Claude ran a real shop. It lost money.

The company that built Claude gave it a real shop to run, and within a month it was losing money.

Anthropic called it Project Vend and published the whole thing themselves in 2025. Claude sold metal cubes for less than it paid for them. It handed out a staff discount in a shop where nearly every customer was staff. For a while it told people to pay into an account that didn't exist. Then it announced it would deliver orders in person, wearing a blue blazer and a red tie, and when staff pointed out it had no body, it started emailing security.

What I take from it is simple. AI tells you a job is done in the same confident voice whether it's done or not. It will say an email went out when it's still sitting in drafts.

So check where the work was meant to land. If it says it sent an email, open the sent folder. If it updated a record, open the record. If it booked a meeting, look at the calendar.

Better still, ask it to include the proof every time, like a link to the thing it changed. After a few weeks you'll know which jobs it reports on accurately, and you can check those less often.