Real Agent Cases

See what real people got AI to do for them

Copy a prompt and use it for your own task.

What is an AI agent case? A real example of someone using an AI agent to get something done, including what happened, the prompt, and a link to the original source. What counts →

Not sure how to start? Try:

656 cases so far, including 106 that failed or only partly worked.

Prompt labels: Original (their words) · Recreated (our rebuild) · Improved (our fix)

All cases

656 of 656 cases, newest first

  • Carlos Mendoza's 4th-grade son marked only B and F as correct in a Santillana exercise on numbers up to the billions. The platform's solution marked B, D and F, but D said '9,615 rounded to the billions is 10,000', which is false. When Mendoza asked ChatGPT to review it, it first said Santillana was right, reading 'billions' as if it were 'thousands'. Only when Mendoza pointed to the exact word did ChatGPT acknowledge the error. The son had been right from the beginning.

    ChatGPT · Personal · Source date:

  • Karan Sharma gave Codex access to all their saved Instagram food reels and asked it to import them into their self-hosted Mealie recipe instance. Codex wrote a pipeline on the fly and ran each reel through it: yt-dlp pulls the audio, Whisper transcribes it, Gemini 3.8 Flash formats it as a recipe and it's sent to Mealie. Sharma calls it very useful.

    Codex · Coding · Source date:

  • For PCMag, Ruben Circelli asked three agents to find J. Kenji López-Alt's no-stir risotto recipe, research the best version of each ingredient at the local Wegmans and add them to an Instacart cart. ChatGPT's Dots, once Instacart was connected, found the recipe, researched the ingredients and filled the cart without issue. Meta's Muse got there slowly, needed Circelli to log in and enter a verification code, and added saffron threads that aren't in the recipe. Gemini's Spark found the recipe and proposed ingredients but couldn't access the account, tried to take over Chrome and froze.

    ChatGPT Dots, Muse, Gemini Spark · Shopping · Source date:

  • Amanda Smith pays $25 a month for a Visible plan and gets a lot of spam calls and texts asking for someone called Andrew, probably the number's previous owner. Visible has no phone line, so Smith asked Muse to check they were formally listed as the number's owner and to find where the number was linked to Andrew. Within seconds it found the number on a data broker page. But Visible blocks bots, and its guest chat wouldn't send Muse's message, then blocked it; taking over the browser led to a 403 error. Muse drafted a message for Smith to paste into the Visible app, and advised requesting removal from WhitePages and FastBackgroundCheck, though the WhitePages form wouldn't let Smith click Next. Visible's advanced support was still pending. Smith found Muse acted more like a chatbot giving instructions, though testing it made Smith finally deal with the task.

    Muse · Personal · Source date:

  • Find a cheaper Medicare drug plan with ChatGPT after a premium rise, then confirm it in Medicare's Plan FinderRecreated promptEvidence: Described by posterBeginner-friendlyReady to run

    A Topretirements blogger's Medicare Part D drug plan wrote to say the 2027 premium would rise from $98.30 to $126 a month. They asked ChatGPT for cheaper options, giving their Florida ZIP code, their medications and preferred pharmacies. It suggested several, including a plan at $5.30 a month in their area, $1,448.40 a year less in premiums. They didn't switch on ChatGPT's word alone: as their insurance agent also advised, they checked it in Medicare's official Plan Finder, which confirmed the switch made sense, and enrolled. They get most prescriptions through the VA, so the cheapest premium suited them; they note others need to compare total yearly cost.

    ChatGPT · Personal · Source date:

  • Dave W Plummer missed an important appearance and asked the ChatGPT desktop app, which has access to their Gmail, to work out how it happened. It found Plummer had noon in mind, the time of an earlier prep call, while the event started at 7:50 a.m. Pacific, with emails also mentioning 8:50 a.m. Mountain. When Codex then reconstructed the email trail, it found the organiser had cancelled the calendar invite sent to Plummer's regular address while sorting out a Teams login mismatch, and the later confirmation email had no new invite. With Codex, Plummer then set up a recurring agent that reads email threads, including changes and cancellations, adds missing calendar entries with the source email and link, and flags contradictions. Test alerts at first didn't notify on the iPhone; after checking alarms, sync and Focus settings, a test finally rang. It now checks every ten minutes while the Mac is awake; its first historical review was still running.

    ChatGPT, Codex · Personal · Source date:

  • Have Claude Cowork prepare late-invoice reminder drafts every Monday night, with a report of who owes whatRecreated promptEvidence: Described by posterBeginner-friendlyReady to run

    Michel, in San Diego, works in a tight-margin business with basic back-office software, and sending late-invoice reminders was repetitive manual work. Claude Cowork walked them through setting up an agent that logs in every Monday night, checks overdue accounts receivable and matches customer names against a Google Sheet of contacts. Using Michel's email template, it adds the contact name, invoice number, amount and due date, puts the emails in the Drafts folder and sends Michel a report of which customers were included, what each owes and the totals. Michel reviews the drafts and sends them, saving about an hour every Monday night.

    Claude Cowork · Ops · Source date:

  • jedberg regularly negotiates OEM agreements with vendors that mostly use the same template and draw the same objections. They tried Claude first, and their lawyer was impressed that it found things the lawyer had missed. jedberg then used Claude to redline that type of contract, and the other side remarked that the lawyer had been surprisingly thorough. The lawyer did the final check and by then basically said it looked good, and jedberg saved a lot of money by doing the first three rounds with Claude.

    Claude · Ops · Source date:

  • Gizmodo's Webb Wright had some badly delayed flights that forced them to pay for Lyft rides home in the early hours, and asked Muse to request refunds. Given screenshots of the receipts and the flight details, Muse drafted decent emails, courteous but firm. But buried inside one draft was the important point that the airlines didn't take refund requests by email, so the drafts would have to be pasted into the airlines' websites. Webb ended up opening a browser and entering the details by hand, though they did use the message Muse had drafted.

    Muse · Travel · Source date:

  • badlibrarian runs a business and sometimes buys things on Amazon. At tax time they used to search for the big business items and categorise them in QuickBooks Online until they gave up, using a janky browser plugin and two monitors full of Excel windows. Now Claude does it and doesn't give up, despite QuickBooks being slow. It will even click 19 times to split a transaction because they'd put a router and shampoo in the same Amazon cart one night in February 2025.

    Claude · Ops · Source date: