All posts
Read time 2 min

ChatGPT wants to do your work. Work is really selling a better task contract.

ChatGPT Work can connect apps and files, break down large tasks, and keep working for hours. Its real significance is that AI assistants now need the same permissions, boundaries, and acceptance criteria as coworkers.

Abstract AI work orchestration with human approval

OpenAI launched ChatGPT Work in July 2026 as an agent for longer, multi-step work rather than another chat model. The official product page describes a system that gathers context from files, the web, connected tools, and desktop apps, proposes a plan, and then creates deliverables such as documents, spreadsheets, presentations, or Sites.

Work is not simply a longer chat

Chat is for questions and quick discussion. Work is for tasks that combine several steps, sources, and finished deliverables. Codex remains the dedicated software-development surface. The distinction is less about which model is “smarter” and more about tools, run time, and the type of output.

Work can gather context, ask for missing information, and present a plan before execution. That does not guarantee a correct result, but it gives the user a chance to narrow the job before a long run begins.

Availability differs by surface

OpenAI's Help Center says Work is gradually rolling out to eligible accounts. Web and mobile Work run in the cloud. Desktop Work can also access local files and desktop apps after the user grants permission. At launch, cloud Work conversations do not appear in desktop Work, and desktop threads and local files remain on that computer.

“Available to all plans” therefore does not mean every account and platform receives identical features on the same day. A team rollout should verify plan eligibility, region, workspace policy, and connector permissions.

Write a task contract, not a vague prompt

A useful Work brief should state:

  1. the file, page, or other deliverable required;
  2. the sources it may read;
  3. the data and applications it must not access;
  4. the actions that require human approval; and
  5. the acceptance criteria for the result.

Good automation candidates are repeatable, reviewable, and easy to reverse: weekly summaries, material organization, and first drafts. Payments, outbound messages, permission changes, and deletion should not become unattended merely because the agent can run for hours.

Sources

Related