2. If you’re already on Cowork, you don’t need to switch. Add Codex to the pilot list as a second tool, mostly because having two real options is healthier than depending on one.

3. Pick one document-heavy task that doesn’t involve client confidential data and run it through one of these tools. Something you’d hand off to a temp without thinking about it. Learn the muscle on the low-stakes version first.

The work didn’t get easier. The math on who does it just changed.

Two real agentic tools is a fundamentally different conversation than one. It means you can pilot, compare, and match the tool to the work without betting the firm on a single vendor's roadmap. If you're sorting out which one belongs where in your practice, or thinking through the contract and associate-training questions that come with either, send me a note at steve@intelligencebyintent.com. The math on who does this work has changed. The decisions about how your firm adapts to that haven't been made yet.

How I’m Using Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro Right Now

TL;DR

Claude Opus 4.7 still owns writing, Word docs, and the memory-based answers that actually feel considered. GPT-5.5 took over my Excel work and beat Claude in a PowerPoint template build I tried in three different Claude surfaces first. Codex has wormed its way into my daily routine, which surprised me. Google is teasing something big. Stack today is roughly 60/30/10 across Claude, GPT-5.5, and Gemini. Ask me again in a month.

Here’s what I’m reaching for, and why.

What Still Belongs to Claude

Writing. Full stop.

GPT-5.5 got better, and I’d put it at a solid 8 out of 10 on prose. But Opus 4.7 still hits a 9-9.5. The difference shows up in voice, in pacing, in the moments where a sentence needs to breathe. I can feel the gap.

Memory is the other big one. When I ask Claude about me, my business, the shape of a project I’ve been working on for months, the answer comes back thoughtful. Not just accurate. Considered. GPT-5.5 has improved here too, but Claude still pulls ahead on the kind of reflection that makes an answer actually useful.

And Word documents are not even close. Out of the box, with no template guidance, the polish I get from Claude is in a different league. Margins, headings, structure, the way a memo reads when I open it. That matters when I’m running a legal demo or sending something to a partner.

Where GPT-5.5 Took Over

Excel is the cleanest example. I’ve moved every spreadsheet task to GPT-5.5. The formulas are tighter, the error checking is better, it follows direction more carefully, and it goes deeper into the work. Claude is still good. GPT-5.5 is great. When the answer needs to add up, I want the tool that adds up best (5.5-Pro!). It’s also amazing at Deep Research.

Then there’s the PowerPoint story.

The PowerPoint Template That Broke Claude

I had Claude Design build me a design spec for a website redesign. It nailed it. I handed that spec to Claude Code and we rebuilt the entire Intelligence by Intent site from it. You can see the result at intelligencebyintent.com.

Then I tried something simple. I asked Claude to take that same design spec and turn it into a matching PowerPoint template. Same colors, same fonts, same look and feel as the website.

Three attempts. Three complete fails.

It would generate slides, sure. But it never touched the actual underlying template. The master slides, the theme, the color palette. None of that got built. I tried in the Claude desktop app. I tried in Claude Cowork. I tried in Claude Code with Opus 4.7 on max thinking. Same result every time. Slides on top of a default theme. Unusable.

Out of frustration, I opened the latest Codex app with GPT-5.5. Same prompt, same spec.

One shot. Exactly what I asked for. Master slides, theme colors, layouts, the whole thing. Done in a single pass.

I came away genuinely impressed. So impressed that Codex has worked its way into my daily routine. I’m extending my memory system with hooks and MCPs to plug into Codex, and I’ll write that up properly soon. Most general users won’t touch Codex, and I get that. But for me, it’s literally become hard to live without over the course of a few days. I prefer it to the Claude desktop app right now. That’s a sentence I didn’t expect to write a month ago.

Where’s Google?

Google is sending signals. They’re hinting at a new model, and from what I can tell, they’re putting real focus on coding. That’s the area where they’ve been weakest. Their design output is already excellent. The raw thinking? Amazing. The raw code, less so. If they close that gap, things get interesting.

I’ll also say what I’ve said before. Gemini in Workspace still feels like a second-class experience compared to consumer Gemini (where’s my NotebookLM integration as the new ‘projects’?). Memory for workspace users? The polish, the pacing, the responsiveness all feel a step behind. I want that to change. I really do because I love this model, but they make it hard to use.

A couple of fresh data points though. Google I/O is around the corner, and they just held their Cloud Next event in Vegas, which I owe you a write-up on. And I just confirmed something new (at least new to me!): you can now generate .xlsx, .docx, and .pptx files directly from Gemini (when did this drop and how did I miss it?). That’s a real shift. Combined with the model rumors, I think we’re about to see something serious from Google. I cannot wait.

Images Update

I was blown away by Nano Banana 2, but now ChatGPT Images 2.0 has become my default. It’s been more consistent than Nano Banana 2 from Google, especially for the kind of professional and brand-adjacent images I generate day to day. Google will catch up. They always do, and I expect a jump in the next month.

What This Means for My Stack

A month ago: 85-90% Claude, 10-15% everything else.

Today: roughly 60% Claude, 30% GPT-5.5 (almost all in Codex), 10% Gemini (almost all in AI Studio with a paid API key and high thinking turned on).

I don’t pick a tool out of loyalty. I pick the tool that gives me the best answer for the work in front of me. As the models change, my stack changes with them.

What to Do This Week

If you’re running a single-tool AI strategy, here are three moves worth making before Friday.

1. Run a head-to-head test on your single most important workflow. Whatever you do most often, push it through both Claude and GPT-5.5 and compare the outputs honestly. Not vibes. Outputs.

2. If you do anything serious in Excel, try GPT-5.5 with extended (or heavy) thinking. Just once.

3. If you’ve never opened Codex, install it and spend an hour on one real task. Not a demo task. Then form an opinion.

The tools are moving fast. The right answer this month may not be the right answer next month. Stay flexible.

More soon. A deep dive on Codex is coming, and a recap of Google Cloud Next is on the way. And if you have a minute, take a look at the new Intelligence by Intent site and let me know what you think.

My stack will probably look different a month from now. I’m okay with that.

The firms making real progress on AI right now aren’t picking a vendor and settling in. They’re running honest tests on their actual work, switching when a better tool shows up, and being willing to be wrong about what they thought six weeks ago. That last part is the hard one. Most partnerships are built to commit to a stack and move on. The cost of doing that is going up, not down.

If you want to compare notes on your stack, or pressure-test the workflow that matters most to your firm, send me a note at steve@intelligencebyintent.com.

Two AI Models. One Credit Agreement. 20 Minutes Each.

TL;DR

I gave a CLE at the Beverly Hills Bar Association this past Wednesday walking attorneys through ten AI-assisted workflows across the transactional arc. The one that got the loudest reaction was the loan covenant analysis. For the CLE session I ran the demo with Claude Opus 4.7 on adaptive thinking. The next day, OpenAI dropped the latest version of ChatGPT (GPT-5.5). I decided to run the same analysis in their new Codex tool with the new GPT-5.5 on xhigh thinking so I could compare results. Each took about twenty minutes. Both produced partner-ready work product. Both caught the day-one default. GPT-5.5 built the deeper spreadsheet. Opus wrote the better Word doc. Different brains for different jobs, and either one beats the three full days a senior associate would have burned on the same task.

The moment every finance partner recognizes

Every finance partner has had this moment. A senior associate hands in a covenant memo. The legal reasoning is solid. The math doesn’t tie. Or the math is right and the defined terms got mangled. Almost no junior lawyer is genuinely strong at both halves of credit work, so partners either do it themselves or run shuttle diplomacy between an associate and a banker until the numbers reconcile.

That’s the bottleneck I wanted to test.

What I showed at the BHBA

The full CLE covered ten transactional workflows across drafting, diligence, and close. Contract redline review on a heavily marked-up MSA. Term sheet to first-draft definitive agreement. Disclosure schedule drafting. SaaS clause libraries. Commercial lease review. Full M&A diligence. Regulatory screening. Closing checklist with dependency mapping. Employment handbook compliance audit. And the one I want to dig into here: loan covenant analysis.

The covenant analysis is the demo that most obviously needs both halves of the brain. Read a 50-page credit agreement, parse defined terms across cross-references, then build the EBITDA bridge and run the math against quarterly financials. It’s the work senior associates either avoid or do badly.

The setup

For the demo I built a sample senior secured credit agreement. $175M Term Loan B, $40M revolver, sponsor-backed industrial manufacturer eighteen months past LBO. Four maintenance covenants. Net leverage stepping down from 5.50x to 4.00x by 2029. Fixed charge coverage of 1.15x. Liquidity floor of $12M. Capex cap of $8.5M. The Adjusted EBITDA definition included all the add-backs that show up in real sponsor-backed deals: restructuring up to $5M, sponsor management fees capped at $2.5M annually, run-rate savings capped at 25% of EBITDA. I built a full year of quarterly financials with messy line items, plus a $5M subordinated seller note with its own tighter senior leverage covenant.

Then I asked both models to produce the same deliverable. Full covenant analysis. Current compliance shown step by step. Headroom by covenant. A 15% revenue stress case. Conflicts with existing debt. Market commentary calibrated to a sponsor-backed mid-market credit. A covenant dashboard. Prioritized markup list with draft language. And a CFO compliance certificate template.

A senior associate would burn three full days on this. Each model finished in about twenty minutes.

What both models surfaced

This is the part that should change how you think about the work.

Both models read the EBITDA definition, walked the add-backs through the caps, and built the bridge from a $4.8M net loss to $29.2M of Adjusted EBITDA. Both pulled Consolidated Total Debt to $186.6M after the $15M cash netting cap. Both calculated net leverage at 6.39x against the 5.50x covenant. Both calculated FCCR at 1.10x against the 1.15x minimum. The deal doesn’t close as drafted. Day-one breach on two covenants, on the actual numbers I gave them.

Both models also flagged what is, in my view, the more dangerous problem: the subordinated seller note’s senior net leverage covenant is set at 4.75x, and the borrower currently sits at 7.72x. Already in default. The new credit agreement’s $2M cross-default threshold catches that note. Unless the seller note is waived or amended at closing and a cross-default carveout is added, the new facility inherits the existing default the day it funds.

Both models flagged the absence of an equity cure right, which is unusual for a sponsor-backed mid-market deal in 2026 and matters because the stress case shows the borrower would need $14M of incremental EBITDA support to stay compliant under a 15% revenue decline. Without a cure, that’s an amendment-and-forbearance conversation, not a sponsor wire transfer.

Both flagged the capex covenant as technically passing but operationally constraining: $7.8M of FY2025 capex against the $8.5M cap is only an 8.2% cushion, and maintenance capex alone is $4.2M. Both noted that the agreement uses an undefined term, “Total Net Funded Indebtedness,” in the incremental debt covenant.

This isn’t pattern matching. This is reading.

Different strengths

GPT-5.5 built the deeper spreadsheet. Thirteen linked tabs. Dashboard, Assumptions, Source Financials, EBITDA Build, Debt and FCCR, Covenants, Stress Model, Sensitivities, Basket Capacity, Reporting, Negotiation, Checks, Sources. Color-coded inputs in blue, formulas in black, cross-sheet links in green. The dashboard pulls live from the covenant tab, which pulls from the EBITDA build, which pulls from the source financials. Drop in a different revenue assumption and the whole stress case rebuilds. The sensitivity tab feeds a chart on the dashboard. This is what a senior banking analyst would build. It’s an actual model, not a report dressed up in spreadsheet form.

Opus 4.7 wrote the better Word doc. The covenant package summary read more lawyerly. The prioritized negotiation list was easier to walk a partner through, and the proposed markup language was tighter and closer to drop-in ready. The cross-checks section caught the same drafting issues but framed them in the order a finance attorney would actually raise them with the agent.

Different brains for different jobs. If I were prepping for the markup call, I’d take Opus’s narrative analysis into the partner meeting and GPT-5.5’s spreadsheet into the call with the financial advisor. The right answer here is to use both models on the same matter and let each do what it does best.

The thing partners ask about first

You should not take any of this output to a client or counterparty without reviewing every number and every defined term. Rule 1.1 competence applies. AI doesn’t have a bar number. You do. The work still needs you. What changes is the first 80%, the part that used to eat three days, now gets done over lunch. You spend your time on the part where judgment actually matters.

Two more things partners ask about. First, plan tier matters. Free and Pro tiers on either platform are consumer-grade and not appropriate for client-confidential work. Use Claude Team or Enterprise. Use ChatGPT Business or Enterprise. If you can’t tell the client where the data sits and who can see it, the data does not go there.

Second, both models will hallucinate a defined-term cross-reference if you don’t give them the actual definitions. I gave both the full definitions section. Skip that step and the output is wrong in ways that look right.

What to do Monday morning

1. Pick a credit agreement currently on your desk and the borrower’s last four quarters of financials.

2. Run the covenant analysis through your firm’s approved AI platform on a Team or Enterprise tier, and check the EBITDA bridge against your own model. That’s where errors hide.

3. Compare the AI’s prioritized markup list to the one you would have built yourself. The differences are blind spots, in either direction.

That last step is the one that will change how you staff your next deal.

The bigger point

A year ago, neither model could have done this. Now both can, and they do it in twenty minutes. One of them quietly built a better financial model than most associates would. The other wrote a better legal memo than most associates would. The cost curve for this work just dropped through the floor.

I have shared the full outputs from GPT-5.5 in a Google Drive folder linked here so you can see what it’s creating these days.

Open the spreadsheet. Imagine the scenario. Then ask yourself how many hours your team billed last quarter on work two machines could do between meetings.

The point of this exercise isn't that AI replaces credit lawyers. It's that the part of the work that used to eat the calendar, reading defined terms, building the EBITDA bridge, running the stress case, has collapsed to twenty minutes per pass. What's left is judgment, and that's the part you were always supposed to be doing anyway. If you want to talk through how this would run on a deal currently sitting on your desk, send me a note at steve@intelligencebyintent.com. The firms figuring this out now will be the ones partners want to be at when the next downturn rewrites every credit agreement in the market.

ChatGPT 5.5 Just Shipped. Here’s What It Actually Changes for Your Law Firm.

TL;DR

GPT-5.5 and GPT-5.5 Pro went live today. The gains for lawyers are in multi-step work, document-heavy tasks, and reasoning that has to hold together across many moving parts. Pro is the variant to test on hard legal work. GPT-5.5 tops the public benchmarks this week but has a documented hallucination problem that is worse than Claude Opus 4.7 or Gemini 3.1 Pro on at least one key benchmark. Test it on real matters. Verify everything. Keep your AI policy current.

Why this release is different

The thing to understand about GPT-5.5 is that the upgrade isn’t really about better answers. It’s about the model not wandering off.

OpenAI’s framing is that GPT-5.5 “carries more of the work itself.” In plain English, you can hand it something with three or four moving parts and it’s more likely to sequence the steps, pick the right sub-task, and check its own output before it hands you a result. That’s the difference between a model that needs a PhD to prompt it and one a senior associate can actually use without making you nervous.

Pro is a genuine step up on the hardest reasoning and research tasks. If you’re paying for it, use it on work that actually needs it. And keep this in mind. GPT-5.4 shipped seven weeks ago. If your firm wrote a three-year AI plan in February, it’s already out of date.

Here’s where I’d actually put this thing to work.

Litigation chronologies stop being a nightmare

I don’t know a litigator who enjoys building a chronology from scratch. Important work. Also awful.

The facts sit across emails, pleadings, texts, medical records, contract files, deposition transcripts, and the client’s memory. Dates don’t line up. The key fact is buried in an attachment. Someone says “late February” when they mean March 3.

That mess is where GPT-5.5’s gains show up.

Don’t ask it to “tell me what happened.” Too vague. Give it a real job. Build the first chronology, cite the source for every entry, flag conflicts between sources, separate hard facts from softer inferences, identify what still needs proof. The lawyer still verifies every entry. But starting with a sourced timeline is a completely different exercise than starting with a blank page. On day 10 of a case that matters, that helps. Before a deposition, even more.

Contract review becomes something closer to business advice

A lot of AI contract demos still feel like parlor tricks. Find the indemnity. Summarize termination. Compare two versions. Fine. Not enough.

Clients don’t need a lawyer to say “there is an indemnity clause.” They need the lawyer to say “this clause makes you responsible for losses you don’t control, and if this deal goes sideways, this is where the money leaks.”

That’s the real work, and GPT-5.5 is better at getting there.

The prompt I’d test isn’t “redline this.” It’s closer to: “Here’s the contract, here’s my client’s preferred position, here’s the deal context. What would make my client regret signing this 12 months from now?”

That’s how deal partners actually think. Assignment, change of control, data rights, payment timing, limitation of liability, audit rights, termination, exclusivity. Legal provisions, sure, but also business choices. The model can help a senior associate surface those choices faster. The partner still decides what to fight over.

Make the model attack you instead of help you

Here’s where I’d use GPT-5.5 Pro. And I wouldn’t use it as a case finder, because that’s still how lawyers end up on the wrong end of a sanctions order.

Use it to stress-test the argument instead.

Most lawyers prompt research the wrong way. “Find cases that support my position.” Understandable. Also how you miss the problem. The better move is to hand GPT-5.5 Pro the issue, the jurisdiction, the facts, and the position you want to take, then tell it to rip the argument apart. Where does this break? What facts would make a judge uncomfortable? What is the other side’s cleanest counter? Which precedent would you hate to see in their brief?

That’s not a research shortcut. It’s a preparation shortcut. Good partners do this instinctively. Junior lawyers take years to build the reflex.

Caveat, because it has to be said: GPT-5.5 does not eliminate the citation problem. Every case it surfaces still needs human verification. Every pin cite. Every quote. We’ll get to that.

Client alerts can become client service

Every firm sends alerts. Most clients ignore them.

Not because the alerts are badly written. Most are fine. The problem is they’re broad by design. They explain what changed. They don’t tell a specific client what to do about it.

Take a regulatory change, an enforcement action, or new agency guidance. Feed the model the update plus a short description of your affected client types. Ask it to separate the generic news from the practical implication. Who should care, why now, what they should review this week, what question the relationship partner should ask on the next call.

A privacy update hits a healthcare company, a SaaS vendor, a retailer, and a PE portfolio company differently. A labor rule hits a 30-person startup and a 12,000-person employer operating in nine states very differently. The old version of this work is “send the alert.” The better version is “turn one alert into 12 client-specific partner notes.” Twenty minutes now, versus three hours.

Clients remember the second kind.

The one most managing partners will sleep on

If AI only helps your lawyers draft faster, fine, that’s useful. If it helps the managing partner see where profit is actually leaking, that’s a different conversation.

Law firms sit on years of useful data. Time entries. Write-downs and write-offs. Staffing patterns. Matter budgets. Realization rates. Collection delays. AFAs that worked, and the ones that quietly bled money. Most firms don’t really learn from it. They feel the pattern. They don’t study it.

5.5 is meaningfully better at working across spreadsheets and documents at the same time. Hand it your closed matters from the last six months and ask where margin leaked. Look for write-off patterns by partner, matter type, office, or phase of the engagement. Compare fixed-fee matters that worked against fixed-fee matters that didn’t, and ask what the actual difference was.

This use case doesn’t get applause at a CLE. It changes how a firm prices work and how partners decide which matters to take. That’s strategy, not software.

What actually worries me

Last summer, lawyers at a well-regarded firm ended up in serious trouble in Johnson v. Dunn after hallucinated AI citations made their way into federal filings. The firm avoided the worst of it largely because it could show real internal AI governance: warnings, escalation, and a written AI policy already in place. That was with an earlier model.

And this is not just a mid-market or one-off problem: this week Sullivan & Cromwell apologized to a federal bankruptcy judge after a filing reportedly contained AI-generated inaccuracies, including fabricated citations and misstatements of law, and acknowledged that its internal AI policies and secondary review processes were not followed.

GPT-5.5 is more confident when it’s wrong than its competitors are, and the gap isn’t small.

On AA-Omniscience, a benchmark specifically designed to measure what a model does when it is outside its knowledge, GPT-5.5 posts the highest accuracy of any model tested at 57 percent. That is the good news. The bad news is that its hallucination rate on the same benchmark is 86 percent. Claude Opus 4.7 is at 36. Gemini 3.1 Pro is at 50.

What that means in practice is simple: GPT-5.5 may know more, reason better, and still be more willing to make something up when it should say “I don’t know.” For legal research, that is not a footnote. That is the caveat.

That is a specific, serious problem for legal research.

If your lawyers are running cite checks through GPT-5.5 without hand-verifying every citation, you are closer to a sanctions order than you were yesterday. My personal view, and I realize not everyone agrees: I wouldn’t use any of these general-purpose models for pure case-finding right now. Use a real legal research platform for that. Use GPT-5.5 Pro to pressure-test what you find.

What to do Monday

Three moves.

1. Pick three associate workflows you run every week. Test GPT-5.5 against your current process. Time both. Score the output. Get real numbers before the next marketing cycle hits.

2. Run the same tests on Claude Opus 4.7. OpenAI tops the leaderboard this week. Anthropic is ahead on grounding. Your practice mix decides which one wins for you.

3. Run the firm management test. Feed GPT-5.5 your last six months of closed matters and ask where margin leaked. See if it finds something you missed. If it does, that one test paid for a year of Pro subscriptions for the partnership.

That’s the test plan. Not a strategy deck.

The model race isn’t slowing down. Your real edge isn’t picking the winner every eight weeks. It’s running disciplined tests on real matters and keeping your AI policy sharp enough that no one in your firm ends up on the wrong side of a sanctions order.

If you’re building that test plan right now and want a second set of eyes on it, reach me at steve@intelligencebyintent.com.

ChatGPT Images 2.0 Is a Real Tool for Lawyers Now

TL;DR: OpenAI shipped ChatGPT Images 2.0 yesterday. The new model, ChatGPT Images 2.0, follows prompts more precisely, renders readable text inside images, and can generate up to eight consistent visuals from a single prompt. Paid tiers get an “images with thinking” mode that plans and verifies before it draws. For the first time, I can build a timeline, a storyboard, or an entity chart in one shot and hand it to a partner without rebuilding it in Illustrator. Five real examples below.

The moment I stopped fighting it

I’ve been trying to use AI image tools in actual legal work for about two years. Every time, the same thing happens. The picture looks great from six feet away. Then you zoom in and the labels say “Plainitff” and “Defendnat.” The timeline reads left to right but the arrows point the wrong way. The entity boxes overlap. You send it back. It gets worse.

So you open Visio. Or PowerPoint. Or you call the graphics team and wait three days.

That’s the part that changed yesterday.

What OpenAI actually shipped

ChatGPT Images 2.0 went live on April 21, 2026, across ChatGPT, Codex, and the API. The model is officially called gpt-image-2 (the internal model name), and it replaces DALL-E 3, which gets retired on May 12. Output goes up to 2,000 pixels wide, aspect ratios run from 3:1 to 1:3 (finally, real infographic dimensions), and the model can produce up to eight consistent images from one prompt. You can get 2K images in the web interface and 4k images in the API.

The big shift is reasoning. Paid plans get an “images with thinking” mode where the model plans the layout, checks its work, and can pull facts from the web mid-generation. That sounds like marketing until you watch it catch its own spelling error on a demonstrative before it finishes rendering.

Why this matters for law firms

Here’s the thing about visuals in legal work. They either do real work or they’re filler. A good trial demonstrative can carry a jury through facts that a 40-page brief cannot. A bad one confuses everyone and costs you credibility. Most firms sit somewhere in between because the cost of making good ones is high and the time pressure is higher.

ChatGPT Images 2.0 changes the math on that tradeoff. Not every visual needs to be Exhibit A quality. A lot of them just need to be clean, readable, and fast. That’s where this model earns its keep.

Let me show you what I mean.

Five things I tried yesterday

1. An infographic

Prompt: “Go research the Beverly Hills Bar Association (BHBA). Create a compelling infographic that they can use in marketing materials that highlights the value the bring, the breadth of content, the capabilities, reasons to join, and more.”

2. An accident storyboard

Prompt: “Create a 4-panel litigation demonstrative storyboard for a fictional rear-end collision at a rain-slick intersection, clearly labeled as an illustration and not evidence. Panel 1: wide overhead view of the intersection with street names and signal lights; Panel 2: driver looking down at a phone as the light turns red; Panel 3: impact sequence with motion arrows and skid marks, non-graphic and neutral; Panel 4: immediate aftermath with police arrival and vehicle positions. Keep the same cars, weather, lighting, and street geometry across all panels. Add small captions, time stamps, directional arrows, and a footer that says “Illustrative reconstruction based on alleged facts.” Style: realistic courtroom demonstrative, sober and precise, not sensational. Photorealistic.”

3. An entity control chart

Prompt: “Create a beautifully clear entity and control chart for a fictional private-company dispute. Parent company: Harbor Peak Holdings LLC. Subsidiaries: Bluestone Logistics Inc. (62%), Cedar Analytics LLC (100%), Meridian Freight Systems Inc. (51%). Individual owners of Harbor Peak: Laura Chen 35%, Omar Reyes 30%, Atlas Ventures LP 25%, management option pool 10%. Show ownership percentages, control arrows, board-seat icons, and a right-side legend explaining voting control vs. economic interest. Use a premium M&A-slide aesthetic: white background, crisp small text, thin connector lines, subtle slate and teal accents, beautifully organized boxes, zero clutter, fully presentation-ready.”

4. A trade secret dispute chronology

Prompt: “Create a premium litigation chronology board for a fictional trade secret dispute between AeroShield Labs and North Ridge Biotech. Facts to visualize: Jan 6, 2026 CTO Dana Kim resigns; Jan 18 joins rival; Feb 2 downloads 1,842 files the night before departure; Feb 14 rival announces a nearly identical product; Mar 1 forensic report confirms USB transfers; Mar 7 cease-and-desist sent; Mar 18 complaint filed. Design a clean 16:9 legal demonstrative with a left-to-right timeline, event cards, short captions, elegant icons, three key-player portraits or silhouettes, and three exhibit-style pull quotes. Style: elite law-firm mediation board, restrained navy/gray palette, crisp readable typography, white background, serious and polished. Add footer text: “Illustrative demonstrative for discussion purposes only.” Do not invent extra facts.”

5. Internal training material

Prompt: “Create a one-page, 6-panel editorial training comic for lawyers titled “Two Prompts, Two Very Different Risk Profiles.” Storyline: Panel 1 associate pastes sensitive client facts into a personal consumer AI account at home; Panel 2 partner notices and looks alarmed; Panel 3 freeze-frame with callouts: privilege, confidentiality, retention, terms of service; Panel 4 the same associate uses the firm-approved enterprise workflow instead; Panel 5 secure review process with redacted facts, policy checklist, and approval steps; Panel 6 ending panel with the caption “Convenience is not the same thing as governance.” Style: sophisticated editorial comic, law-firm setting, readable captions, consistent characters, modern and serious with a light touch of wit, no goofy cartoon energy.”

The tradeoffs you need to know about

This isn’t a replacement for a trial graphics team. You still need one of those for courtroom-quality work that has to survive cross. What this replaces is the thirty minutes you used to spend in PowerPoint making a mediocre version yourself, or the $800 you paid a freelance designer to turn a rough sketch into something presentable.

A few things to watch. The model still hallucinates. It will invent a statute number or a date if you aren’t specific, and it will misspell a party name on roughly one out of every few generations. Proofread every word that lands inside an image.

But the bigger risk is the thinking mode that pulls from the web during generation. Great for accuracy on generic diagrams. Bad for confidentiality. Do not hand it privileged facts and ask it to go look things up. Use standard mode for anything client-related, or use an enterprise tier with the right data protections in place.

Third tradeoff worth flagging: cost. API pricing runs about $0.21 for a standard 1024×1024 image. Thinking mode adds reasoning tokens on top, which can push a layout-heavy visual meaningfully higher. For most firms that’s still cheaper than a designer, but it’s not free. Budget accordingly.

What to do Monday morning

If you’re a managing partner or a practice group leader, here’s a short list of moves that matter this week.

1. Try it yourself on a real piece of work. Pick one demonstrative or training slide you’d normally outsource and rebuild it in ChatGPT. Twenty minutes, tops.

2. Update your AI usage guidance. If your policy says “no image generation for client matters,” it’s time for a revision. The right guidance is probably about confidentiality mode, not a blanket ban.

3. Flag it for your CLE and training team. Internal education materials are the lowest-risk, highest-value use case. Start there.

4. Put it in front of your trial graphics team, not around them. They’ll know within an hour what it can and can’t do, and they’ll use it to draft faster.

The part that sticks

The old complaint about AI image tools was that they made pretty pictures for people who didn’t need pictures. This one makes useful pictures for people who do.

That’s a different product.

For years, AI image tools have been a sidebar for legal work. Cute, sometimes clever, rarely usable. What shipped yesterday is the first version I'd put in front of a partner without apologizing for it, and the first one that changes the real math on how firms build visuals. If your team is thinking about where this fits, what to ban, and what to start piloting this quarter, I'm happy to talk it through. Reach me at steve@intelligencebyintent.com. The firms that figure out the right rails for this in the next ninety days will spend the rest of the year pulling ahead of the ones that don't.

Claude Design Just Launched. Here’s What It Means for Law Firms.

TL;DR: Claude Design isn’t going to replace your trial graphics firm, and that’s not the interesting question. The interesting question is what it does to the enormous middle of visual work at firms that currently sits in PowerPoint purgatory, gets redrawn five times, and never quite looks like it came from the same place.

Most of what comes out of a law firm visually is either very good or very bad.

The very good stuff gets outsourced to trial graphics specialists who charge $8,000 to $25,000 for a single animation and take weeks to deliver. The very bad stuff gets made in PowerPoint at 10 PM by a paralegal who didn’t sign up for this. The interesting question about Claude Design isn’t whether it can beat the specialists. It can’t, at least not yet, and that’s not the point.

The interesting question is what happens to the giant middle. And there’s a lot of middle.

Timelines. Chain-of-command charts. Statutory element breakdowns. Deposition excerpt callouts. Damages layouts for mediation. The internal case narrative deck that gets passed between partners six times before anyone figures out the theory. Pitch decks. CLE slides. One-pagers. The marketing deck your practice group has been meaning to update since 2023. All of it visual work. None of it visual enough to justify a vendor.

Here’s the short version. Claude Design launched April 17, powered by Claude Opus 4.7, Anthropic’s newest model. It’s included with Pro, Max, Team, and Enterprise plans at no additional cost. You describe what you want, it asks a few questions, it builds a first version (or several, if you ask for variations), you refine through conversation or direct edits or these custom sliders it generates for each design that control things like spacing, density, color warmth. Export to PDF, PPTX, HTML, or push to Canva for final polish.

Anthropic is pitching it at founders, product managers, designers working under deadline, and marketers. Not at lawyers. Which is kind of the point.

The thing worth noticing

Most AI tools give you a blank box and hope your prompt is good. Claude Design asks questions first. What’s the main role here? Who’s the audience? How many variations? For each question it gives you five or six possible answers rather than a blank box.

That’s not a design interaction. That’s a product manager interaction. If you’ve ever briefed a good associate on a deck and watched them come back with three clarifying questions before touching the template, you know this feeling.

I think that’s the actual story. Not the sliders. Not the Canva export. Not the fact that Figma’s stock took a hit the day it launched, though it did, and Mike Krieger resigned from Figma’s board three days before, so draw your own conclusions. The story is that this thing behaves more like a thinking partner than a tool.

Greg Isenberg ran an unscripted test right after launch and gave it 9/10 for wireframing and 8.7/10 for deck research and design. The video part is still weak (4.5/10). Everything else is closer than most people expected.

Where this actually lands for firms

Claude Design is not going to make your medical animation for trial. It’s not going to produce the exhibits you’d put in front of a jury in a bet-the-company case. The specialists are still the specialists and should stay that way.

What it does is the middle. The timeline that doesn’t need to be beautiful but does need to be right. The damages chart for mediation that currently gets built three times because nobody can decide what to include. The case narrative deck that gets used internally first, then at the settlement conference, then never. And all the non-litigation work. Pitch decks for client pursuits. CLE slides. Client reports. Practice group one-pagers. Internal training.

Two things worth trying this week.

The first is a case narrative deck. Pick an active matter. Feed Claude Design the operative complaint, a rough theory of the case, a few key dates. Ask for twelve slides that tell the story. Use it internally first. Use it in the next mediation. You’re not replacing a demonstrative specialist. You’re replacing the third draft of the deck nobody wanted to make.

The second is a damages visualizer. Future medical costs, lost earnings, life care plan. The hard part with numbers is always showing how they add up, and adjusters and mediators respond to clean visual logic. Hand it a spreadsheet and ask for five different framings. Pick the one that lands. If it saves you one mediation cycle, the tool paid for itself.

A side note while I’m here. If your firm has a marketing team, this changes what they can do at the same headcount. If your firm doesn’t have a marketing team, you just got one.

The one caveat that matters

Confidentiality. This is where attorneys need to get the tier structure right. Free, Pro, and Max are consumer terms. They can train on your inputs depending on your opt-in settings, and retention runs up to five years if training is on. Team and Enterprise are commercial terms. Training is contractually prohibited, not a toggle you have to remember to flip. For client data, you want Team or Enterprise, not Pro or Max. This isn’t new and it isn’t specific to Claude Design. It applies to every AI tool your firm touches. If your associates are using personal ChatGPT accounts for client work, that’s a much bigger problem than which design tool you pick.

Everything else is normal first-draft discipline. Review the dates. Check the numbers. Assume the output is wrong until you’ve verified it. Treat it like a second-year associate’s draft.

Monday

Confirm your firm is on a commercial Claude plan. Pick one matter and one deliverable, nothing that touches a client yet. In Claude Design’s organization settings, set up a design system by uploading a brand guide, a template deck, or a few pages from your website. Every project after that will inherit your firm’s look. Tell one person they’re the reviewer before anything ships.

Then go look at your last six months of external design spend and figure out what you’d bring in-house.

That last one is the one nobody will actually do. Which is why it’s the one that matters.

The point isn't the tool. It's the middle. Every firm has a stack of visual work that sits between "not important enough to outsource" and "too important to ignore," and for the first time that stack has a credible in-house answer. If you want to talk through what this looks like at your firm, or what to try first without touching client data, reach out: steve@intelligencebyintent.com. The firms that figure this out in 2026 won't be the ones with the biggest design budgets. They'll be the ones who stopped paying for work they could already do themselves.

Every Lawyer Should Have These Five Skills (The Claude Kind)

TL;DR: Skills are small instruction files that load automatically when Claude sees a task that matches. They beat Projects and custom GPTs for most lawyer work. Here are five to build first and how to actually do it.

The question I keep getting after trainings

Every training I run, some version of this happens. A partner pulls me aside at the break and says: “I built a Project in Claude. Uploaded the templates. Wrote the instructions. Added all the examples. I used it twice. Then I stopped.”

I hear this a lot. It’s not a Claude problem. It’s a container problem.

Projects and custom GPTs were the right answer eighteen months ago. They are not the right answer anymore. And the better option is sitting one setting away.

It’s called a Skill. Almost none of my clients have heard of them.

What a Skill actually is

A Skill is a small folder with a file called SKILL.md inside it. The file tells Claude: when you see a task that looks like X, load these instructions. Claude decides when to pull it in. You don’t have to remember. You don’t have to be in a specific Project. You don’t paste the same 500-word prompt at the start of every chat.

A Project is a room you have to walk into. A Skill is a reflex that fires anywhere in the building.

Anthropic ships a few Skills by default (Word, Excel, PowerPoint, PDF). You build your own for everything else. They work in Claude.ai, in Claude Code, and through the API.

Why this beats what you were doing

The big thing is that Skills travel. A Project locks you into one workspace. A Skill fires anywhere you’re working in Claude, even in a fresh conversation with zero setup. That alone changes the math on whether it’s worth building one.

They also stack. If you have a firm style Skill and a contract review Skill, Claude uses both at once when you paste in a draft agreement. Projects can’t do that. You pick one container and live in it.

And when you need to change something? Edit the file once. Everyone using it gets the new version the next time they ask Claude for help. Nobody has to remember to rebuild anything or re-upload their templates. If you have ever tried to get fifty attorneys to update the same Word template, you understand why this matters.

One more thing worth knowing. Skills are built to an open standard, so the same file should transfer to other platforms that adopt it. I wouldn’t bet the firm on that yet. It’s just a reason not to obsess over lock-in.

The five that every lawyer should have

Firm voice and style

This is the one I’d build first for any firm that writes a lot. It captures how your firm actually writes. Brief structure. Memo headings. Citation format. Caption layout. The house rules nobody writes down: no passive voice in the intro, no “clearly,” no rhetorical questions in motions, no two-sentence paragraphs when the client is reading.

Every associate eventually turns in something that sounds off, and the senior associate or partner rewrites it. That rewrite is the expensive part.

I built one of these for myself, for my Substack writing. Took me a weekend to get close, and I’m still editing it most weeks. Which is kind of the point.

Contract review playbook

Every firm has positions. Indemnification. Limitation of liability. IP assignment. Governing law. Termination. Confidentiality. Most of those positions live in a few partners’ heads and in a Word doc nobody can find.

A contract Skill captures the firm’s positions plus preferred alternate language. Paste in a draft agreement. Claude produces a first-pass memo with the firm’s positions, flagged risks, and suggested revisions. It is not a finished redline. It is a starting point that saves an associate an hour per agreement and gets every junior attorney pointed at the same North Star. The partner who reviews the work stops having to teach the same five things every week.

Billing narratives

This one pays for itself in two weeks. Maybe less.

Most firms have clients with billing guidelines that nobody reads until something gets kicked back. No block billing. ABA codes. Specific formatting. Banned phrases like “reviewed documents” and “worked on case.” Attorneys break the rules, usually at 11 PM when they’re entering time from a phone. Billing catches some. The client catches the rest and writes off the time.

A billing narrative Skill takes raw time entries and reformats them for each client’s rules. Applies the ABA codes. Flags block-billed entries. Rewrites the lazy descriptions. “Worked on case” becomes “drafted opposition to motion to compel; researched authority re: work-product doctrine.”

If your firm writes off even five percent of fees to guideline violations, nothing else on this list comes close on ROI. And most firms are writing off more than five percent. They just haven’t measured it.

Plain-English client communications

Your general counsel and business clients keep telling you the same thing. Your emails read like briefs. They want the answer, the risk, and the recommended action. Short. In a language they can forward to the CEO without a translator.

This Skill takes a piece of legal analysis and rewrites it for a non-lawyer executive. Short summary up top. Risk rated plainly. Recommended action. No hedging paragraphs that sound like malpractice insurance talking.

Pair it with your firm style Skill and Claude produces two versions of every communication at once: the careful memo for the file, and the short email for the client.

Depo and discovery summaries

Dep summaries and document review are where associate hours go to die. Every firm does them a little differently. Every associate does them a little differently again.

A summary Skill defines the output format once and enforces it. Issue tags. Witness. Date. Page-line cites. Admissions flagged, contradictions with prior testimony marked. When the case moves toward trial, everything is already in the format you need for the cross outline. No rework.

What you’re giving up

A couple of honest tradeoffs.

Skills work best small and single-purpose. One for contract review. A different one for billing. A third for client emails. If you try to build one giant “everything for my practice group” Skill, it will do everything badly. This is the biggest mindset shift for people used to one-big-Project thinking.

On individual plans (Free, Pro, and Max), Skills are per-user. Team and Enterprise admins can push them out centrally. If your firm is on individual plans, each attorney uploads the file themselves. The files are portable, so it’s not a crisis. But someone needs to own the master version, or you end up with six slightly different contract Skills drifting around the firm.

How skill-creator actually goes

The official instructions are clean. The real process is messier. I have built a dozen or so of these at this point, for myself and for clients, and here is what actually happens.

First, make sure code execution is enabled at Settings > Capabilities. Then go to Customize > Skills and turn on skill-creator. Start a new conversation and say, “Help me create a new skill.”

It starts asking questions. What does the Skill do. When should it trigger. What inputs will come in. What should the output look like. This is where most people rush and regret it. Give it real examples. Three or four actual outputs from your practice, redacted if you need to. Not summaries of examples. Actual examples. The quality of your examples sets the ceiling for the quality of the Skill. Generic examples produce generic Skills.

After the interview, it drafts a SKILL.md file. Read it carefully. It will describe how you say you practice, which is not always how you actually practice. Edit the description. Edit the instructions. Get it closer to reality.

Then you test it on three live matters. You will find that it misfires. Either it doesn’t load when it should, or it loads and does something slightly wrong. Under-triggering is the most common issue, and Anthropic’s own team flags this in their docs. The fix is usually a more specific description of when the Skill should activate, plus a nudge in your prompt the first few times (”use my contract-review Skill on this”).

When you're happy with it, zip the folder up, go to Customize > Skills, hit the plus button, and upload the ZIP. Yes, it has to be a ZIP. The skill folder itself won't work. If you're on Team or Enterprise, an admin can push it out to everyone from organization settings, and you stop playing version-control whack-a-mole across fifty attorneys.

An hour in, you have something working. A week in, you have something good. A month in, after you have run it against thirty real matters and edited it six times, you have something you will not want to work without.

Monday morning

Pick the Skill that hurts the most right now. If billing write-offs are bleeding the firm, start with narratives. If associate drafts are always off-key, start with style. If you redline five agreements a week and explain the same thing every time, start with the playbook.

Spend an hour. Test it on three real examples. If it saves you thirty minutes a week, it already paid for itself.

The attorneys who pull ahead over the next twelve months aren’t the ones using Claude the most. They’re the ones teaching Claude how their firm works, once, and letting it remember.

The firms that pull ahead over the next twelve months won't be the ones using Claude the most. They'll be the ones that sat down once, wrote down how their firm actually works, and let Claude remember it on every matter after. Five small files. A weekend of real work. A practice that runs a little sharper every week after that. If you want a second set of eyes on a Skill you're building, or help picking which one to start with at your firm, email me at steve@intelligencebyintent.com. The attorneys who figure this out first will be very hard to catch.

Claude Now Lives Inside Word, Excel, and PowerPoint. Here’s Why That Changes Everything for Law Firms.

TL;DR: Anthropic’s Claude add-ins for Microsoft Office now share context across Word, Excel, and PowerPoint in real time. Your spreadsheet data, your narrative documents, and your presentations can all talk to each other through a single AI conversation. For law firms, this kills the most tedious part of complex deals: the endless copy-paste-reformat cycle. Here’s what it looks like in practice.

The Problem You Already Know

You’ve lived this moment. It’s 9 PM, you’re finalizing a client deliverable, and someone on your team updates a key number in a spreadsheet. Now you need to hunt down every place that number appears in the Word doc. And the PowerPoint deck. And you need to make sure the formatting doesn’t break when you paste it in.

Nobody went to law school for this. But everyone in a law firm does it. Partners billing $800 an hour do it. Associates spend half their evenings doing it. Paralegals have built entire careers around being really, really good at it.

That’s the problem Claude’s Office add-ins just solved. Or at least took a serious swing at.

What Actually Changed

As of this week, Anthropic completed the set. Claude now runs natively inside Word, Excel, and PowerPoint through official Microsoft add-ins.

But here’s the thing that matters: it’s not three separate AI tools. They share context.

If you’re working in Excel and ask Claude to analyze a rent escalation schedule, then switch over to Word where you’re drafting a lease, Claude already knows what’s in that spreadsheet. No CSV export. No pasting a table. No re-explaining. The conversation carries across all three apps.

The shared context feature shipped for Excel and PowerPoint back in March. Word arrived on April 13. Now the full loop is closed.

I keep coming back to this framing: instead of three apps that happen to have AI bolted on, you’ve got one AI that can see across your entire document workflow at once. That’s a fundamentally different thing.

Why Law Firms Should Care About This More Than Anyone Else

Law firms run on documents. I don’t mean that metaphorically. Documents are literally the product. And the documents almost never live in just one format.

The data sits in spreadsheets. The analysis goes into Word. The client presentation lands in PowerPoint. Every deal, every matter of any complexity, touches all three.

So the old workflow goes like this: an associate manually transfers numbers from Excel into a Word table, reformats everything, prays they didn’t transpose a digit, and then does it all over again when the numbers change two hours later. A paralegal rebuilds a slide deck by hand every time the deal terms shift. On a complex transaction with dozens of iterations, you’re burning hours per week per person on work that adds zero analytical value.

That’s what shared context eliminates. Claude reads data from your open files and writes it directly into the document or deck. One conversation, three applications, no copy-paste.

Where This Actually Shows Up in Practice

Commercial Leasing: Two People, One Living Document

This is the scenario that sold me on the concept. Picture a commercial real estate deal. The partner is drafting a 40-page lease in Word. Down the hall (or across the country), the financial analyst is running rent escalation models, CAM reconciliation projections, and TI amortization schedules in Excel.

Every time the analyst updates a number, the partner needs it in the lease exhibits. In the old world, that’s an email. Or a Slack message. Or the analyst walks over with a printout. Then the partner manually updates Exhibit B, double-checks the math, and hopes nothing got lost in translation.

With shared context, the partner just asks Claude to pull the latest Year 3 base rent and escalation percentages from the analyst’s spreadsheet and update the exhibit. Done. And when the client’s board wants a summary deck? Open PowerPoint. Claude already has the lease terms from Word and the projections from Excel. The bones of the deck assemble from the data in your open files.

I realize that sounds almost too convenient. But that’s literally what the shared context feature does.

Building a Board Deck Without Losing Your Mind

Different scenario, same pain. An in-house corporate lawyer needs to prep the quarterly board deck. The raw material is scattered across three places: financial data in Excel, a management narrative in a Word memo, and last quarter’s branded PowerPoint template that absolutely cannot have its formatting broken because the CEO’s EA will send it back.

This used to be an afternoon. Maybe longer. Pulling numbers into slides, retyping key points from the memo, making sure the fonts match, aligning text boxes that refuse to cooperate.

Now the lawyer opens all three files, tells Claude to build the Q2 deck from the financial summary and strategic highlights, using the existing slide master. Claude reads the spreadsheet for numbers, pulls narrative from Word, and generates slides in the company’s fonts and colors. The lawyer reviews and refines. Forty-five minutes instead of four hours, and most of that time is actual thinking, not formatting.

M&A Due Diligence: The One That Gets Interesting

This is where I think the real leverage is. (Not leverage in the banned-buzzword sense. Actual leverage.)

A mid-market M&A deal generates a staggering amount of information spread across formats. The deal team tracks findings in an Excel workbook with tabs for financial metrics, contract summaries, risk flags, and compliance items. Associates draft sections of the due diligence report in Word. Partners need a summary deck for the client’s investment committee.

In every deal I’ve seen, the most junior person on the team spends days reconciling the tracker with the report. Making sure every flagged issue in the spreadsheet shows up in the right section of the narrative, with the right context, and nothing falls through the cracks. Then they rebuild key findings into slides. It’s meticulous, important, and completely manual.

Here’s where shared context changes the math. The senior associate can ask Claude to review the diligence tracker, cross-reference it against the draft report, and flag gaps. Then build a 10-slide summary from both files. Claude handles the cross-referencing and formatting. The associate handles the judgment calls: Is this finding material? Does this risk need escalation? That’s the work that actually requires a law degree.

Quarterly Compliance: Same Report, New Numbers, Every 90 Days

I almost didn’t include this one because it’s less dramatic than the M&A example. But honestly, it might be the highest-ROI use case for a lot of organizations, precisely because it’s so repetitive.

A compliance team tracks regulatory metrics in Excel: capital ratios, transaction volumes, SAR counts, training completion rates. The compliance officer writes the narrative portions of quarterly filings in Word. The general counsel presents a compliance summary to the board in PowerPoint.

Same structure every quarter. Different numbers. And every quarter, someone spends two days updating the narrative report with fresh figures, cross-checking them against the tracker, and rebuilding slides.

This is almost criminally well-suited for shared context. Update the Excel tracker, have Claude refresh the Word filing with current numbers, then refresh the board slides and flag any metric that moved more than 10% from last quarter. The repetitive mechanics are handled. The compliance officer focuses on analysis and exceptions, which is what you’re actually paying them for.

The Honest Tradeoffs

I want to be fair here.

The Word add-in launched days ago. This is beta. There will be rough edges, and you should not use this on final client documents without careful human review. That said, Anthropic made a smart design choice: when you use Claude’s suggested edits mode in Word, every edit shows up as a tracked change. That forces someone to accept or reject each revision individually. Built-in guardrails.

The data question matters too. Shared context means your information is moving between applications through Claude’s conversation. For firms with sensitive client data, you need to understand the data flow. The add-ins support enterprise LLM gateways through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure, so firms can route traffic through their existing compliance infrastructure. That’s the right approach. But it means your IT team needs to be in the room for this conversation, not just your practice groups.

And one more thing that should be obvious but I’ll say it anyway: Claude doesn’t replace judgment. It can pull numbers, format tables, and cross-reference data all day long. It cannot decide whether a lease term is market-rate, whether a diligence finding is material, or whether a compliance metric warrants board-level escalation. The value is in clearing away the mechanical work so your people can focus on exactly those decisions.

What to Do Monday Morning

1. Install the add-ins from Microsoft AppSource. Search “Claude by Anthropic” under Insert > Get Add-ins in any Office app. A quick note on availability: Claude for Excel is on all paid plans (Pro, Max, Team, Enterprise). PowerPoint requires Max, Team, or Enterprise. Word is currently limited to Team and Enterprise only.

2. Test on internal work first. Build a memo from a spreadsheet, refresh a recurring deck, something where a mistake costs you time but not a client relationship.

3. Get your IT and security teams involved before you go firm-wide. They need to evaluate the data flow, especially if you’re running an enterprise LLM gateway.

The Bigger Picture

For years, the AI conversation in law firms has been about generating text. Write me a first draft. Summarize this contract. That’s useful, but it’s one-dimensional.

What the Office integration does is connect information across the formats where legal work actually lives. Spreadsheets, documents, and presentations aren’t three separate workflows. They’re one workflow that happens to span three file types.

Now you have an AI that finally sees it that way too. And that landed this week, not next year.

Here's what this comes down to. The AI conversation in professional services has been stuck on "generate text" for two years. That's useful, but it misses the bigger problem. Your work doesn't live in one app. It lives in three, and the gaps between them are where time, accuracy, and money disappear. Claude's Office integration is the first tool I've seen that closes those gaps in a way that actually fits how legal teams work. If you want to think through what this means for your firm, or you're trying to figure out where to start, drop me a line at steve@intelligencebyintent.com. The firms that figure this out first won't just be faster. They'll be working on entirely different problems than everyone else.

I have one last really important question I would love some feedback on. I’m thinking of putting all of these examples on YouTube. Would that be valuable? Where prompts/issues/demos can be compared across ChatGPT, Claude, and Gemini, I will do that; or if there’s a specific capability to highlight (e.g., Claude Cowork or NotebookLM), I will focus on that. Please let me know!

PS: I’m doing a special BHBA CLE this upcoming Monday on the brand-new Claude add-in for Word if you are a member. Details here: https://myemail.constantcontact.com/Claude-for-Word--AI-Powered-Drafting-Comes-to-Microsoft-Word.html?soid=1102591689817&aid=W5U9-39dXqA

Google Built Three Different Geminis, and Business Users Are Getting the Worst One

TL;DR: Google’s Gemini is actually three different products depending on whether you’re using consumer, Workspace, or Enterprise. Consumer gets the best stuff first. Workspace, the version your firm is paying for, is missing memory, custom instructions, and real NotebookLM integration. The web app also appears to throttle the model’s thinking depth. I love Gemini. I just want the business version to be as good as the one I use on my personal account.

Three Products, Three Teams, One Confusing Brand

I have a consumer Gemini account. A Workspace account. And an Enterprise account. I use all three, every day, advising law firms on AI strategy.

They feel like three completely different products.

I don’t mean small differences. I mean features that exist in one version and are completely absent in another, with no obvious logic to explain why. The models themselves are excellent. I think Gemini’s models are among the best available right now, full stop. But the product wrapping around those models? It’s a mess.

And the version getting the short end of this is the one business users depend on (workspace!).

Consumer Gemini Gets the Good Stuff First

The consumer version, tied to your personal Google account with a Google AI Pro or Ultra subscription, is clearly where Google’s product energy goes.

Personal Intelligence shipped in early 2026. Memory works, and it works well. You can manage individual items, delete things you don’t want retained, connect apps like Gmail, Calendar, Drive, Photos, YouTube, GitHub, and more. You get custom instructions so you can tell Gemini how you want it to respond. The NotebookLM integration is genuinely exciting. Google folded NotebookLM into the Gemini Projects experience, so your notebooks become living sources inside Gemini conversations. I’ve been a NotebookLM power user since launch, and this is exactly the kind of integration that makes you want to go all-in on Google’s tools.

Scheduled activities work. Memory works. Custom instructions work. It all hangs together.

Now go try that in Workspace.

Workspace: Where the Work Happens, Minus the Features

This is the version your firm is probably running. This is where billable work gets done, where client data lives, where you actually need AI to be smart about context.

Memory? Doesn’t exist. Every conversation starts cold. Gemini has no idea who you are, what you were working on yesterday, or what matters to your practice. Consumer users get persistent memory. Workspace users get amnesia.

Custom instructions? Nope. You can’t tell Workspace Gemini your role, your preferences, how you like things structured. The consumer version has had this for months.

The NotebookLM situation is particularly frustrating. NotebookLM itself works great in Workspace, I use it constantly with clients. But that deep Projects integration from the consumer side? It hasn’t crossed over. You can add a notebook as a source, sure. But it’s a file reference, not the woven-together experience consumer users get.

Even the app connectors are thinner in some ways. Yes - there are different business connectors here - including MailChimp, Hubspot, and Salesforce (that is awesome!) But YouTube Music! Not YouTube? No Google Photos?

Scheduled activities do work, I’ll give it that. But try explaining to a managing partner why the consumer version their associate uses at home has features the firm’s paid Workspace subscription doesn’t. I’ve had that conversation. It doesn’t go well.

Enterprise Goes Its Own Way

Enterprise Gemini is a different animal entirely. More business connectors, which makes sense. A rudimentary agent builder, which is promising. And it actually has memory, though it works differently than the consumer version. You can reference past chats, connected data, and manage saved memories.

But the agent builder needs work. I tried building what should be a simple morning briefing: “At 8AM every morning, read my overnight emails, look at my calendar, and prep me for what I need to know.” I walked through the whole setup. It built the agent. Then I went back to edit it, and the schedule had vanished. Gone. Consumer and Workspace both handle scheduled tasks just fine, but the Enterprise agent builder, the one with “Enterprise” in the name, can’t hold onto a schedule.

Three builds. Three sets of assumptions about what users need. Three incomplete products that could be one great one.

The Magnus Problem (or, Why Gemini Live Is Both Amazing and Maddening)

Let me tell you about my dog.

Magnus is my St. Bernard. If I open Gemini on my consumer account and type “how old is Magnus,” it knows exactly who I’m talking about. Pulls from memory, gives me the answer. That also works on Enterprise. Doesn’t work on Workspace, obviously, because Workspace doesn’t have memory.

But here’s the thing that really gets me. If I switch to Gemini Live, the voice mode, that same consumer account that just answered my question about Magnus has no idea what I’m talking about. Memory doesn’t carry over to Live. At all.

I use Gemini Live constantly. In the car. On vacation in Europe and Japan. The voice interaction is phenomenal. It blows ChatGPT’s voice experience out of the water! But it’s a completely separate experience from text-based Gemini. No shared context. No shared memory. No continuity.

I know this is a hard engineering problem. Streaming voice inference with real-time memory lookup across connected data is genuinely difficult. But here’s what I keep coming back to: nobody else can do this except Google. They have the email. The calendar. The documents. The photos. The search history. The voice infrastructure. All the raw material is right there. It’s just not connected yet.

When it is, it’ll be incredible. Right now it’s a collection of incredible pieces that don’t always talk to each other.

Are We Getting the Full Model?

There’s something else going on that’s harder to pin down, and it’s been bugging me.

The Gemini web app, even when I select 3.1 Pro, often gives me what feels like a half-effort response. Surface-level. Quick summaries instead of analysis. The kind of answer you’d get if someone was trying to finish a conversation quickly.

About a month ago, users started surfacing evidence of a hidden effort level parameter. The default across the app, including for Ultra subscribers paying $249.99 a month, appeared to be medium (0.50).

Now, I want to be careful here. There’s real debate about whether the model is accurately reporting an actual hidden parameter or just producing a confident-sounding hallucination. Google hasn’t confirmed anything. But my own experience lines up with what people are describing.

I have a go-to test. I ask the model to create an SVG of a St. Bernard riding a bicycle, wearing a baseball cap, and it’s raining. I’ve done this test dozens and dozens of times. (Yes, I make everything about my dog.) In the Gemini web app with 3.1 Pro selected, I get something basic. Take the exact same prompt to AI Studio, same model, set thinking to high, and the result is dramatically better. More detailed, more creative, more technically accomplished.

Here’s the SVG from the web app:

Same prompt now from AI Studio with thinking set to high (same model):

I will say that when 3.1 Pro first launched, the web app gave me versions much closer to what I get from AI Studio now… But it’s definitely felt downgraded since.

Same model. Very different output. That gap shouldn’t exist for paying users.

I’m not asking for Deep Think on every query. Most prompts don’t need it. But when I want the full model, actually engaged, I shouldn’t have to switch to a developer tool to get it. A simple toggle would fix this.

What This Actually Means for Your Firm

My legal clients love Gemini’s capabilities. The image processing is excellent. Analyzing video depositions is a genuine differentiator that other platforms can’t match. The 1M token context window opens up document review workflows that would choke every competitor. When they can access the full model, the reasoning and writing are strong.

But then I have to explain why the consumer version of the product has features that the business version doesn’t. That’s not a great pitch when you’re trying to get a firm to commit to a platform for the long haul.

What I Want to See

I’m writing this because I love Gemini, not in spite of it. NotebookLM is incredible. Nano Banana 2 for image generation is outstanding. The voice capabilities are the best I’ve used on any platform. I don’t want to switch. I want Google to bring the business versions up to what I know they can do.

1. Ship memory and custom instructions to Workspace. Business users need persistent context more than anyone.

2. Unify the NotebookLM integration across all tiers. If it works on consumer, bring it to Workspace and Enterprise.

3. Give paying users a way to access the full model’s depth in the web app. A toggle. A button. Something.

Maybe what Google needs is a PM for the PMs. Someone whose whole job is making sure these three product lines don’t keep drifting apart.

Because right now, the best version of Gemini is the one nobody’s using for work. And that’s exactly backwards.

If you read this far, you’re probably not asking whether your firm should use AI. You’re asking whether the platform you’re paying for is actually giving you what you think it is.

That’s the conversation I have every day with managing partners, CIOs, and practice leaders who are deep enough into implementation to feel the friction but aren’t sure if the friction is normal or a warning sign. If that’s where you are, tell me what you’re seeing. steve@intelligencebyintent.com. I’ll tell you what I’d do, what’s working at other firms, and what’s not ready yet.

AI Agents Are Here. Your Competitors Already Know.

TL;DR: AI agents aren’t chatbots. They run multi-step workflows on their own, and the market moved fast while most firms were still figuring out prompts. Claude Cowork, OpenClaw, Perplexity Computer, ChatGPT’s agent mode. All live. All shipping. For law firms, this changes how clients evaluate your fees. For PE firms, it’s the operational lever sitting inside every portfolio company that nobody’s pulling yet. Here’s what I know, what I’m still figuring out, and what you should do about it.

You’ve used ChatGPT. Maybe Claude or Gemini. You opened a tab, asked a question, got an answer, closed the tab.

That was the old world. And I think it ended sometime in the past year, though honestly it’s hard to pin an exact date because everything moved at once and then kept accelerating.

We’re in the agent era now. The gap between “I’ve heard of AI” and “AI runs parts of my operation” is getting wider every week. If you’re a managing partner at a law firm or you’re running a PE portfolio, this isn’t theoretical anymore.

So What’s an Agent?

Simplest way I can put it. A chatbot answers questions. A copilot helps while you work. An agent does the work.

“Summarize this PDF” is a prompt. Any chatbot handles that. But “Research 10 companies, score them on five criteria, build a comparison table, and email me the results” is an agent task. That second one involves planning, using tools, looping back when something breaks, and delivering a finished product. No human in the loop until the output lands.

The practical gap is enormous. A chatbot saves you minutes. An agent saves you hours. Sometimes days. And I’m not being hyperbolic. I’ve watched my own agents do in 15 minutes what used to take me an entire day.

Over the past year, and accelerating hard in recent months: Claude Cowork went generally available with enterprise controls. OpenClaw, an open-source agent, rocketed past 300,000 GitHub stars and became a genuine phenomenon. Perplexity launched Computer, which orchestrates 19 different AI models from the cloud at $200 a month. OpenAI folded its Operator product into ChatGPT’s new agent mode. Google is pushing agentic capabilities into Gemini everywhere it can.

That’s not a trend. That’s a land rush.

Three Ways In

Not every firm needs the same thing, and I want to be careful here because the temptation is to jump straight to the most impressive-sounding option. Don’t.

No-code desktop agents. Claude Cowork is the one I demo most often. It runs on your desktop, interacts with files, browser, apps. You tell it what to do in plain English. “Every Monday at 8am, pull my calendar, summarize open items from our project tracker, and draft my weekly status update.” That’s the whole instruction. It just runs. Cowork now ships with enterprise features like role-based access, spend limits, usage analytics. If you’re a practice group leader or you run an ops-heavy team without engineering resources, this is where you start. Not where you end up. Where you start.

Custom builds. This is what I did with Project Ollie. Fourteen specialized agents running on a $600 Mac Mini. I have agents for AI research, legal research, CRM management, email scanning, calendar intelligence, content pipeline, security monitoring. All autonomous. All delivering through Telegram and email. Monthly API cost is basically nothing beyond my existing subscriptions. Here’s the thing, though. The barrier isn’t money. It’s thinking clearly about what each agent should and shouldn’t be allowed to touch. I’ll come back to this in the security section because I made mistakes and I’d rather you learn from mine.

Managed cloud agents. Anthropic launched Managed Agents on April 8th. Perplexity Computer runs entirely in the cloud. ChatGPT agent mode, same idea. Someone else handles the sandboxing, sessions, permissions, observability. Your team focuses on the task, not the infrastructure. For engineering teams or companies that want production agents fast, this path makes a lot of sense. The tradeoff is control. You’re renting, not owning.

What Your Clients Already Expect

I want to be direct here because I think some firms are still treating AI as a “nice to have” conversation.

It’s not.

A Harvard Law School Forum piece from last month said it plainly: clients aren’t asking whether firms use AI anymore. They expect to see the benefits passed to them. More insight, more speed, more value per dollar. Not cheaper rates, necessarily. Better outcomes.

Harvey just raised $200 million at an $11 billion valuation. More than 25,000 custom agents on their platform. Over 1,300 customers across 60 countries, most of the Am Law 100. That tells you where institutional money thinks legal work is going.

And it’s not just the big legal tech players. Y Combinator’s 2025 Request for Startups told founders, in effect, to build AI-native service firms that compete with incumbents, using a law firm as the prime example. Not sell software to firms. Compete with them. AI-first firms like Crosby and Avantia are already operating on fixed-price models with no billable hours. They’re small. They won’t stay small.

The one that should really get your attention, though, is a prediction from the Debevoise Data Blog. They said corporate legal departments will increasingly bring routine legal work in-house, generate drafts with AI, and send those drafts to law firms to review. Think about what that does to the relationship. Who’s responsible for accuracy? Who bears the malpractice risk on an AI-drafted document that a client generated and a firm reviewed? Nobody has clean answers yet. But the question is coming, and it’s coming this year.

Meanwhile, Gartner says 40% of enterprise apps will feature task-specific AI agents by end of 2026. Colorado’s AI Act takes effect in June. EU AI Act compliance deadlines are currently set for August, though proposed amendments could push some obligations into 2027. If you don’t have an AI governance policy ready for your next client RFP, you’re already behind regardless of which deadline lands first.

The PE Angle

If you’re sitting on a portfolio of 15 companies, agents are the operational multiplier hiding in plain sight.

Where they create value: ops efficiency across portfolio companies, data consolidation after acquisitions, faster due diligence, automated reporting. Every one of those is a defined, repeatable workflow. That’s exactly what agents eat for breakfast.

The readiness question is simpler than people expect. Clean data. Defined workflows. Access controls. Someone accountable for AI at each company. Budget.

And the hardware barrier? Six hundred dollars. A Mac Mini running 24/7 with Claude Code proves the concept. You don’t need fourteen agents like I built. You need one. Pick the task someone does every single day that’s boring and well-defined. Automate that. See what happens. Scale from there.

Security. The Part Nobody Wants to Talk About.

The gap between demo and production is security. Full stop.

I learned this the hard way. One of my agents executed a command it shouldn’t have. Nobody got hurt. But it proved that guardrails aren’t something you add later. They’re the architecture.

OpenClaw is the cautionary tale here. It went viral in late January, 100K GitHub stars in days, the whole developer community losing its mind over it. And then a critical remote code execution vulnerability dropped (CVE-2026-25253, CVSS 8.8) that put over 15,000 publicly exposed instances at risk of one-click compromise. Cisco’s security team tested a third-party OpenClaw skill and found it was exfiltrating data without user awareness. One of OpenClaw’s own maintainers warned publicly: if you can’t run a command line, this is too dangerous for you.

For law firms handling privileged communications? For PE firms touching confidential deal data? That’s not a footnote. That’s the conversation.

My rules, earned through screwing up: read-only access first, always. Write access is earned, not assumed. Agents never write directly to primary accounts. API keys go in encrypted, permission-locked files. Log everything. Treat every failure as a system improvement, not a bug.

On Hallucinations

Agents don’t fix the hallucination problem. They make it worse. An agent running for hours can compound a wrong assumption across dozens of steps before anyone notices. In legal work, that’s malpractice. In diligence, that’s deal risk.

You don’t avoid agents because of this. You design them with checkpoints and human review gates. The best legal AI tools in 2026 aren’t maximizing autonomy. They’re constraining it. Structured workflows. Clear handoffs. Scoped authority.

One legal tech leader I heard recently put it this way: “Show me your guardrails” now means “show me your workflow.” I think that’s exactly right.

What to Do Monday Morning

1. Pick one task and automate it. Not your most complex workflow. The boring one. The one someone does every day that makes them mutter under their breath. Weekly status reports. Client news monitoring. Board prep. Calendar briefings. Start there.

2. Audit your data. Agents need clean, accessible data. If your documents live across five systems with no naming convention, fix that first. Nothing else works until the data does.

3. Get your governance ready. Write an AI policy. Or update the one you wrote 18 months ago that’s already stale. Colorado hits in June. EU deadlines are in flux but coming. Your clients will ask. Have an answer.

Where This Lands

I’ll be honest. I don’t know exactly how fast this moves. Some of the predictions floating around feel breathless. “90% of legal documents AI-generated by end of year.” Maybe. I’m skeptical of the timeline, less skeptical of the direction.

What I do know: agents work. They’re shipping. The security risks are real but manageable if you’re disciplined. And the firms and portfolio companies that treat them like junior team members, with supervision, clear boundaries, and regular check-ins, will pull ahead of the ones still debating whether to experiment.

The window’s open. I don’t think it stays open as long as people assume.

If you read this far, you’re not wondering whether AI agents are real. You’re trying to figure out which ones are safe to deploy and how fast you need to move. That’s the right question, and it’s the one most of the breathless coverage skips entirely.

That’s the conversation I have every day with managing partners, GCs, and PE operating teams who are past the hype and into the hard decisions. If you’re working through where agents fit in your operation, or you’re trying to separate what’s ready from what’s still a demo, send me a note at steve@intelligencebyintent.com. Tell me what you’re sitting with. I’ll be direct about what I’ve seen work and what I’d wait on.

Share Intelligence by Intent

Claude Just Moved Into Microsoft Word. Lawyers Should Pay Attention.

TL;DR: Anthropic launched Claude for Word as a beta add-in that lives inside your document, edits with track changes, reads comments, preserves formatting, and shares context across Word, Excel, and PowerPoint simultaneously. It’s built for document-heavy professionals, and Anthropic isn’t being subtle about who they mean: the first use case on the product page is legal contract review. I’ve been using it. It’s genuinely good. And it’s better than what Microsoft currently offers natively.

I installed it yesterday morning. By lunch, I’d used it to tear through a 30-page document, rewrite two sections, and respond to a string of comments. All without leaving Word. All with track changes on.

That last part matters more than anything else I’m about to tell you.

Why Track Changes Is the Whole Ballgame

If you work with lawyers, you already know this. Track changes isn’t a nice-to-have. It’s the operating system of legal document work. Every edit needs to be visible. Every revision needs to be reviewable. Every change needs an audit trail that a partner, client, or opposing counsel can walk through line by line.

Here’s what Claude for Word actually does: when you flip it into suggested edits mode, every change Claude makes shows up as a native Word tracked revision. Deletions in red. Insertions in green (unless you’re like me and have changed your colors!). You accept or reject each one in Word’s own review pane, the same way you’d handle markup from a junior associate or co-counsel.

Think about what that means. You can tell Claude to rewrite an indemnification clause, make it mutual, cap damages at 12 months of fees. And instead of getting a blob of text you have to copy-paste and manually compare, you get a clean redline you can review in 30 seconds.

I’ve been testing AI tools in legal workflows for over two years now. This is the first time the output arrives in a format that matches how lawyers actually work.

It Reads Your Document Like a Colleague Would

Most AI integrations treat your document like a pile of text to be summarized. Claude for Word does something different. It reads your heading styles, your numbering schemes, your defined terms, your cross-references. When it edits a section, the surrounding formatting stays intact. No blown-up numbering. No reformatted headers. No wrestling with styles after every AI touch.

This sounds like a small thing. It’s not. Anyone who’s ever had to rebuild a 50-page agreement’s numbering after a bad paste knows exactly what I’m talking about.

And then there’s comment-driven editing. Leave comments in your document where you want changes. Claude reads each comment thread, understands what text the comment is anchored to, edits that specific passage as a tracked change, and replies to the comment thread explaining what it did. For review workflows where a partner drops notes like “tighten this” or “make this mutual” or “I don’t like this language,” Claude can work through the entire stack.

Cross-Document Context Is a Quiet Superpower

Here’s a feature that hasn’t gotten enough attention yet. Claude for Word shares context with Claude for Excel and Claude for PowerPoint. In a single conversation, Claude can see what’s open across all three apps.

Think about what that means for deal work. You’ve got a financial model open in Excel, a presentation deck in PowerPoint, and a memo or agreement in Word. You can tell Claude to pull specific numbers from the model into the memo. Or to draft an executive summary in the deck based on the agreement’s key terms. No copy-paste. No re-explaining what’s in each file.

For firms doing M&A, private equity, or any transaction work that involves bouncing between financial data, slide decks, and legal documents, this is a significant time savings. And it’s genuinely hard to find this capability anywhere else right now.

It’s Better Than What Microsoft Offers. That’s Not Just My Opinion.

I want to be direct about this because I know some of you are already paying for Microsoft 365 Copilot and wondering whether you need another AI tool in Word.

Early indications suggest Claude handles long documents, complex reasoning, and careful editing well compared to Copilot’s current offering, though head-to-head reviews are still emerging. Claude’s context window lets it process entire contracts and lengthy reports in a single pass, something that trips up other tools. And the tracked changes implementation, where every edit lands as a reviewable revision, is cleaner and more reliable than what Copilot currently delivers.

In my testing, Claude for Word feels less like a tool bolted onto the application and more like it actually understands Word’s document model. The formatting preservation, the comment-thread awareness, the style inheritance when filling templates. These are details that matter enormously to professionals who live in Word all day.

It’s worth noting that Microsoft is now integrating Claude models into its own Copilot product. When a company like Microsoft builds your models into its platform alongside its own, that says something about where the technology stands.

What You Should Know Before Rolling This Out

A few practical realities to keep in mind.

Claude for Word is in beta. Anthropic is upfront about the limitations: don’t use it for final client deliverables without human review, don’t treat it as a replacement for legal judgment, and be cautious with highly sensitive documents. These are reasonable guardrails, not dealbreakers.

The add-in works on Word for Windows, Mac, and web. It requires a Claude Team or Enterprise subscription. For firm-wide deployment, your IT team can push it through the Microsoft 365 Admin Center or use a manifest file for organizations that have the Office Store locked down.

One important caveat: Anthropic warns against using Claude for Word with untrusted documents from external sources. Prompt injection attacks, where hidden instructions in a document trick the AI into unintended actions, are a real risk. If you’re reviewing a counterparty’s redlined draft, be thoughtful about that.

Data retention is also worth understanding. Inputs and outputs are deleted within 30 days on Anthropic’s backend, but chat history doesn’t persist between sessions. The add-in doesn’t yet connect to Enterprise audit logs or the Compliance API. For firms with strict data governance requirements, that’s a conversation to have before deployment.

What to Do Monday Morning

1. Have someone on your team install the add-in from the Microsoft Marketplace and test it on a non-sensitive document. The setup takes five minutes.

2. Try the tracked changes workflow on a real contract review. Ask Claude to flag off-market provisions, rewrite a clause, or work through a set of reviewer comments. See how the output compares to what a junior associate would produce.

3. Talk to your IT and data governance teams about deployment logistics, data retention policies, and whether the current beta limitations are workable for your firm’s risk tolerance.

This isn’t the kind of tool you evaluate by reading about it. You need to see a redline appear in your document, in your formatting, with your defined terms preserved, to understand why this one feels different.

Anthropic is clearly going after the legal market. The product page leads with contract review. The demo videos show NDA redlining. The prompt examples read like a first-year associate’s task list. And they’re building it on top of the application where lawyers already spend most of their working hours.

That’s not a coincidence. That’s a strategy.

If you read this far, you’re not wondering whether AI belongs in your document workflow. You’re trying to figure out whether this specific tool is ready for the way your team actually works, and whether the risks are manageable enough to find out.

That’s the conversation I have every day with firm leaders who are past the hype cycle and into the hard questions. If you’re working through deployment logistics, data governance concerns, or just trying to figure out what’s worth testing and what isn’t, send me a note at steve@intelligencebyintent.com. Tell me what you’re sitting with. I’ll give you a straight answer about what’s ready and what still needs time.

Google Just Merged Gemini and NotebookLM. Here’s What That Means for Your Firm.

TL;DR: Google launched “Notebooks” in Gemini this week, connecting it directly with NotebookLM. Think of it as Google’s answer to Projects in ChatGPT and Claude. It’s a dedicated workspace where you can organize sources, files, and conversations around a single topic. The sync between the two apps is genuinely useful. But the privacy picture is more complicated than Google is advertising, and it hasn’t rolled out to Workspace (business) accounts yet. If you’re advising clients or handling sensitive material, you need to understand what changed before you start using it.

I Was Standing in Front of 40 Attorneys When This Dropped

I was presenting to a group of attorneys earlier this week, walking them through just how powerful NotebookLM can be. Upload your deal documents, your case files, your regulatory filings. Ask questions. Get answers grounded in your actual sources, with citations. No hallucinated case law. No wandering off into the internet.

It’s one of the tools I recommend most to my legal and PE clients. I’ve been using it since it launched, and I’ve watched dozens of firms build it into their daily work.

And then, mid-presentation prep, Google announced they’re merging NotebookLM directly into Gemini.

My first reaction: this is great. My second reaction: wait, what happens to the privacy model?

Both reactions turned out to be right.

What Google Actually Built

Google introduced a feature called “Notebooks” inside the Gemini app. The naming is a little confusing since NotebookLM already has notebooks. But here’s the simple version.

You can now create a dedicated workspace inside Gemini, attach files to it (PDFs, Google Docs, website URLs, YouTube videos, pasted text), set custom instructions, and have all your conversations within that workspace stay organized in one place. Those notebooks sync automatically with NotebookLM. Add a source in Gemini, it shows up in NotebookLM. Update your custom instructions in NotebookLM, it syncs to Gemini.

Google’s own language is telling. They’re calling notebooks “personal knowledge bases shared across Google products.” That’s not a minor label. They’re positioning this as a persistent layer of context that follows you across their entire product suite, starting with these two apps.

If you’ve used Projects in ChatGPT or Claude, you’ll recognize the concept immediately. A contained space. Dedicated context. No information bleeding across workspaces. Every one of my clients cares about that separation, especially when they’re handling client matters or confidential deal information.

Why the Two-App Model Is Actually Smart

Here’s what makes Google’s approach different from what OpenAI and Anthropic have done with Projects.

NotebookLM and Gemini are good at different things. NotebookLM is a research tool. It stays grounded exclusively in your uploaded sources. It won’t go to the web. It won’t hallucinate from training data. When it answers a question, it pulls from what you gave it and shows you exactly where.

Gemini is a general-purpose assistant. It can search the web, write drafts, reason across topics, and use Google’s full model capabilities. But until this week, Gemini had no dedicated project workspaces. You could reference past chats and add NotebookLM as a source, but there was no persistent, organized container for all your project materials in one place.

Now they talk to each other. You can upload your source documents in NotebookLM, generate an Audio Overview or a slide deck from them, and then jump into Gemini to ask follow-up questions about those same documents while also pulling in live web data. No exporting. No re-uploading. No copying files between apps.

Custom instructions sync across both. Sources sync across both. It’s genuinely useful for anyone running multi-week projects.

The Privacy Problem Nobody’s Talking About

Here’s where it gets complicated. And this is the part that matters most if you’re a law firm, a PE shop, or any business handling confidential information.

NotebookLM has always had a clean privacy story. Your uploaded sources stay private. Your chats are not used to train Google’s models. Period. That hasn’t changed.

But the moment you start chatting with your notebook through Gemini’s consumer app, different rules apply. Google’s own support page spells it out: chats in Gemini are governed by your “Keep Activity” settings and the Gemini Apps Privacy Hub. Translation: if you haven’t disabled activity tracking in the consumer version of Gemini, your inputs and outputs can be used for model training.

Let me say that more plainly. Your documents in the notebook? Not trained on. Your chats about those documents in NotebookLM? Not trained on. Your chats about those same documents in Gemini? Potentially trained on, unless you’ve changed your settings.

That’s a meaningful distinction. And it’s one that most users won’t notice.

I spent some time chatting with Gemini 3.1 Pro about this to confirm. The notebooks themselves, just because they’re connected to Gemini, aren’t being trained on. But the conversation layer in Gemini operates under Gemini’s privacy rules, not NotebookLM’s.

There’s another wrinkle worth knowing. Your NotebookLM chats don’t appear over in Gemini. But your Gemini chats do appear in NotebookLM, as read-only context in your sources panel. And if you’ve opted in to “shared context,” collaborators on a shared notebook can see those Gemini chats too.

The Deletion Quirk You Need to Know

One more thing that tripped me up. When you delete a notebook, it destroys the container, the source documents, and the NotebookLM artifacts. But it does not delete your Gemini chat history. Those conversations get ejected from the notebook and returned to your main Gemini chat list.

For anyone managing client matters with retention policies, that’s a detail you can’t ignore. Deleting the project doesn’t mean deleting the conversation trail in Gemini.

A Few Things I Discovered Poking Around

Shared notebooks don’t show up in Gemini. At all. I have several shared notebooks, some that colleagues created and shared with me, others that I created with public link sharing enabled. None of them appeared in my Gemini notebook list. When I turned off the public link on one of my notebooks, it immediately showed up.

Only notebooks that are truly yours, unshared, private, appear in Gemini. That’s actually a reasonable privacy choice by Google, but it’s not documented anywhere obvious.

Also worth noting: Gemini can’t generate NotebookLM Studio artifacts. No Audio Overviews, no Video Overviews, no Infographics from within Gemini. You still need to go to NotebookLM for those. The sync is for sources and context, not for the full feature set.

The Big Gap: Workspace Users

Here’s the thing that will matter most to my readers. This feature is currently rolling out to consumer accounts only. Google AI Ultra, Pro, and Plus subscribers on the web. Google’s own footnotes confirm that Notebooks in Gemini are not available for Workspace or Education accounts yet.

That’s the gap. The businesses that would benefit most from this, law firms, consulting shops, financial services, are exactly the ones who can’t use it yet.

Google Workspace accounts already have strong privacy protections. Content isn’t used for model training. Human reviewers don’t see your data. Enterprise-grade security is baked in. If Google brings this same Notebook integration to Workspace with those protections intact, it could be a significant differentiator.

But we don’t know when that’s coming, and we don’t know if the privacy model will carry over cleanly. I truly hope it does. Because right now, Workspace users who need project-style workspaces are still better served by Claude Projects or ChatGPT Projects, both of which are available today with business-grade privacy.

What to Do Next

1. If you’re a consumer Gemini user, check your Keep Activity settings right now. Go to myactivity.google.com and review what Gemini is tracking. Decide whether you’re comfortable with that before you start chatting with notebooks in Gemini.

2. If you’re already using NotebookLM, your existing notebooks will appear in Gemini automatically. Test the integration, but be deliberate about which conversations you have in Gemini vs. NotebookLM.

3. If you’re advising clients on AI tools, update your guidance to reflect the split privacy model. NotebookLM chats and Gemini chats about the same notebook operate under different rules.

4. If you’re on a Workspace account, sit tight. This hasn’t rolled out to business accounts yet. Keep using NotebookLM directly, and watch for announcements about Workspace integration.

The Bottom Line

Google built something genuinely useful here. Connecting a grounded research tool with a general-purpose AI assistant, sharing context across both, syncing sources and instructions automatically. That’s the right architecture.

But the privacy story is messier than it should be. Two apps, one notebook, two different sets of rules for what happens to your conversations. For professionals handling sensitive information, that’s not a detail. That’s the whole ballgame.

I’ll be watching the Workspace rollout closely. That’s where this feature either becomes essential or stays a consumer curiosity.

If you read this far, you’re not someone who saw “Google update” and moved on. You’re the person at your firm who has to decide whether this is safe to use before everyone else starts using it on their own.

That’s the conversation I have every day with managing partners, GCs, and ops leaders who are trying to stay ahead of exactly this kind of problem. If you’re working through how to evaluate AI tools when the privacy story keeps shifting underneath you, send me a note at steve@intelligencebyintent.com. Tell me what you’re sorting out. I’ll tell you what I’m seeing and where I think the real risks are, straight.