OpenAI shipped a product on Thursday whose output is somebody’s finished work product. ChatGPT Work, launched July 9 alongside the general release of GPT-5.6, connects to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars and CRMs, reads what it finds there, and hands back the deliverable: the spreadsheet, the deck, the report, even a working web app. It went live on desktop across every ChatGPT plan the same day, with web and mobile following.
The coverage filed it as a shot at Microsoft Copilot, and it is one. OpenAI selling an enterprise agent puts it squarely in Microsoft’s yard, days after the two spent a quarter building rival deployment armies. But the Copilot fight is between two giants who can afford it. The exposure nobody priced is further down the stack: the thousands of software companies that spent the last two years building this exact product on top of OpenAI’s models.
The app layer built its roadmap on the lab’s model
Since 2024 the standard SaaS play has been an AI layer over your own product. Connect the customer’s data, wrap GPT around it, ship “your copilot for finance” or legal or sales, and charge for the seat. The whole stack ran on one assumption: the model company stays a supplier. On July 9 the supplier shipped the wrapper. Context gathering across the customer’s apps, connectors into Slack and Drive and the CRM, a finished artifact at the end — that was the roadmap of a few hundred funded startups and the AI-features slide in every SaaS renewal deck I’ve seen this year.
When the IPO filings landed in June, I wrote about Benedict Evans’s argument that models trend toward commodity and the money moves to whoever owns the workflow. The labs read the same memo everyone else did. ChatGPT Work is OpenAI acting on it, against its own ecosystem.
When the agent owns the screen, the seat loses its argument
Per-seat software priced a person sitting in front of an interface, and that model was already dying. ChatGPT Work is built to be the interface. If your product becomes one of the connectors on its menu, something the agent reads from and writes to, the customer still needs your data and your system of record. What they stop needing is as many people logged into your screen. And the screen time was what the seat count billed.
I sold at Fastly against cloud platforms that bundled a version of what we did. You can win those fights, and I learned where: never on the feature list, always on the depth the platform treats as an afterthought — the edge case that was our whole business and a line item in theirs. That lesson is about to get relearned by every vertical SaaS team whose differentiation was “we put AI on your data.” The generalist agent does that now. It comes with the model.
The procurement question ships faster than the capability
None of this lands overnight. ChatGPT Work inherits every permission problem the enterprise already has; an agent that reads everything a marketing manager can touch is a security review, and security reviews are slow. Microsoft will fight with the bundle, and the bundle usually wins the mid-market. A generalist producing a hundred kinds of deliverables will make a worse financial model than a product tuned for exactly that. The demo is ahead of the deployment, as usual.
The question arrives anyway. Some CFO is going to ask “why are we paying for this tool when ChatGPT Work makes the same report” well before ChatGPT Work makes a good version of that report, the same way “why not just use ChatGPT” showed up in software deals two years before an agent could do anything about it. If you sell software, write your answer now, and make it an outcome you own instead of a feature you have: the workflow depth the generalist won’t reach, the proprietary data it can’t see, the number you’re accountable for. And if your AI feature is a wrapper on the supplier’s model, take a hard look at your renewal calendar. OpenAI just showed everyone its roadmap, and yours was on it.