ChatGPT Work Turns the Chatbot Into a Worker You Delegate To

OpenAI has launched ChatGPT Work, a new agent inside ChatGPT that takes action across your apps and files, stays with a project for hours, and turns a goal into finished work. It is powered by Codex and GPT-5.6. For growth leaders, this is the moment the chatbot stops being a thing you prompt and becomes a thing you delegate to.

The distinction matters. You no longer ask for an answer and copy it into a document. You describe an outcome, and the agent produces the document, deck, analysis, site, or report, drawing context from the apps and files you connect.

What ChatGPT Work actually does

According to the OpenAI announcement on X, ChatGPT Work can take on entire workflows from a single request. It understands the goal, uses context from selected apps and files, and keeps work moving while you stay in control.

The GPT-5.6 model behind it is built to reason through complex, multi-step tasks and to match your templates, reference files, and preferred style. You describe the result you want without spelling out every step. That is the practical difference between an assistant and an operator.

When it rolled out

ChatGPT Work went live on 9 July 2026. On web and mobile it rolled out that day to Pro, Enterprise, and Edu plans, with Plus and Business plans following over the next few days.

In the desktop app, Chat, Work, and Codex are available on every plan, including Free, globally on Windows and Mac. Existing Codex app users update as usual, and their app becomes the new ChatGPT desktop app.

How the agent mechanism works

Think of it as a persistent worker rather than a single reply. The loop looks like this:

1. Goal: you state the outcome, for example a competitor teardown deck or a monthly performance report.
2. Context: the agent reads the apps and files you have connected to ground its work in your real data.
3. Execution: it takes actions across those tools, holding the task for hours if needed.
4. Output: it hands back a polished, template-matched artefact you review.

Where it fits in a marketing and ops stack

The obvious fits are the repetitive builds that eat analyst hours: weekly reporting, campaign briefs, audience research syntheses, and first drafts of landing pages. This is the same agentic direction we covered in GPT-5.6 landing inside Microsoft 365 Copilot, and it reinforces a theme from why your data needs more work than your vendor list: agents are only as good as the context you feed them.

The governance to lock down first

An agent that acts across live accounts needs guardrails before, not after, deployment. Set these three:

• Scoped access: connect read-only or sandbox environments before granting write access to ad accounts, CMS, or CRM.
• Human sign-off: require review on anything that publishes, spends, or emails.
• Audit trail: log what the agent did and where its inputs came from, so output is traceable.

OpenAI's own product materials stress that you stay in control. Treat that as your obligation, not just a reassurance.

The concrete action

Pick one bounded, low-risk workflow this week: a report or a research summary that does not touch spend. Run it through ChatGPT Work with read-only access, then judge the output against your best analyst. That test tells you exactly what to delegate next.