He gave ChatGPT his inbox, Slack, and phone. OpenAI wants your boss to do the same.

Andrew Ambrosino runs OpenAI's desktop app, and he hands the company's new work-tier agent his inbox, his Slack, his Notion, his Figma files, and his phone. "You could imagine a situation," he told TechCrunch, "where the model pulls something from a DM that's not meant to be shared, and surfaces that." He made the trade anyway.

OpenAI is now asking every white-collar worker to make the same one. ChatGPT Work shipped last month at the bottom of the subscription ladder: $20 per month, the same price as the consumer chatbot, no separate seat license, no enterprise sales call required.

The product is a re-skinned version of Codex, pointed at the apps non-engineers already use. Where Codex writes code in a repository, Work reads email, drafts Slack messages, scrolls through documents, and asks for permission before it acts. In its launch post, OpenAI calls the bet a move "beyond answering questions to helping everyone turn their biggest ideas into reality."

Tibault Sottiaux, who leads the core product group that includes Work, told TechCrunch the company is now testing whether the same agentic loop that worked for software engineers can scale to the more permissioned workflows of accountants, investors, doctors, and the rest of the white-collar economy. The shape of that test is what the new product is for.

The commercial case for the bet is straightforward. Coding is a small share of professional work. The big labs have spent the last two years training models that are excellent at it, and the per-user revenue curve on a coder using a coding agent is real but narrow. To justify the next training and compute cycle, OpenAI needs the same usage pattern to cross over into the inbox, the calendar, the spreadsheet, and the contract review.

A longer-running agent burns more tokens. A non-coder who lets an agent draft twenty emails, schedule three meetings, and triage a Slack channel is paying OpenAI, in inference cost, several times what a consumer chatbot user pays for a few questions. Per-user revenue is the variable that has to move. Adoption outside engineering is the only place it can.

A16z has argued that closed providers risk missing agentic value without work-surface access — the economic logic that makes OpenAI's white-collar push commercially consequential. ChatGPT Work is that bet, at $20, on a tier that anyone with a credit card can buy.

The work that has to happen for the test to pass is mostly invisible. The agent has to know which Slack channel is the one where a typo costs a deal, and which is the one where it doesn't. It has to know which Notion page is a draft and which is a contract. It has to know when to ask. Ambrosino, who built the desktop app the agent runs on, told TechCrunch he is the canary: the failure mode he names is the failure mode a non-technical user would never think to ask about.

OpenAI's own evaluation work, branded GDPval, is built around measuring whether models can perform real economically valuable tasks end-to-end, the kind of work that has a market wage attached. A preprint posted in June examined what happens when agentic systems are pointed at those tasks and given long horizons. The short version of the result is that capability scales with horizon, and the worst failures tend to be the ones the user can't see.

That is the rubric a non-engineer reader can actually use. The product is real and cheap. The agent runs on your actual work surface. The engineer who built it is using it on his own, and he has named the worst thing that could happen. The question is not whether the tool works on a benchmark. The question is whether the company you work for will hand it the keys to your inbox, and what you would need to see before you let it.

For now, the answer is sitting at the bottom of OpenAI's pricing page.