Training

ChatGPT agent mode: what it can do in your work, and what not to let it do

ChatGPT agent mode no longer exists under that name, because OpenAI has replaced it with ChatGPT Work. Here is what it does, which tasks your team can hand over, which to keep and which controls to switch on.

What ChatGPT agent mode was, and what replaced it

Plenty of people still look for the ChatGPT agent mode button in the menu, and in November 2026 it is gone. OpenAI's Help Centre says ChatGPT agent is no longer available and that ChatGPT Work is the option for longer, multi-step tasks with a finished deliverable.

ChatGPT agent arrived in 2025. You could hand it a whole task, such as finding suppliers, comparing prices and building a table, and the assistant browsed and used tools on its own. Work, launched on 9 July 2026 according to OpenAI's release notes, takes that idea further. It researches and analyses information, works across connected apps and files, and delivers finished documents, spreadsheets, presentations or reports.

According to the same notes, Work is available on web and mobile on paid plans other than Go, and in the desktop app when your plan and workspace include it. It shares one usage allowance with Codex. On Plus and Business, that usage is capped per five-hour window and per week.

Tasks you can delegate at work

It works best on tasks with a clear outcome, a few sources and a deliverable someone can review.

  • Research with a report

    Gather public information on five competitors, their prices and terms, and deliver a table with the source of each figure.

  • Documents from your files

    Turn meeting minutes, a spreadsheet and two emails into a board presentation or a report in the company's format.

  • Data analysis

    Clean a CRM export, group sales by region and flag anomalies, with the formulas visible so you can check them.

  • Scheduled tasks

    Prepare a weekly round-up of industry news or order status and leave it as a draft for someone to review before it is shared.

  • Meeting preparation

    Read a client's history in the connected apps and write a brief with the context and open issues.

What not to let it do

These are sensible red lines for a small or mid-sized company, even where the tool is able to cross them.

  • Sending messages to clients

    It can draft the email. A person sends anything to clients, suppliers or public bodies. Put it in the instruction: "stop before sending anything".

  • Payments, purchases and contracts

    Actions with financial or legal consequences stay out. OpenAI provides a separate confirmation for these cases, and your company rule should still be that nobody delegates them.

  • Accounts with sensitive access

    Keep it away from online banking, government portals and your website's admin panel. In Enterprise and Edu workspaces, signing in to websites through the cloud browser is not available.

  • Sensitive personal data

    Payroll, health, job applicants or client files should not go through a delegated task without a clear legal basis and without knowing which app receives the data.

  • Hidden instructions in what it reads

    A web page or a document can contain text written to give the assistant orders, known as prompt injection. The more it acts on its own, the more it matters to limit which sites it visits and what it can do there.

The controls worth knowing

Based on OpenAI's documentation in November 2026. Setting names can change.

  • Permission per website

    By default, ChatGPT asks before visiting a new website. Users can choose to always ask, auto-approve or always allow. Allowing a site does not approve every action on it.

  • Confirmation of sensitive actions

    ChatGPT can ask for a separate confirmation before an action with financial, legal or account consequences. While it works, you follow its progress, answer its questions, change direction and approve important steps.

  • Connected apps

    In Business workspaces, apps are on by default, and in Enterprise and Edu they are off. Admins can restrict an app to read-only or to approved actions.

  • Computer access

    For Work to use files and apps on a computer, the workspace owner must enable local access. Without it, the task runs in the cloud and cannot see local files.

The ten-minute test before you give the ChatGPT agent a task

Run it on every new task before suggesting it to the rest of the team.

  1. 01

    Write the expected outcome (2 minutes)

    One sentence with the deliverable, the format and who it is for. If you cannot write it, the task is not ready to delegate.

  2. 02

    Mark where it must stop (1 minute)

    Add stopping points to the instruction: before sending, buying, deleting or sharing anything outside the company.

  3. 03

    Run it on dummy data (4 minutes)

    Use a made-up client or a file with no personal data. Watch which sites it asks to visit and which apps it uses.

  4. 04

    Check the deliverable (2 minutes)

    Check three figures at random against their source. If one is wrong, the task will need systematic human review.

  5. 05

    Give it a colour (1 minute)

    Green, it is delegated with a final review. Amber, only the research and the draft are delegated. Red, the task stays with a person.

Example

A Toulouse office supplies distributor hands over its weekly report

In this typical case, an office supplies distributor in Toulouse with fifteen staff uses ChatGPT Business. Every Friday, the sales manager spends the morning on a report covering the week's sales, open orders and three competitors' prices.

He runs the task through the ten-minute test with a sales export stripped of client names. Work produces the table and a four-page presentation, but mixes up two product codes. He rates it amber: Work gathers the data and prepares the draft, and he checks the figures.

The workspace admin connects only the sales sheet and the reports folder, read-only. The scheduled instruction ends with "leave the report as a draft and do not share it". Visits to competitors' websites are approved site by site.

To measure it, he logs for two months the time he spends on the report, review included, and how many errors he corrects each week and of what type. If errors fall and the review gets shorter, he extends the task to open orders.

Before you open Work to the whole team

With a feature like Work, problems usually come from habit: someone approves a site without looking, connects their whole inbox or delegates a send because they are in a hurry. Written rules and practice on real cases cut those mistakes.

Obinova is an AI agency based in Barcelona and Paris. Our ChatGPT training starts from your team's own tasks: what to delegate to Work, which controls to switch on, how to write instructions with stopping points and how to check the result. We also cover Claude, Gemini and Microsoft Copilot, in person at your company or on our dedicated video platform.

If a task needs several internal tools connected with fixed rules, our n8n automation service builds custom workflows and agents, documented so your team can keep improving them.

Frequently asked questions

How do I turn on ChatGPT agent mode?

You can't any more, because OpenAI has withdrawn it. Its successor is ChatGPT Work: you switch to Work and describe the task and the outcome you expect. In a company workspace, access and controls depend on how the admin has set things up.

What is the difference between ChatGPT agent and ChatGPT Work?

Work is what OpenAI now offers for the tasks the agent used to handle. It is built to deliver a finished result, such as a document, a spreadsheet or a presentation, and it works with your files, connected apps and a cloud browser. It can also run scheduled tasks that repeat or watch for changes.

Is the ChatGPT agent safe to use in a company?

It can be, with clear rules. Connect only the apps you need, read-only where possible, keep approval on for every new website and write in each instruction where it must stop. Sending, payments and sensitive data stay with a person. The workspace admin decides which apps and what local access are allowed.

How much does ChatGPT Work cost?

There is no separate price. It is included in paid ChatGPT plans other than Go, and its usage comes from the same allowance as Codex. On Plus and Business it is capped per five-hour window and per week, and on Enterprise the admin can set per-user limits. Check your consumption in your account's usage settings.

Which tasks should you not delegate to an AI agent?

Anything that cannot be undone or that commits the company: messages to clients, payments, purchases, signatures, deleting data or decisions about people. It is also wise to keep it from signing in to sensitive accounts. For those tasks, the agent can prepare the draft and a person decides.

Let's decide what your team can hand over to ChatGPT

Book a free 30-minute video call. We run two or three of your team's tasks through the ten-minute test and work out what training each person needs.

Free 30-minute diagnosis →
Free diagnosisWhatsApp

WhatsApp · Reply within 2 hours during business hours