AI agents can do more work. Where does that work go when they stop?

TechCrunch reports that OpenAI released ChatGPT Work last month on its $20-per-month starting subscription tier. The product is meant to take the sort of multi-step work that coding assistants do for engineers and put it in front of people whose days are full of email, spreadsheets, calendars, customer systems and documents.

The company is not hiding the ambition. An OpenAI desktop-app engineer told TechCrunch that his own setup can reach his inbox, Slack, phone, Notion, Figma and more. That may be a reasonable arrangement for someone testing the product. It is not a normal starting point for a person who just wants to stop manually copying school dates or reconciling one weekly report.

A broad permission grant is not the same thing as a useful assistant. It is the beginning of a relationship that needs a shape.

Suppose an AI assistant finds five dates in an email. A success message is not enough. The person needs to know which calendar received them, whether any dates were ambiguous, what it skipped and where to correct a mistake. If the assistant has created a draft, a task, a calendar event and a spreadsheet row across four services, the user should not have to replay its reasoning just to locate the result.

This is where many agent demos get slippery. A scrolling activity feed can prove that something happened. It does not necessarily answer the ordinary questions: What still needs me? What changed outside this screen? Can I close the laptop now?

Give each job a named landing place before it runs. For a calendar task, that could be a reviewable list of proposed events with the source email attached. For a weekly report, it could be one page with the report, source dates, unanswered items and a clear owner. The assistant can work across the mess. The person should not have to.

OpenAI told TechCrunch that more than a billion people use ChatGPT online, while the combined Work and Codex app has 20 million users. Those are company figures, not an apples-to-apples usage study, but the distance between them matters. Plenty of people will ask a chatbot a question. Far fewer will connect it to the records that make a day of work possible.

The reason is not that regular people need a command line. It is that their real work has consequences: a meeting put on the wrong calendar, a private note pulled into the wrong document, a follow-up that looks handled until nobody receives it. The more an assistant can do, the more the end state has to be legible.

The product test is almost embarrassingly plain. After the agent stops, can a tired person find the result, see the unresolved bit and decide what happens next without opening six tabs?

Theo Marlow would keep the 20-million figure in its lane. It refers to a combined application, and it comes from the company through TechCrunch. It shows that an audience exists; it does not tell us how many people gave the tool real permissions, returned after a first run or got time back. Those are different questions.

Jun Vega is less interested in the permission screen than in the next morning. If an assistant put a preschool date on the calendar, a parent should be able to open one place and see the original message, the event, any uncertainty and a fast way to fix it. A feature that saves five minutes on Tuesday but creates a search party on Thursday did not save the household much.

Both points pull against the same lazy conclusion. A powerful assistant is not automatically a settled habit. It earns its place when people can tell what it carried forward and what it left for them.

If you are trying an AI assistant for work automation, choose one job with a concrete finish line: turn one invoice email into a draft record, collect open questions before a client call, or turn a school notice into proposed calendar events. Decide where the result belongs and what must remain a draft before connecting anything broad.

Then test the failure you are actually likely to have. Change one source detail. Remove access halfway through. Ask it to handle an ambiguous date. See whether the result says it stopped, guessed or needs a decision. You do not need a grand personal assistant to learn whether a tool is trustworthy enough for one small chore.

The compelling version of AI help is not a system that disappears into every app. It is one that leaves the work easier to pick up than it found it.