On the 9th of July, OpenAI put its cards on the table. It announced ChatGPT Work, a new mode that sits beside the normal chat box and the developer-focused Codex, and the whole pitch is that it does not answer you. It goes away, works for hours, and comes back with the finished thing: a spreadsheet, a deck, a document, a small web app. It runs on GPT-5.6, a new model family that reached general availability the same day.
What it actually does
The idea is simple enough to be suspicious of. You hand it an outcome (“check whether we are ready to launch”), it goes off across your connected apps, breaks the job into steps, and grinds through them in the background while you get on with your life. OpenAI showed it off with early users at Nvidia, Zapier and Virgin Atlantic doing jobs like sifting thousands of sales leads and cross-checking a product launch against Jira tickets and go-to-market plans. The agents run in the background and can be shared across a team, which is the bit big companies actually pay for.
The catch for the little guy
Read the name again. It is ChatGPT Work, an enterprise platform aimed at teams with connected tooling and a compliance department, not a magic button on your free account. And “owns your workflow” is marketing for “runs unsupervised for a long time”, which is exactly when a confident wrong answer does the most damage. An agent that spends two hours building a launch-readiness report is also an agent that can spend two hours being wrong in a very tidy-looking spreadsheet. Nobody at Nvidia is signing off on autopilot, and neither should you.
What you can do this week
You do not need the enterprise tier to steal the good idea here. The shift that matters is writing outcomes instead of prompts. Stop asking for “a summary” and start asking for “a three-tab sheet: leads, scored 1 to 5, with the reason in column C”. Give the agent a bounded job with a checkable result, connect only the two or three apps it genuinely needs, and treat whatever comes back as a first draft you sign, not a finished deliverable you forward. If you are on a plan with any agent or background mode, run one real task through it this week and time how long the checking takes. That number, not the demo, tells you whether it is saving you anything.
The honest read: this is a real step, and the background-and-share model is where serious tools are heading. It is also a machine that will do a week of mediocre work in an afternoon if you point it badly. Point it well.
Did you know: the word “robot” comes from the Czech robota, meaning forced labour, coined for a 1920 play in which the artificial workers eventually stop taking orders.