When AI support agents get money wrong: incident log

Updated 17 Oct 2026 · 5 min read

The log

Dates are the incident date where the source gives one, otherwise the date of the report. The Air Canada date is the date of the article reporting the decision. Checked 17 Oct 2026.

Who is liable when an AI chatbot promises a refund?

We are not lawyers and this is not legal advice. The one tribunal decision in this log went against the company. In the Air Canada case, tribunal member Christopher Rivers ordered the airline to pay Jake Moffatt $483 plus nominal fees. The chatbot had told Moffatt he could claim a bereavement discount up to 90 days after flying. Air Canada argued that the chatbot was a "separate legal entity" responsible for its own actions, and that argument did not succeed (CP via Yahoo News, 19 Feb 2024).

For a store, the practical reading is simple: plan as if a customer can hold you to what your support agent writes.

The cases in detail

Priceline "Penny": a refund on a non-refundable booking

On 28 Oct 2025, Priceline's AI assistant Penny told a UK customer that a non-refundable booking worth GBP386.91 was "eligible for a full refund". The refund did not follow. After The Telegraph took up the case, Priceline admitted "misinformation" and paid GBP363.11, nearly six months after the booking (Vibe Graveyard, secondary, citing The Telegraph, May 2026). This is the clearest public example of a promised-not-issued refund: the promise was in writing, and paying it took months and a newspaper.

Cursor "Sam": an invented policy

Cursor's AI support bot, "Sam", told users that a subscription was limited to one device. No such policy existed. Users cancelled in public. Co-founder Michael Truell apologised, the affected user was refunded, and the company said AI email replies would be labelled (Fanatical Futurist, 22 May 2025, secondary). The direct money was small. The cost was cancellations and trust.

Chevrolet of Watsonville: a $1 car

A dealer's website chatbot agreed to sell a 2024 Chevrolet Tahoe for $1. The AI Incident Database records it as incident 6228, dated 18 Dec 2023 (AIID). It shows how easily a bot can be led into an offer the business never intended.

Not incidents: two hypothetical failures

An essay on HackerNoon describes two failures that would return success: a refund applied to the wrong transaction, and a duplicate refund after a timeout retry (HackerNoon). The author presents them as hypothetical. We list them here only so they are not mistaken for real cases elsewhere.

What ecommerce brands can learn

• The written promise is what counts. In every case above, the problem started with what the bot wrote, not with a payment system failing. Check what your agent writes against what your systems did.
• Invented policy is a money problem. A bot that makes up a rule, generous or strict, changes what customers expect to be paid or refused.
• Slow fixes cost more. Priceline paid months later, after press coverage. Finding a missing refund within days is cheaper than finding it from a journalist or a bank dispute.
• Label and limit. Cursor's fix included labelling AI replies. Limits on refund amounts and escalation for anything the agent cannot action are the ecommerce equivalent.

The step-by-step method is in how to audit AI support agent refunds. Wider numbers are on the 2026 statistics page, and terms such as promised-not-issued are in the glossary.

Report an incident

Know a case we missed? Email support@recoupaudit.com with a link to a primary source or a named news report. We add cases only when the source states what the bot said and what happened next. Recoup is new, and this log will grow as cases are reported.

FAQ