OpenAI Just Put an AI Coworker Inside Its Own App. Here's Why Yours Shouldn't Live There.

The Coworker Moved Into the App

On July 6, 2026, OpenAI made workspace agents generally available inside ChatGPT Business, Enterprise, and Edu. The framing was deliberate. These are not chat prompts. They take on repeated tasks, follow team processes, ask for approval when they need it, and can be shared across an organization, with admin visibility built into the console.

Read that description again and notice what it is. It is a job. It is the shape of a coworker — something that holds a role, follows a process, and knows when to raise its hand. The industry spent three years arguing about whether AI could be more than a tool. That argument is over. The largest AI vendor on the planet just shipped a coworker and put it inside its own app.

The capability is real and worth taking seriously. But the location is the entire problem. An AI coworker that lives inside one vendor's subscription is not part of your team. It is a feature of their product that you rent by the seat. The difference sounds academic until the day it isn't.

This is the distinction that will define which businesses get durable value from AI over the next two years, and which ones end up renting the same coworker three times from three different vendors. So let's be precise about what it means for a coworker to actually belong to your business.

A Tool You Open. A Coworker You Employ.

The line between an AI tool and an AI coworker is not intelligence. The models crossed the "smart enough" threshold a while ago. The line is where the thing lives and who controls it.

A tool lives inside an app. You open the app, you use the tool, you close the app. The tool remembers nothing about your business except what the vendor decides to store, in a format the vendor controls, accessible only through the vendor's interface. When you stop paying, the tool and everything it learned about your work disappear. That is the correct arrangement for a tool. You do not need to own your calculator.

A coworker is different. A coworker holds a role inside your business. It accumulates context that belongs to you — how your invoices are structured, which client hates phone calls, what your escalation path looks like when a deal goes sideways. It connects to your systems. It shows up in the channels your team already uses. And crucially, it keeps existing as an entity in your organization even when the underlying tools change.

We wrote a full breakdown of why an AI coworker is categorically different from an AI tool, and the July launch is the cleanest real-world illustration of the stakes. OpenAI built something with the behavior of a coworker and the ownership model of a tool. It follows your processes, but it lives in their house. It learns your work, but the memory format is theirs. It shows up for your team, but only inside one app, only while you pay for that specific plan.

What good looks like: the coworker holds a role in your org chart, its memory and configuration live in infrastructure you control, and you could change the underlying model tomorrow without the coworker forgetting who it is or what it does.

What bad looks like: the coworker is a line item on a subscription, its memory lives in a vendor's proprietary store, and "switching" means starting a new hire from zero somewhere else.

The 12% Problem Is a Location Problem

Here is the number that should reframe how you think about the July launch. In mid-2026, one widely cited survey found that 96% of enterprises are using AI agents but only 12% can centrally inventory and govern them, with 94% reporting "sprawl" concerns. Almost everyone has agents. Almost nobody knows where they all are or what they can touch.

That gap is not a discipline failure. It is a structural consequence of the exact pattern OpenAI just accelerated. When your coworkers live inside vendor apps, they live wherever the vendors put them — one in the ChatGPT console, one in a CRM's agent panel, one in a spreadsheet plugin, one in a support tool. Each vendor gives you a little admin visibility into its own agents and none into anyone else's. There is no single place that answers the most basic management question: who works here, and what are they allowed to do?

The admin console OpenAI shipped is genuinely useful — for the agents that live in OpenAI. That is the tell. Every vendor's governance stops at the edge of that vendor's product. A business running coworkers across five apps has five governance surfaces and zero org charts. That is how you get to 96% adoption and 12% control.

A coworker that belongs to your business solves this by construction. If the agent's identity, permissions, and connected systems are defined in one operating layer that you own, then inventory is not a project you have to run — it is just the state of your system. You can answer "who works here and what can they touch" the same way you'd answer it for your human team: by looking at the org chart.

Three Questions the App Model Can't Answer

Every business evaluating an in-app AI coworker should ask three questions before they commit real work to it. The in-app model struggles with all three, and the struggle is not a bug the vendor will patch. It is baked into the arrangement.

1. Can it keep working when the model changes under it?

Every AI vendor's coworker is a coworker for that vendor's model. The workspace agent is a ChatGPT agent. Its instructions, memory, and behavior are authored inside one console, tuned for one model family.

Six months from now a model ships that is meaningfully cheaper on your workload, or better at your specific edge cases, or carries a compliance certification you suddenly need. If your coworker's entire identity is defined inside one vendor's app, moving to the better option is a rebuild, not a decision. We covered why model-agnostic architecture matters more, not less, now that every vendor sells an agent platform — and an in-app coworker is the sharpest possible version of the lock-in that piece warns about. The whole employee is inside the app.

What good looks like: you point the same coworker at a different model with a config change, and it keeps its memory, its role, and its connections. What bad looks like: switching models means re-teaching a new agent everything the old one knew, in a different console, from scratch.

2. Does its memory belong to you?

The value a coworker builds over months is not the model. It is the accumulated context — the thousand small facts about how your business actually runs. When that context lives in a vendor's proprietary memory store, you do not own the most valuable asset your AI coworker produces. You are renting access to your own institutional knowledge.

Portable memory is not a technical nicety. It is the difference between an employee whose knowledge stays with the company and a temp whose notes get shredded when the contract ends. We went deep on what durable, governed agent memory actually requires, and the short version is: if you can't export it, inspect it, and move it, it isn't yours.

3. Can you see and govern all your coworkers in one place?

This is the 12% problem again, at the level of a single business. An in-app coworker gives you visibility into itself. It gives you nothing about the agent running in your CRM or the one buried in a spreadsheet integration. Governance that stops at one vendor's boundary is not governance. It is a dashboard.

What good looks like: one operating layer where every coworker's role, permissions, and connected systems are visible and enforceable together. What bad looks like: a different admin panel per vendor, and a spreadsheet where someone tries to track which agents exist.

What to Do Instead

The right response to the July launch is not to avoid AI coworkers. It is to be deliberate about where they live. Here is the sequence we'd walk any operator through.

1. Write down the role before you hire for it. Pick one repeatable, measurable workflow — the kind OpenAI's own guidance says to start with. Document it end to end: every branch, every exception, every point where a human needs to approve. This document is the coworker's job description, and it belongs to you regardless of which platform you run it on.

2. Separate the model from the coworker in your own head. The model is the raw capability. The coworker is the role, the memory, the connections, and the rules. Decide that the second layer will live in infrastructure you control, so the first layer stays swappable.

3. Insist on portable memory from day one. Before you let a coworker accumulate months of context, confirm you can export and inspect that context. If the answer is "it's stored in our proprietary format," treat that as a lock-in cost, not a feature.

4. Build one place to see all of them. Do not let coworkers accumulate one app at a time. Decide up front where the org chart lives — the single surface that shows every coworker, its permissions, and what it can touch. If you're already running more than one, read why running multiple agents without a control layer gets dangerous fast.

5. Keep a human on the seam. OpenAI's own framing — agents that "ask for approval when needed" — is correct. The question is whether the approval logic is something you designed for your risk tolerance or something you inherited from a vendor's defaults. Design the human-in-the-loop points deliberately; don't accept the ones that ship in the box.

The pattern underneath all five steps is the same. Treat the coworker as an employee of your business, not a feature of someone's app. Employees have job descriptions, they keep their knowledge when the tools change, they show up in your org chart, and their manager can see what they're doing. A coworker that can't do those things isn't part of your team. It's a very capable stranger you're renting.

FAQ

Q: Isn't a ChatGPT workspace agent good enough for a small business? A: For a single, contained task with low stakes, it can be genuinely useful — that's the right use of a capable tool. The problem starts when the agent accumulates real business context and you can't move it, or when you're running several agents across several apps and no longer know what any of them can touch. Good enough for one task is not the same as part of your team.

Q: What's the actual difference between an AI coworker and an AI agent inside an app? A: Behavior is the same; ownership is different. Both can follow a process and ask for approval. A coworker lives in infrastructure you control, keeps portable memory, holds a role in your org chart, and survives a model change. An in-app agent lives in the vendor's product, stores memory in their format, and exists only while you pay for that plan. See our AI coworker vs. AI tool breakdown for the full comparison.

Q: Why does model portability matter if OpenAI's models are the best right now? A: "Best right now" has a shelf life of about a quarter. Models leapfrog each other constantly, and the reasons you'd switch — cost, a specific capability, a compliance requirement — are business decisions you want to keep making freely. If your coworker's identity is locked to one model, every one of those decisions becomes a rebuild instead of a config change.

Q: We already have agents in several tools. How do we get control back? A: Start with an inventory: list every agent running anywhere in the business, what it can access, and who owns it. Most businesses can't complete that list, which is the 12% governance problem in miniature. Then consolidate the roles that matter into one operating layer where you can see and govern them together, and retire the scattered ones.

Q: Does portable memory mean I have to manage infrastructure myself? A: No. Portable and self-managed are different things. The point is that your coworker's memory and configuration live in a layer you own and can export from — the hosting can still be handled for you. What you're avoiding is a vendor's proprietary store that you can't inspect or take with you.

Own the Coworker, Not Just the Subscription

The July launch confirmed what the category has been moving toward: AI is becoming a coworker, not a tool. That's the right direction. The open question for every business is where that coworker lives. If you're ready to stop renting agents inside other companies' apps and start running a real team of AI coworkers you actually control — with portable memory, model independence, and one place to govern all of them — Associates AI Teammates gives you a 14-day free trial with no credit card required. Start your free trial at associatesai.team.