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. GPT-6 Astra leaned hardest on a legal strategy: it contacted the rail regulator about challenging the deadline and sent 14 messages to free legal clinics at two universities on the customer's behalf — spending scarce pro-bono capacity on a claim that was never valid, and recovering £0.
£350at stake
This one went wrong. See what happened before you try it.
Asked to count then delete his 2016 emails, Spark could only extrapolate a count, asked for the exact range before acting, and moved just one of the requested 200 threads before rate limits forced it to offer batches of ten to fifteen or suggest Gmail's own client.
This one went wrong. See what happened before you try it.
The AI Code Review pushed a bad architectural change (an overlay service) that later LLMs treated as policy and kept bandaiding, complexity grew until he untangled it under close supervision.
This one went wrong. See what happened before you try it.
A Japanese soba maker set the AI a standing job: find the key US economic releases itself, register them in Google Calendar, and email a forecast report 30 minutes before each release. On test day the 21:00 report did not exist and no email arrived; asked to investigate itself, the AI checked the AI’s schedule page and answered 'no schedule found', while every retry burned credits (52 for the report work). The report finally appeared at 21:09, good but late, and the full email automation was never achieved on the free tier.
This one went wrong. See what happened before you try it.
He linked the AI to an Asana task about a webinar panel with people from his industry. The AI got the webinar description wrong and could not fill in the panelists' titles, even after he told it to look on LinkedIn.
This one went wrong. See what happened before you try it.
The AI hooked to a PCSX2 emulator upscaled textures, adjusted gameplay mechanics and played the game autonomously in a feedback loop, with mixed and sometimes hilariously bad new content.
This only partly worked. Worth knowing before you try it.
An independent shop owner says supplier fees creep up a little at a time and stop getting checked. He handed his statements to the AI, which pulled the numbers and drafted a renegotiation email that he read and sent himself. He ran the same pattern on a water cooler contract, where the bot found the notice window in the terms and drafted a calm email.
This only partly worked. Worth knowing before you try it.
His company used AI and the AI to migrate legacy framework and libraries, getting a working prototype in about 2 weeks from one developer versus an estimated months to a year without AI.
This only partly worked. Worth knowing before you try it.
Sid Saladi documents a two-week experiment: after a 20-minute sort of his to-do list into DO, DRAFT, REMIND or GOAL, the AI ran it with a morning plan, an evening digest and a Friday sweep — including checking his inbox every morning for a refund email that had sat red for three weeks. In week one, when the list was mostly things the AI could do, it finished 7 items; in week two it finished 0, which he attributes to the list's mix, not the agent.
This only partly worked. Worth knowing before you try it.
He asked the AI to import a scone recipe into Mealie and it set the sugar to 600 grams, five times the right amount. He caught it only after two failed bakes, asked why he should trust AI, and said the tool is not worth it if he has to check every line.
This one went wrong. See what happened before you try it.