agentusecasesAll 965 use cases
Agents

What people do with Operator

OpenAI's browser-using agent. 14 real use cases so far, most often for travel and shopping. Each links to its source and comes with a prompt to try.

Also called OpenAI Operator, ChatGPT Operator.

14 use cases · 7 kinds of job · Updated Oct 8, 2026

All 14 Operator use cases

  • Reserve a restaurant table through OpenTable

    In a demo, Operator booked a 7 p.m. table at Beretta in San Francisco via OpenTable, first searching the wrong region before correcting and asking for approval.

    How it went Operator found the right setting without help and stopped for approval before booking. It did not find the restaurant on the first search. This was a staged demo, not a real-world test.

  • Book flights with AI trip planners

    A blog tested five AI trip planners on completing flight bookings; Kayak's agent finished through its own checkout, while ChatGPT Operator only partly completed bookings via form-filling and needed manual cancellation.

    How it went Kayak completed the booking, though changes and cancellations go through its support flow. Operator failed in two of three runs because of session timeouts, CAPTCHAs and price changes. Gemini and Mindtrip only redirected.

  • Find London tours and order groceries

    Operator found top-rated London walking tours on TripAdvisor, but its Instacart order went wrong by searching in the wrong city.

    How it went The tour search took about a minute and worked. The Instacart task failed, and Newton finished it himself. Adding bananas, seltzer and raspberries took 15 minutes. Free ChatGPT gave a comparable tour list faster.

  • Make a dinner reservation with Operator

    Marily recorded her first try using Operator to make a dinner reservation, liking that it skipped searching and detail entry, but found it slower than expected.

    How it went The reservation got made and she called it pretty amazing. It was slower than she expected, though it didn't bother her much. She wanted it to be more proactive, with memory, and wanted voice interaction.

  • Find a toy and start checkout with Operator

    OpenAI's Operator picked a toy matching criteria and opened Target.com but did not complete the purchase, walking the user through checkout.

    How it went Operator found a toy that met her criteria, but it did not complete the purchase. It led her through checkout, while Perplexity in the same post did complete a purchase on its own.

  • Order pens and search for reservations and resorts

    Operator found a 144-pack of pens on Staples after Amazon blocked it; restaurant booking failed and the resort search chose random dates.

    How it went It found 144 BIC blue pens for $23.39, and the author finished the payment. Restaurant booking failed: wrong location at first, then blocked pages. The resort search picked random dates and locations without asking questions.

  • Order groceries with computer-use agents

    The author tried Operator and two other computer-use agents on grocery shopping and found the experience too poor for ordinary consumers.

    How it went Operator misread the list, then added only 13 of 16 items. The second attempt took over 30 minutes and several interventions, versus under four minutes by hand. The two other agents did worse.

  • Handle moving errands and reservations

    A TechCrunch reporter used Operator for a week; a restaurant reservation needed many interventions and an electrician booking failed.

    How it went It found a suitable restaurant but needed more than half a dozen answers along the way. TaskRabbit blocked it, and it hallucinated garage distances (20 and 30 minutes' walk) after entering a wrong address.

    2 accounts

  • Test Operator on legal paperwork, party planning and shopping

    The author tried Operator on five use cases; party planning was 'meh' and Amazon shopping failed because it couldn't add items to the cart.

    How it went Party planning was 'meh'. Amazon shopping failed because Operator couldn't add items to the cart. The freelancer task went badly, which she blames on her own prompt. Operator did better when 'good enough' sufficed.

  • Test an agent on SEO research and outreach tasks

    SE Ranking tested Operator on pulling SERP results, backlink checks and outreach email management; it was slow and error-prone, so they did not recommend it for complex SEO work.

    How it went Simple SERP pulls and outreach drafting worked best. Competitor research was abandoned partway, keyword research covered one keyword at a time, and the backlink check wrongly marked some pages as having links.

  • Research sales leads with Operator

    Testers ran Operator on lead research; after a CAPTCHA handover it lost its mission and returned generic company profiles instead of qualified leads.

    How it went The first attempt returned generic company profiles, not qualified leads. Each time control came back, the testers restated the source being checked, what to look for, and which companies were already found. It worked for one prospect at a time.

  • Collect financial data on 100 independent schools

    A writer had Operator gather financial data for independent schools in the background; it was slow and was stopped after about 25 schools, then the data went to o1 for analysis.

    How it went Operator was very slow and was stopped after about 25 schools. It rounded numbers inconsistently and sometimes picked the wrong organization. o1's dashboard was usable, though basic and needing label fixes, and the writer chose not to share it.

  • Find trending articles and generate LinkedIn post variations

    A tutorial shows Operator searching Google for trending eCommerce articles, picking one, and using EasyGen to generate 10 LinkedIn post variations, pausing for the user to log in.

    How it went Operator produced 10 different LinkedIn posts from the chosen article. The source gives no quality assessment, timing, or cost.

  • Play Wordle in a browser with a computer-using agent

    Researchers had OpenAI's Computer-Using Agent play NYT Wordle; it solved only 5.36% of about 200 runs, mainly from misreading tile colors.

    How it went The agent solved only 5.36% of runs. The main failure was that it recognized tile colors inconsistently, and its accuracy depended on the context.

Five new use cases in your inbox every morning

The best things people got an AI agent to do, each with the prompt to try it.

Agents used for similar jobs

  • Perplexity's agentic browser.

    Most often: travel · sales & marketing · research

  • OpenAI's agent mode inside ChatGPT.

    Most often: shopping · travel · research

  • OpenAI's always-on agent that works on its own cloud computer, also written Dot.

    Most often: travel · work · software

  • Amazon's assistant that takes actions like bookings and orders.

    Most often: shopping · life admin

All 22 agents

Browse by kind of job