From Prompts to Products: What Actually Matters from OpenAI's DevDay
Last year everyone made a custom GPT. This year you can actually ship something people will use on purpose
Here’s what actually matters from OpenAI’s DevDay yesterday, minus the fog. If last year was a neat demo, this year felt like a platform you can build on. The headline is simple: ChatGPT is becoming the place where people do things, not just ask things. If you lead a team, that should change your roadmap this quarter.
1) Apps inside ChatGPT: your website didn’t move, your customer did
OpenAI introduced an Apps SDK that lets developers build chat-native apps that run directly inside ChatGPT. Think mini apps a user can open, authorize, and use without leaving the conversation. There’s an app directory, submission and review, and early brand partners. Translation: a new distribution channel where intent is already high because the user is mid-task.
Practical uses I would ship now:
Sales: a pricing configurator that lives in the same chat where a prospect is asking questions. Capture requirements, generate a quote, and push the draft to Salesforce in one flow.
Support: a warranty app that checks purchase history, runs quick triage, schedules a return label, then follows up with a summary email. No context switching.
Legal: an NDA intake app for business users. Paste the counterparty paper, get a standardized version, see risky clauses explained in plain English, accept firm-approved edits, then file the clean copy to iManage or NetDocuments with the right matter number and permissions. The user never leaves the chat.
Compared with last year’s plugins and custom widgets, this looks closer to a real marketplace with clearer permissions and discovery. If you run product, mark this line: distribution beats features.
2) AgentKit: build task doers, not just chatty helpers
AgentKit includes Agent Builder (a visual canvas to connect nodes/tools), ChatKit for embedding agentic chat UIs, and expanded Evals (datasets, trace grading, automated prompt optimization, third-party model support) to measure and improve behavior over time. If you have used n8n or Make, the mental model will feel familiar. Only now, the “nodes” can reason, and the handoffs can include natural language.
Where I’d point it first:
Finance ops: an invoice-chaser that watches your ERP aging report at 7 a.m., drafts three follow-ups in the right tone, posts them for AE review in Slack, then logs outcomes back to the opportunity.
IT: a joiner-mover-leaver agent that reads HRIS updates, opens tickets, sets permissions, and confirms completion with the manager.
Legal ops: matter-intake triage that converts free-form requests into structured records, routes to the right practice group, and sets a 24-hour clock. Lawyers see fewer pings and better context.
If you tried custom GPTs last year and walked away, I get it. This is different because you can wire real steps, set evaluation criteria, and treat the agent like a product, not a toy.
3) GPT-5 Pro and Sora-2 in the API: more headroom for real work
GPT-5 Pro in the API means better reasoning, tighter instruction following, and enterprise-friendly control. For teams that need consistency across long workflows, that matters more than a flashy demo.
Sora-2, OpenAI’s new video model, is now part of the story as well. The public messaging includes visible watermarking and C2PA metadata plus safety limits spelled out in a system card. I’ve seen social posts claiming the API output is clean of watermarks. Treat that as rumor until official docs are crystal clear. For planning purposes, assume watermarking and content restrictions exist. Safer that way.
What you can do today:
Marketing: generate 6-second and 10-second product loops that match your brand kit, then A/B them in paid social. Add a simple rule so the agent rejects disallowed content or celebrity likenesses.
Training: turn SOPs into short video explainers for frontline teams. Keep them task-specific, one objective per clip.
Legal and compliance: quick fact-pattern simulations for staff training. Build an approval step by counsel and a log that stores prompts, outputs, and reviewers for audit. Your policy owners sleep better.
4) Codex, now GA, is not just for engineers
Codex graduated to general availability with an SDK, Slack integration, and admin controls. Yes, this helps developers move faster. The interesting part for a COO is how many small code tasks live outside engineering. That Excel macro your sales ops lead is scared to touch, the nightly CSV cleanup someone runs by hand, the weekly SharePoint export that breaks when column names change, these are Codex snacks. Put them on a backlog and clear ten a week.
A quick reality check on “who gets disrupted”
It’s easy to say half of SF just got wiped out. Some tools will. App-in-chat collapses many thin wrappers: products that only pass data between services with a pretty UI, or assistants that live and die on top-of-funnel traffic. On the other hand, companies with real distribution, data rights, or industry depth will do fine. If you sell legal research with citation networks, you bring something a chat app cannot clone overnight. The rest should assume ChatGPT is now a channel, then design for it.
What I’d do this quarter
Ship one app that saves time for non-technical staff. Make it boring and valuable: NDA intake, pricing config, returns. Four weeks is enough if you keep scope tight.
Stand up one agent with a clear KPI. Example: days-sales-outstanding down by two days, or first response time on legal intake under two hours. Give it an owner and a weekly evaluation rhythm.
Write a light policy for AI-generated video. Cover watermark expectations, review steps, and where these assets can ship externally. Keep brand and counsel aligned.
If you remember nothing else, remember this: last year we talked about prompts. This year we are talking about products. The teams that turn one or two of these into real workflows will bank the gains first.
Moving Forward with Confidence
The path to responsible AI adoption doesn’t have to be complicated. After presenting to nearly 1,000 firms on AI, I’ve seen that success comes down to having the right framework, choosing the right tools, and ensuring your team knows how to use them effectively.
The landscape is changing quickly - new capabilities emerge monthly, and the gap between firms that have mastered AI and those still hesitating continues to widen. But with proper policies, the right technology stack, and effective training, firms are discovering that AI can be both safe and transformative for their practice.
Resources to help you get started:
In addition to publishing thought AI leadership on a regular basis, I also work directly with firms to identify the best AI tools for their specific needs, develop customized implementation strategies, and, critically, train their teams to extract maximum value from these technologies. It’s not enough to have the tools; your people need to know how to leverage them effectively.
For ongoing insights on AI best practices, real-world use cases, and emerging capabilities across industries, consider subscribing to my newsletter. While I often focus on legal applications, the broader AI landscape offers lessons that benefit everyone. And if you’d like to discuss your firm’s specific situation, I’m always happy to connect.
Contact: steve@intelligencebyintent.com
Share this article with colleagues who are navigating these same questions.