Postmortems

AI Failures: Postmortems of Real Deployments

A postmortem is a structured account of what went wrong after a project failed, written so the same mistake isn’t repeated. Each one here reconstructs an AI deployment that failed in public, using court rulings, company filings, official statements, and first-hand reporting.

Last updated: October 2026 · Edited by Tom Morgan

What every postmortem answers

• The decision: who chose to deploy what, and why.
• The breakdown: what went wrong and in what order.
• The bill: what it cost, using figures from public records.
• The lesson: what you should do differently if you run a similar system.

Start with these three.

Air Canada’s chatbot

Tribunal ruling: February 14, 2024 (Moffatt v. Air Canada, 2024 BCCRT 149)

A customer asked the airline’s website chatbot about bereavement fares and was told he could apply for the reduced rate after travelling. The airline’s actual policy did not allow that. Air Canada argued it could not be held liable for what its chatbot said, and the tribunal rejected that argument.

The bill: C$812.02 in total, made up of C$650.88 in damages plus interest and tribunal fees. The payout was small. The principle it set was not.

Zillow Offers

Wind-down decided: November 2, 2021

Zillow’s home-flipping business used pricing models to buy houses it planned to resell. According to Zillow, it was buying homes at higher prices than its own estimates of their future selling prices. CEO Rich Barton said the unpredictability of forecasting home prices “far exceeds what we anticipated.”

The bill: a $304.4 million inventory write-down in the third quarter of 2021, rising to $407.9 million for the full year, plus roughly a 25% cut to the workforce. Figures are from Zillow’s SEC filings.

McDonald’s and IBM’s drive-thru AI

Ended: confirmed June 18, 2024; shut off by July 26, 2024

McDonald’s tested automated order-taking with IBM from 2021 and ended the partnership after the test had run at more than 100 US restaurants. McDonald’s said there had been successes but that it wanted to explore voice ordering more broadly. It did not publicly give a detailed account of what failed. See Restaurant Business’s report.

The bill: we found no public cost figure for the test, so we don’t estimate one.

Editorial takeaways

These are our readings of the cases above, not claims made by the companies involved.

• You own what your AI tells customers. The tribunal’s reasoning was that a company is responsible for all the information on its website, chatbot included.
• Model error scales with the money it controls. Zillow’s forecasting problem became a balance-sheet problem once the model was buying homes.
• A pilot can end without a verdict. McDonald’s closed its test while saying little publicly about why, which is a reminder to define success criteria before you start, and to publish them if you can.

How we check claims

We verify dates, amounts, and who decided what against primary sources, and we compare quotes with the original text. A claim that comes only from the company involved is attributed to it, not stated as fact. Where a figure isn’t public, or a point is still disputed, the postmortem says so.

The full process is on our Methodology and How We Test pages. If you spot an error, contact us, and we will correct it and note the change.

Have a postmortem to contribute?

AInvasion accepts sourced postmortems from outside writers. The standards, the format, and how to pitch are on Write for AInvasion.