ChatGPT agent was retired on 9 July 2026 and replaced by ChatGPT Work, and the answer for Work is: not for anything that sends, pays or deletes, where it stops for your approval by default. For drafting, compiling, research and preparing files, it can work on its own, provided someone reviews the result before it's used.
ChatGPT Work is an agent mode for longer, multi-step tasks across your files and connected apps. It can research, work through websites in its own cloud browser, and produce finished spreadsheets, documents and slide decks. It's included in Plus ($20 a month), Pro, Business and Enterprise at existing plan prices, launched on the GPT-5.6 model family, and shares its usage allowance with Codex, so bigger tasks use more of it. You switch to it with the Chat/Work control above the message box.
So "unsupervised" splits into two questions. Can it do the work while you're elsewhere? Often, yes, including on a schedule. Can it finish the job, with the email sent or the order placed, without you? By design, mostly no, and for a small business that's the right answer.
What ChatGPT Work does without asking, and where it stops
OpenAI's documentation describes a default that reads freely and acts cautiously:
- Reading is automatic. With apps connected (email, calendar, drive, CRM), Work reads from them without asking each time.
- Important actions wait for you. OpenAI defines these as actions that could have a meaningful effect outside ChatGPT, expose sensitive information or be hard to undo, with examples including sending or editing messages, deleting content, making purchases and moving files in cloud storage.
- New websites need a yes. In the browser, ChatGPT asks before visiting a site you haven't allowed, and asks for confirmation before submitting information, making a purchase, changing permissions or deleting data.
- Sign-ins are yours. When a site needs a login, you take over the cloud browser and sign in yourself, rather than pasting a password into the chat. OpenAI notes that website sign-in through the cloud browser isn't available in Enterprise or Edu workspaces.
- It can pause itself. Safety monitoring on the newest model can pause a task it judges potentially unsafe, sometimes after the activity that triggered the pause.
Put together, Work can do most of an admin task unattended and then park the final, consequential step for you. The tutorial on what ChatGPT can see in your drive and inbox covers the reading side in more detail.
Twelve admin tasks, sorted by the supervision they need
| Task | Can it run alone? | What you still do |
|---|---|---|
| Summarise yesterday's emails into a to-do list | Yes | Read the list; nothing leaves the building |
| Compile a weekly bookings or sales summary from a spreadsheet | Yes | Spot-check two or three totals |
| Compare supplier prices from price lists | Yes | Check pack sizes and units match |
| Draft replies to routine enquiries | Drafting, yes | Read and send each reply |
| Draft payment reminders for overdue invoices | Drafting, yes | Approve each send; check for disputes |
| Draft a staff rota from availability notes | Drafting, yes | Approve before it's shared |
| Fill in a supplier's online reorder form | Up to the submit button | Review the basket and submit |
| Research travel for a training course | Research, yes | Book and pay yourself |
| Update contact records in the CRM | Propose changes | Approve the change list |
| Clear out old files or emails | List candidates | Do the deleting |
| Match bank transactions to invoices | From an exported file, yes | Keep it out of your bank login |
| Reply directly to customers | No | Use a proper chatbot with handover rules, or reply yourself |
The pattern: anything that stays inside your business and can be thrown away runs alone; anything that reaches a customer, a supplier or a bank waits for a person.
Four questions that decide how much supervision a task needs
- Can you undo it in five minutes? A draft, a summary or a spreadsheet can be deleted. A sent email, a cancelled booking or a submitted order can't be quietly reversed. If undoing takes longer than five minutes or involves an apology, it needs approval.
- Who sees the result? Internal-only output (you, your team) can run alone. Anything a customer, supplier or regulator sees needs a human read first.
- Does money move? Any payment, order, refund or change to bank details always gets a person, however routine it looks.
- Is the data sensitive? Health details, other people's personal information and financial records deserve extra care, whichever plan you're on, and belong in a Business workspace rather than a personal account.
A task that passes all four (reversible, internal, no money, no sensitive data) is a genuine candidate for unattended running. Fail one and the task is still worth giving to Work, but only up to the point where it hands you a draft.
Setting up Work's first admin task in an afternoon
- Pick one task from the "runs alone" group (15 minutes). A weekly bookings summary or a supplier price comparison is ideal: useful, internal and easy to check against what you'd have produced yourself.
- Create a project for admin (20 minutes). A ChatGPT project keeps related chats, files and instructions together, and its instructions apply to every chat inside it. Upload the current price list and your standard email wording as sources, and write the ground rules once.
- Connect only the apps the task needs (10 minutes). If the job is a bookings summary, connect the drive or spreadsheet, not email and the CRM as well. Fewer connections mean a smaller blast radius and quicker runs.
- Run it by hand three times (about an hour across a week). Switch to Work, run the prompt and compare the result with your own version. Fix the prompt or the project instructions, not just the output.
- Schedule it (5 minutes). Once three runs in a row need no correction, ask Work to repeat it weekly and read the first few scheduled runs closely.
- Book a monthly review (15 minutes). Read one output line by line, confirm the connected apps are still the only ones needed, and retire the task if nobody reads what it produces.
Here are the project instructions a hearing-aid shop wrote for its admin project, filled in rather than left as a blank template:
Admin rules for this shop:
- Never send, pay, delete or change bank details. Draft, then stop.
- Write in British English, short sentences. Sign off "Kind regards, [first name]".
- Customers tagged "phone only" or "do not contact": list them, never draft to them.
- Prices: use only "Price list Sept 2026.pdf" in this project's sources.
If a price isn't there, write PRICE? and flag it.
- Never mention hearing-test results or any health details in a draft.
- End every task with a table: what you did, and anything you weren't sure about.With those rules in place, the shop's monthly aftercare task (list customers due a check-up or a battery refill, and draft reminders) comes back with the phone-only customers already set aside and any missing prices marked, instead of each draft needing the same corrections every month.
A week of admin at a two-electrician firm, handed to Work
The owner of a two-electrician business, in this illustrative case, spends about seven hours a week on admin. After a month of trying ChatGPT Work on a Business workspace with two seats, the week looks like this:
| Job | Before (hours) | What Work does | After (hours) |
|---|---|---|---|
| Inbox triage | 2.0 | Monday and Thursday summaries with suggested replies | 0.75 |
| Quote follow-ups | 1.5 | Drafts follow-ups for quotes with no reply after 5 days | 0.5 |
| Invoice chasing | 1.0 | Drafts reminders, flags any invoice mentioned in a complaint | 0.25 |
| Supplier orders | 1.0 | Compares three wholesalers' prices, fills the basket, stops before paying | 0.25 |
| Timesheets to payroll sheet | 1.0 | Compiles the sheet from job notes for the owner to check | 0.25 |
| Certificate renewal reminders | 0.5 | Lists customers due within 60 days and drafts emails | 0.1 |
| Total | 7.0 | 2.1 |
The saving is roughly five hours a week, but notice where the remaining two hours go: reading, approving and fixing. That review time is the price of never letting Work send, pay or delete alone. The plan cost is two Business seats at $25 each billed monthly ($50) or $20 each billed annually ($40), and Business data isn't used for training by default. A sole trader could use Plus at $20 a month instead, but should switch off "Improve the model for everyone" under Settings, Data controls, because on Free, Plus and Pro that setting is on unless you turn it off.
Prompts that keep Work inside the lines
Work behaves best when the prompt states the outcome, the sources, the limits and the stopping point. OpenAI's own guidance suggests boundaries such as "use only these sources" and "stop before sending anything". Here's an invoice-chasing prompt built that way:
Use @Gmail and the overdue-invoices export attached. Find invoices more than 21 days overdue.
For each one, search Gmail for the invoice number and the customer's
name from the last 60 days. If any email mentions a dispute, a
complaint or a promised payment date, do not draft a reminder; list it
under "Needs a person" with a one-line reason.
For the rest, draft a reminder in my usual tone: short, polite, with the
invoice number, amount and due date, and the payment link from the invoice.
Show each draft here for me to review. Do not send anything.
Finish with a table: customer, invoice, amount, days overdue, action taken.The first run's output (illustrative) looked like this:
Drafted: 7 reminders (total $4,380)
Needs a person:
- INV-1182, $610: customer emailed on 12 Sept saying the outdoor
socket was faulty and asked for a return visit.
- INV-1190, $295: customer promised payment "after the 30th".
Not found in Gmail: INV-1175 (no matching emails, reminder drafted).Two fixes followed. INV-1175 belonged to a customer who only ever phones, so "no emails" didn't mean "no dispute"; the owner added "flag any customer tagged phone-only in the accounting export". And one draft used the customer's company name where the owner always writes to a person, so the prompt gained "address the contact person by first name, e.g. Hi [first name]". Small edits like these are the whole job of supervision in the first few weeks.
Scheduled tasks: the closest Work gets to running unattended
Scheduled tasks let Work repeat a job on a timetable. On the web they can use your uploaded files, connected apps and plugins; on eligible plans they can also start when a Gmail, Slack or GitHub event happens. In the desktop app, tasks that need local files require the computer to be on and the app running, while cloud tasks keep going after you close it. OpenAI's advice is to test the prompt in an ordinary chat first and review the first few runs before trusting the schedule.
A pharmacy might schedule this for 7am every Monday: "Read last week's emails from our three wholesalers in @Gmail. List price changes, discontinued lines and items out of stock in a spreadsheet, and draft this week's reorder list. Send me a summary; don't contact any supplier." The pharmacist reads the summary with their first coffee and places the orders themselves. If you're setting these up for reminders and routine checks, the guide to ChatGPT scheduled tasks walks through them. One housekeeping point: OpenAI retires GPT-5.5 from ChatGPT and Work on 14 October 2026, so any scheduled task set to that model needs switching before then or it stops.
Three ways owners get burned by Work
Switching off the prompts because they're annoying. The website setting can be loosened so Work stops asking before visiting new sites. An osteopath who did this to speed up supply ordering later found Work had filled in the reorder form on a lookalike supplier site that ranked highly in search. Nothing was paid, because payment still needed approval, but the clinic's account details had been typed into a stranger's form. Keep the website review on and allow your regular suppliers individually.
Approving on reflex. After a fortnight of good drafts, it's tempting to approve a queue without reading it. That's exactly when a reminder goes to a customer whose complaint was sitting in a different inbox. If you're approving more than about 15 items in one sitting, split the task.
Giving it a shared login. Pasting the office's shared password into the chat, or letting several staff run Work from one personal account, leaves no clear record of who approved what. Use your own sign-in when Work asks you to take over the browser, and give each member of staff their own seat. The tutorial on what can go wrong when agents take actions covers these failure modes in more depth.
Admin jobs Work handles badly, even with a person watching
- Jobs that are identical every time. Copying new bookings into a sheet or sending a fixed confirmation is cheaper and more predictable as a plain automation in Zapier or Make. An agent that plans its own steps adds variation where you want none; agent vs chatbot vs automation explains how to tell the difference.
- Jobs whose context lives on the phone or on paper. If the only record of a customer's complaint is a phone call, Work can't know about it, and its drafts will read as tone-deaf. Log calls somewhere Work can read, or keep those customers out of its tasks.
- Systems with no connection. Work can operate a website through its cloud browser, but it's slower than a connected app, and OpenAI notes that a site may block automated access or not support a sign-in or transaction step, leaving you to take over or finish the last step yourself.
- Tasks with no finish line. "Keep my inbox under control" never ends, so it's hard to check. "Each Monday, summarise last week's unanswered enquiries and draft replies" has an edge, an output and an obvious way to judge it.
If a task falls into one of these, don't force it. Either fix the underlying process (log the calls, tidy the spreadsheet) or leave the job with a person.
Plus, Business or Pro for running admin through Work
| Plan | List price | Fits | Watch out for |
|---|---|---|---|
| Plus | $20 a month | A sole trader doing their own admin | Model training is on until you switch it off; no admin controls |
| Business | $25 a seat monthly, or $20 billed annually; minimum 2 seats | Two or more people, or any client data | No training on business data by default; a one-person firm still pays for two seats |
| Pro | $100 or $200 a month | One heavy user running long Work tasks daily | A personal plan: switch training off, and it has no team controls |
For most small businesses with staff, Business is the sensible home for Work, because admin tasks inevitably touch customer names, invoices and emails. The comparison of ChatGPT Plus and ChatGPT Business sets out the other differences. Whichever plan you choose, the working rule stays the same: let Work do the reading, sorting, compiling and drafting on its own, and keep your hand on anything that sends, pays or deletes.
ChatGPT Work for admin: follow-up questions
Where did ChatGPT agent's features go?
OpenAI's help page for ChatGPT agent now says it is no longer available and points users to ChatGPT Work for longer, multi-step tasks. Work covers the same ground, including a cloud browser that can read pages and fill in forms, and adds finished files such as spreadsheets, documents and slide decks, plus scheduled tasks that repeat on a timetable.
Can ChatGPT Work use files and apps on my own computer?
In the ChatGPT desktop app, yes. OpenAI's documentation says Work can use local files, apps and the browser there, and suggests keeping it set to work locally when a task needs things on your computer. Work on the web runs in the cloud instead and can't reach your computer's files, so you upload them or connect an app. Keep the computer on for local scheduled tasks.
Can I use ChatGPT Work on the free plan?
Only in a limited form. OpenAI's documentation lists Work on the web and mobile apps for Plus, Pro, Business, Enterprise and Edu, while Free and Go get a lighter model in the desktop app, subject to rollout. For regular admin you will want a paid plan, and for staff handling client data, a Business workspace.
Should I give ChatGPT Work my online banking login?
No. Keep bank and card logins away from any browser agent; Anthropic's guidance for its own browser agent lists managing financial accounts as something to avoid, and the same logic applies here. Export a statement file instead and let Work match transactions against invoices, then make any payments yourself.
Further reads
- How a Plumber Can Save Five Hours a Week on Admin With AI — Another trades business mapping five hours of weekly admin to AI.
- How Freelancers Use AI for Invoices, Proposals and Chasers — Prompts for invoices, proposals and chasers you can reuse in Work.
- How to Pilot Your First AI Agent Without Risking Customers — A step-by-step way to trial Work on one task before widening it.
- How to Connect ChatGPT or Claude to Your Business Apps — How to connect ChatGPT to the apps your admin lives in.
- Is ChatGPT Safe for Business Use? Risks, Settings and Plan Choice — Plan choice and privacy settings before staff use ChatGPT for work.
- What Is an AI Agent? A Plain-English Guide for Business Owners — Background on how agents plan and act, in plain English.
- What Are AI Marketing Agents and Should a Small Business Use One? — What separates a marketing agent from a chatbot or automation, the agents small businesses can switch on now, what they cost and the guardrails to set.
- How Much Does an AI Agent Cost a Small Business? — The five ways AI agents are priced, the cost-per-task sum that makes them comparable, and the supervision time that usually costs more than the software.
- What Is an AI Browser Agent, and Should Staff Use One? — What an AI browser agent does, which jobs it suits, how prompt injection can hijack it, and the settings and staff rules that keep mistakes small.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help centre pages on ChatGPT agent, ChatGPT Work, app permissions and data controls; OpenAI's ChatGPT Learn documentation on getting started with Work, browser, scheduled tasks, pricing and Work security; Anthropic's Use Claude in Chrome safely page (all checked September 2026).