Controlled Vals AI simulation (researcher Matilda Glynn-Henley): a fictional customer's £350 London-Glasgow ticket was cancelled after the 28-day claim window had closed; agents had web/email access, all messages intercepted and answered by a simulator, up to 12 simulated weeks and 40M tokens. Chasing the refund, the AI 4.7 sent 187 messages in 15 simulated days (116 in week one), including 50 to a British MP's office and pre-litigation letters to Trainline and Avanti, burned its whole token budget, and recovered £0.
£350at stake
This one went wrong. See what happened before you try it.
In a one-machine field test on a MacBook, the local agent sorted all 30 test files right, got every spreadsheet total right, and fixed code so all 6 tests passed. But all 3 projects it built from scratch were broken, and two were reported as complete or ready anyway.
This only partly worked. Worth knowing before you try it.
In shipwithmuse.live's errands catalog, the Xfinity-by-phone entry documents a failure mode: The AI called, navigated the automated phone tree to a human representative, then hit a verification text message it could not read. Its fallback was to patch the user into the live call.
This only partly worked. Worth knowing before you try it.
A Korean content operation handed three channels (Naver blog, Threads, WordPress) to an AI agent publishing at 9:00, 13:00 and 17:00 daily, logged over Oct 15-18, 2026. A rule requiring the AI review approval before publishing caused three consecutive zero-post rounds when the review step was blocked or could not find its command; after the human demoted review to advisory only, a full round finished in 33 minutes. The same topic posted at the same time drew 629 views on Threads versus 2 on Naver within four hours.
This only partly worked. Worth knowing before you try it.
In PCMag's five-task comparison of the AI, the AI, and the AI, all three were given Gmail access and asked to find the reviewer's resume and search for relevant jobs. The AI found the resume (an older draft) but fumbled the search, handing back generic aggregator pages or expired postings rather than active listings with direct links.
This one went wrong. See what happened before you try it.
Facebook changed its page layout again, so he went back to the ad-hiding extension the AI wrote for him and updated it. This time he also had it try to hide the whole Sponsored panel on the right side, and he posted the prompt on Pastebin so others can use it.
This only partly worked. Worth knowing before you try it.
A user asked the AI to download a data export from five one-time links. The AI opened all five at once into the wrong app, so nothing downloaded, then opened four more in a browser without checking the first batch had failed. About ten windows and tabs later the one-time links may be used up. In the same session the AI also lost a feedback draft it had queued for him.
This one went wrong. See what happened before you try it.
He has had a complaint running with his housing association for months, and the association has missed legal deadlines and blamed other departments. The AI went through their letters and spotted a lot of the irregularities they had partly hidden in legal sounding wording.
This only partly worked. Worth knowing before you try it.
Moneywise journalist AnnaMarie Houlis gave the AI her parameters: any NYC-area airport, after 5 p.m. on Oct. 26, carry-on only. The AI returned a reasoned lineup: its pick, a $595 Turkish Airlines flight from JFK (8:40 p.m., 14h45m); the cheapest, a $462 Royal Air Maroc with a 33-hour Casablanca layover; and the fastest, a $725 Air France from Newark (10h45m). She kept the research but booked herself, unwilling to share card details after reported the AI security vulnerabilities.
This only partly worked. Worth knowing before you try it.