What is ChatGPT Work, and is it right for your team?

ChatGPT Work, launched by OpenAI on July 9, is not a new subscription. It is a way of working inside ChatGPT: instead of answering one message at a time, an agent (an AI system that pursues a goal over many steps and chooses its own tools along the way) takes a task, breaks it into smaller steps and carries them out. It runs on GPT-5.6 and borrows the technology of Codex, OpenAI's coding agent.

The output is meant to be finished work rather than a chat reply: a budget variance analysis in a spreadsheet, a campaign brief, a slide deck that follows your company template, or a shared web page. You can watch its progress, answer its questions, change direction and approve important actions before they happen.

What it connects to and what it can do

ChatGPT Work gets its information through plugins, connections to tools such as Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars and customer-relationship systems. It decides which plugin to use from your request, or you can point it at one by typing @ and the app's name.

Three features extend it:

Scheduled Tasks run a job once, on a repeating schedule, when an event occurs, or to watch for changes, for example refreshing a weekly meeting agenda from new Slack messages.
Sites in ChatGPT, in public beta, turns a result into an interactive page or small web app, such as a project tracker, that you share by link.
On the desktop app for Mac and Windows, it can use a built-in browser and Computer Use, meaning it can click, type and move files in your own applications.

Because a single task can run for a long time, usage is counted differently from ordinary chat. OpenAI says it is metered the way Codex is, and bigger tasks draw more from a plan's included allowance.

Is it right for your team? A quick check

The fit is best when the work is repeatable, spread across several tools, and easy to judge when it is done. Use this table to sort your candidate tasks.

Question	Good sign	Warning sign
Do you do this task often?	Weekly or monthly	Once a year
Is the information in connected tools?	Drive, Slack, email, a CRM	Paper, phone calls, people's heads
Can you tell a good result quickly?	You have past versions to compare	Quality is a matter of taste or debate
What happens if it is wrong?	A draft someone reviews	Money moves or a client receives it directly
Who owns the result?	A named person checks it	Nobody in particular

Two or more warning signs on a task suggest keeping it manual for now, or using ChatGPT Work only to prepare a draft.

A worked example: a two-week pilot

OpenAI suggests learning the tool on a task you already know well. For a team, that becomes a short, controlled pilot.

Pick two tasks from the table with mostly good signs, such as a monthly expense summary and a weekly client status report.
Connect only the plugins those tasks need. If the report uses Drive and Slack, leave email and the CRM off.
Write the request with a stop point:

Prompt

Using the "Client status" folder in Google Drive and this week's
#client-updates channel in Slack, draft the weekly status report in
the same format as last week's report in that folder. Flag anything
you could not confirm. Do not send or share anything; leave the draft
for my review.

Compare every draft with the version a person would have written, and record time saved and errors found.
Only then add a Scheduled Task, and keep reviewing the first few automatic runs before trusting it.

Try it: List three tasks your team repeats every month and score each against the five questions in the table. Start your pilot with the one that has the most good signs.

Settings an administrator should look at

OpenAI's announcement describes central controls for Enterprise and Edu workspaces. Admins can decide who gets access, which company information is in reach, which tools may be connected and what the agent is allowed to do. On the web, they can manage plugins, set rules for browsing and network access, and block risky actions in linked systems. In the desktop app, ChatGPT Work follows the governance model already used for Codex.

Two further layers matter for cautious teams. Auto-review uses OpenAI's strongest models to check important actions in connected tools before they run, to catch unauthorized sharing of sensitive data. Spend controls in the Admin Console let admins set workspace defaults, limits for groups and individual exceptions, and review requests for more credits. The Compliance API gives oversight of ChatGPT Work conversations and actions across the organization.

How is it different from Team and Enterprise?

They answer different questions. Team and Enterprise are plans: they decide who pays, how the workspace is managed and what protections apply. ChatGPT Work is a feature that runs inside those plans and others. First, a naming point: since August 29, 2025, the plan once called ChatGPT Team has been named ChatGPT Business. OpenAI described it as a name change only, with the same features, prices and limits, and business data excluded from model training by default.

What changes from plan to plan is how you get ChatGPT Work and how much control you have over it. On web and mobile it reached Pro, Enterprise and Edu first, with Plus and Business following within days of launch. The desktop app offers Work, alongside Chat and Codex, to every plan, Free included. Free and Go users work with the GPT-5.6 Terra model, while paid plans, including Business and Enterprise, can pick any of the three tiers and adjust the effort level. Some extras are limited to higher plans: the "ultra" setting, which runs several agents in parallel, is available in ChatGPT Work to Pro and Enterprise.

The biggest difference for a growing team is governance. The detailed admin controls, spend limits and compliance tools described at launch are presented for Enterprise and Edu. A small team on Business can still benefit from ChatGPT Work, but it should check which of those controls appear in its own admin settings, and set clear house rules about plugins and approvals, before connecting sensitive systems.