ChatGPT Work: A Practical Guide for Teams

ChatGPT Work is the agent mode inside ChatGPT, built for multi-step business tasks. OpenAI launched it in July 2026. You give it a goal plus the files or apps it needs, and it works through the job on its own until it has something finished to hand back: a report, a spreadsheet, a slide deck, a document or a small web app.

ChatGPT is now split into three surfaces. Chat is for questions and quick help, Work is for longer jobs that end in a deliverable, and Codex is for software development. Work is available on every plan in the desktop app, including Free with limited usage, and on the web and mobile apps for paid plans.

People searching for "ChatGPT for work" often mean something else: the business plans. ChatGPT Business is for teams and ChatGPT Enterprise is for large organisations. Both keep company data out of model training by default and add admin controls. I cover both below, because the plan you pick decides how much control your company has over what Work does.

What Work does differently from Chat

Chat answers one message at a time. Work plans a task, breaks it into steps, uses tools, and keeps going until the job is done or it needs your input. OpenAI's help center describes it as the place to research a topic, analyse information, or create a document, spreadsheet, presentation, report or Site.

It launched alongside GPT-5.6 and a new desktop app that merged Codex into ChatGPT. The old desktop app was renamed ChatGPT Classic. A typical run looks like this:

You describe the goal and attach context
        ↓
Work proposes a plan (Plan mode) and you approve it
        ↓
It reads files, searches, uses connected apps
        ↓
It pauses for approval before sensitive actions
        ↓
You get a finished deliverable to review

Zapier put it well: Work is basically Codex for people who don't write code. The same agent engine sits underneath, minus the terminal and git. For developers, that framing tells you where the line is. Code changes, diffs and tests stay in Codex; the documents around the code go to Work.

The tools it can use

Work connects to business tools through plugins, including Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs and project trackers. ChatGPT picks a plugin automatically, or you can name one by typing @ followed by the app name. If an app is missing, you can add a custom MCP server.

Beyond plugins, a few capabilities are worth knowing by name:

• Plan mode makes Work propose a strategy first, which you approve or adjust before anything real happens. Use it for anything that touches shared files or other people.
• Scheduled Tasks run a job once, on a recurring schedule, when an event happens, or as ongoing monitoring. A daily briefing at 8 a.m. or a Friday status report are the obvious uses.
• Sites (public beta at launch) turns output into a shareable interactive page or web app that can update as the underlying data changes.
• Computer use lets Work operate your machine from the desktop app: clicking, typing and moving files.
• A built-in browser pulls in websites and online files. OpenAI also released a Chrome extension and said it will sunset the standalone Atlas browser.
• Projects give every chat inside them the same files, context and instructions, so a client or a recurring report doesn't need its background repeated each time.

Work also stops and asks for explicit approval before actions such as emailing your team or editing a shared file, and OpenAI added an automatic review layer that checks important actions before they run.

Desktop or web: where Work runs changes what it can touch

This is the part that confuses most new users. The same mode behaves quite differently depending on where you start it.

One thing the table can't show: even when Work runs locally, messages and task context can still be stored in the cloud. "Local" does not mean nothing leaves your machine.

On desktop, the default permission setting asks before Work runs commands outside its workspace or connects to the internet. The more permissive options, "Approve for Me" and "Full Access", trade safety for speed. I'd leave the default in place and only loosen it for trusted, repeatable tasks.

Which ChatGPT plan you need for Work

OpenAI changes plans and limits often, so check the official pricing page before you buy. This is the lineup as reported in October 2026.

Work is metered like Codex. A complex task uses far more of your plan's included usage than a short chat, so heavy Work users hit Plus limits quickly. GPT-6.1 Terra has also been added to the models available in Work. Business plans offer a higher-usage seat type, reported as Business Premium; confirm the current seat options with OpenAI sales.

If you'll use Work with company data, the deciding question between a personal and a business plan is control, not features. A personal Plus or Pro account belongs to you, not your employer. Your company can't manage it, audit it, or remove access when you leave. Business (the plan formerly called ChatGPT Team) and Enterprise put the account under the company's admin console and keep business data out of training by default.

What admins can control

Business covers the basics a team needs: a shared workspace with an admin console and user management, SAML single sign-on and multi-factor authentication, no training on workspace data by default, encryption at rest and in transit, SOC 2 Type 2 compliance, and admin control over which apps and connectors members can use.

Enterprise adds the things a security or compliance team will ask for:

• SCIM user provisioning and role-based access control
• Custom data retention and data residency options
• Enterprise Key Management (EKM) for customer-managed encryption keys
• A Compliance API for oversight of conversations and agent actions at scale
• Priority support with an SLA

For Work specifically, OpenAI said at launch, as The Next Web reported, that Enterprise and Edu admins can decide which users get Work, what company context the model can see, and which tools and plugins members can connect. They can also limit which actions are permitted, including browser use, network access and sensitive actions on the web. Codex's governance of local files, apps and browsers extends to Work on the desktop app, and the Admin Console has spend controls, group limits and individual overrides.

On personal plans (Free, Go, Plus and Pro), you can stop your conversations from being used to improve models in Settings → Data controls, and temporary chats aren't saved to your history. That's fine for your own material. For anything confidential that belongs to your employer, use a company-managed account rather than relying on personal settings.

Jobs where Work earns its allowance

Because Work is metered, save it for multi-step tasks that end in a file you can review. A quick question belongs in Chat.

A weekly status report built from raw data is the canonical example. Notice that the prompt reads like a brief to a colleague, not a question:

Goal: Build this week's sales status report.
Inputs: /Reports/sales_week44.xlsx and last week's report (attached).
Steps: Clean the data, compare with last week, flag any region down more than 10%.
Output: A 2-page document and a 6-slide deck in the same style as last week's.
Ask me before emailing anything.

A daily briefing works well as a scheduled task:

Every weekday at 8:00:
- Summarise unread emails from my manager and key clients
- List today's meetings with one line of context each
- Check the project tracker for tasks due in the next 48 hours
Send the briefing to me only.

Other jobs that suit it:

• Turning a meeting transcript into decisions, owners and deadlines, with draft follow-up emails. Keep sending as an approval step so you see every email first.
• A two-page research brief on a vendor, market or regulation, with a link for every claim. Open the important links yourself, because agents still misread sources.
• Cleaning an export with messy dates, duplicates and mixed currencies. Ask for a cleaned sheet, a summary tab, and a note listing every change it made, so the clean-up can be audited.
• For developers, the paperwork around code: release notes from merged pull requests, an incident post-mortem from logs and chat history, a migration plan for stakeholders, or a status page as a Site.

For how the newest models handle the code itself, see GPT-6 Astra for C# development.

How to brief it

The prompts above share a shape worth copying. State the goal as what "done" looks like. Name the exact files, folders or apps to use, and which to ignore. Give the constraints: format, length, tone, deadline, and anything it must not touch. Say when it should stop and ask you. Then specify the output's file type, name and save location. Focused prompts that include only the relevant files also use less of your allowance.

Where people get burned

Most problems with Work come from trusting it too much, too early. Giving it full access on day one is the big one: more access means more damage when something goes wrong, so start with default permissions and narrow folders.

Prompt injection is the less obvious risk. A malicious web page, email or document can contain instructions aimed at the agent rather than at you. Keep approval switched on for sending, deleting and sharing, and back up important folders.

A few smaller traps catch people regularly. Slides and documents from Work are good first drafts, not finished work, so check the numbers and polish the design yourself. Desktop projects that use local files stay on that machine and won't follow you to the web. And a personal account is the wrong place for company data, because your employer can't control retention, access or training settings there.

How it compares with other work agents

Work competes with Anthropic's Claude (which folded its Cowork agent mode into the main Claude app), Microsoft Copilot inside Microsoft 365, and Perplexity's Computer and Personal Computer agents. The right choice usually depends on where your company's files and identity already live, more than on any single feature. For a head-to-head on the agent side, read Perplexity vs ChatGPT computer agents.

If you're evaluating Work for a team, start on a Business workspace with default permissions, pick one recurring report as the pilot, and keep every approval step on until you've reviewed a few weeks of output.