Course contents
Rolling Work Out to a Team
Know the controls an admin uses to roll out Work (access, plugins, browsers, local access, approvals, usage limits and Team Tasks) and plan a pilot your admin will approve.
Rolling Work out to a team is mostly about deciding, in advance, what agents may touch and who reviews what they do. As of November 2026, OpenAI's admin docs give workspace owners controls over who can use Work, which plugins and actions are available, whether agents can use a browser, the network or a computer, how often they must ask before acting, and how much they can spend.
You don't need to be an admin to use this lesson. If you're the person proposing a pilot, these are the questions your admin will ask, and knowing the answers is what gets a pilot approved.
The controls, in plain English
| Control | What it decides | Worth knowing |
|---|---|---|
| Work access | Which members, groups or roles can use Work | Grant it to a pilot group first |
| Plugins and apps | Which plugins exist, who can use them, and their action settings | Off by default in Enterprise and Edu; on by default in Business |
| App permissions | How often agents ask before acting in apps: Always ask, Any changes, Important actions or Never ask | Under Any changes, reads run and changes need confirmation |
| Cloud computers | Use a browser, Allow network access and Allow computer use, under Let agents use cloud computers | Enterprise admins must enable the cloud browser |
| Website approval | Which of Always ask, Auto approve and Always allow members can choose | Admins can disable Always allow workspace-wide |
| Local computer access | Whether Work Cloud can use members' connected computers | A workspace owner turns on Allow local computer access |
| Usage limits | Monthly user and group limits, overrides, a workspace overage limit, and alerts | Alerts notify; they don't stop spending |
Two details deserve a second look. A shared or agent-owned connection uses the connected account's permissions, which may exceed what the person asking can access, so limit that account's scopes and choose who can use it. And allowing a website never approves every action on it: consequential actions can still require separate confirmation, whatever the website setting.
Match each admin worry to the control that addresses it.
Pick one on the left, then its partner on the right.
Team Tasks: automation that belongs to the team
Teams give ChatGPT users a shared place to organize work, and Team Tasks let members set up work that runs on a schedule or when an event occurs. Create a team from Settings > Teams > Create team. Create a team task from Scheduled > + New Task with the Team toggle on, then set the trigger, instructions, plugins and model. Use Run now to test and Previous runs to review. Admins decide who can create teams and team tasks through the Create teams and Create and manage team tasks permissions.
Team Tasks run in the cloud with the team's service account and configured app connections, so they don't depend on anyone's laptop. They don't inherit the creator's personal memories, custom instructions or chat history: everything a task needs has to be in its instructions. Unattended runs can't complete a new app sign-in and stay subject to action-approval requirements. Two cautions from OpenAI's docs: connected accounts may access information a member's own account cannot, and new members can see all earlier runs and generated files. Pausing stops future runs, but don't rely on it to interrupt a run that's already going.
A team lead drafted this Team Task for the sales team. Click the settings that create risk.
Plan the pilot
OpenAI's admin guide sets out a rollout in order: assign owners for each area, configure access and identity, set up plugins and connected capabilities, then governance and monitoring. Test with a member who has the intended permissions, and check that an allowed action works and a blocked action fails before anyone relies on the setup. Review each plugin's source, owner, data access and actions, and test with non-sensitive data and the least access that works.
Measure value the way OpenAI recommends: compare a workflow's time, quality and rework before and after adopting Work, and watch usage in the workspace's analytics. OpenAI's own planning example for a marketing team shows how widely usage varies: about $53 per employee a year at the median, $309 on average, and $1,154 at the 90th percentile. Those are planning figures, not prices, and the spread is the lesson: set limits with the heavy users in mind, not the typical one.
Usage alerts fire at 80% of the monthly budget. What happens when the team reaches 100%?
Plain-English definitions from the AI Coding Dictionary.