Last reviewed: October 6, 2026. Summaries on this page come from our own source-checked case files. Each one links to the full write-up, with primary sources and a confidence label for every claim.
Air Canada lost a small-claims case over C$812.02. Knight Capital’s misfiring orders cost about $440 million by the company’s count, and over $460 million by the SEC’s. McDonald’s never published the accuracy rate of its drive-thru AI. The costs differ by orders of magnitude, but the habits behind them look alike, and so do the ways the stories get retold wrong.
Short answer: An AI case study is a documented account of what happened when an algorithm or AI system was deployed: what failed, who bore the cost, and what the record does and does not prove. This hub indexes nine of them, from 2012 to 2026, across finance, housing, retail, government, aviation, hiring and consulting. Each is graded for evidence, labeled by the type of consequence (court ruling, regulator penalty, financial loss, or reputational only), and annotated with the claims that tend to be exaggerated. Not all nine involve machine learning, and the page says which do.
| Case | Year | Sector | What failed | Consequence | Evidence grade | Full case |
|---|---|---|---|---|---|---|
| Knight Capital | 2012 | Trading | Software deployment and testing controls | SEC penalty of $12 million | High | Eight-case roundup |
| Dutch childcare benefits | 2013 to 2021 | Government | Self-learning fraud-risk model that used nationality | Cabinet resigned; data protection fines | High (victim counts vary by source) | Eight-case roundup |
| Amazon recruiting tool | 2014 to 2018 | Hiring | Résumé model trained on a male-dominated history | Project disbanded; Amazon says recruiters never used it to evaluate candidates | High (deployment extent disputed) | Eight-case roundup |
| Robodebt | 2015 to 2019 | Government | Rule-based income averaging with no human check | Court findings, Royal Commission, two settlements | High | Eight-case roundup |
| Zillow Offers | 2021 | Housing | Home-price forecasting at scale | Business wound down; $304.4 million Q3 write-down | High (company filings) | Eight-case roundup |
| Air Canada chatbot | 2022 to 2024 | Aviation | Chatbot gave a wrong refund rule | Tribunal ruling; C$812.02 owed | High | Read the full case |
| RealPage | 2024 to 2026 | Housing | Alleged algorithmic coordination of rents | Federal settlement; state claims continue | Medium (litigation active) | Eight-case roundup |
| Deloitte Australia | 2025 | Consulting | Generative AI report with fabricated citations | Partial fee refund | High | Eight-case roundup |
| McDonald’s and IBM drive-thru | 2019 to 2024 | Fast food | Voice ordering short of its accuracy target | Pilot ended; second attempt began in 2026 | High for the pilot, Low for 2026 claims | Read the full case |
A case enters this hub only if its core facts rest on a filing, a regulator, a court or tribunal decision, or at least two independent outlets. Where sources disagree, the disagreement is printed, not smoothed over.
Every claim in a full case file carries one of three labels:
Cases that are still moving, such as RealPage, are marked as developing and rechecked before each update. The full process is on our Methodology page.
Calling every entry “AI” makes a punchier headline and a wrong one. We sort cases into four tiers by what actually failed.
| Tier | What it means | Cases here |
|---|---|---|
| Classical software | No learning involved; a deployment or control failure | Knight Capital |
| Rule-based automation | Fixed logic applied at scale, with no human review | Robodebt |
| Machine learning and generative AI | Learned or generated outputs that were wrong, biased, or invented | Dutch childcare model, Amazon, Zillow, Air Canada, Deloitte, McDonald’s |
| Algorithmic coordination | The software did its job; the alleged harm is market conduct | RealPage |
Zillow’s technology is publicly described as ML-based, but Zillow has not published the model behind its purchasing decisions, so “pricing algorithm” is the safest label.
A software installation error sent erroneous orders into the market at the open. According to the SEC, the firm’s router sent more than 4 million orders in the first 45 minutes while trying to fill 212 customer orders. Knight reported a realized pre-tax loss of about $440 million; the SEC’s order says more than $460 million. Both figures are legitimate, because one is the company’s accounting and the other is the regulator’s. The SEC charged inadequate controls for code deployment and testing, and Knight paid a $12 million penalty.
What it does not show: an AI failure. It is included because it shows how automation multiplies a small error faster than people can react. Full case in the roundup
The Dutch tax authority introduced a self-learning risk model to flag suspected benefit fraud, and the model treated non-Dutch nationality as a risk factor. Roughly 26,000 families were wrongly accused, though counts vary by source, and the entire cabinet resigned in January 2021. The Dutch data protection authority fined the tax administration €2.75 million in December 2021 and a further €3.7 million in April 2022.
What it does not show: that the bias was an accident. The inputs were chosen by people. Full case in the roundup
Reuters reported in October 2018 that Amazon had been building résumé-screening software since 2014 and found by 2015 that it did not rate technical candidates in a gender-neutral way. The model was trained on ten years of résumés, mostly from men, and penalized résumés containing the word “women’s.” Amazon edited out those terms but could not guarantee it would not find other proxies, and the project was disbanded.
What it does not show: a tool deployed and discriminating at scale. Amazon says recruiters never used it to evaluate candidates. The accurate framing is built, tested and abandoned. Full case in the roundup
Australia’s scheme automatically issued debt notices when tax-office income data did not match welfare recipients’ reported earnings. It dropped the earlier manual check and averaged income across 12 months. In 2019 the Federal Court found that averaging could not by itself establish debts. A class action settled for A$1.8 billion in 2021, and the 2023 Royal Commission called the scheme a costly failure of public administration. A second settlement of about A$475 million in additional compensation was approved on June 23, 2026.
What it does not show: a single combined total. The first settlement was mostly refunds and wiped unlawful debts; the second is new compensation for harm. Do not add them into one total. Full case in the roundup
Zillow’s board decided on November 2, 2021 to wind down its home-flipping business, citing unpredictable home prices, capacity limits and operational problems, with about a quarter of the workforce affected. The company took a $304.4 million inventory write-down in the third quarter of 2021 alone, because it had bought homes at higher prices than it could sell them for.
What it does not show: that “the algorithm broke.” Zillow’s own framing is forecasting uncertainty, and it kept refining the Zestimate. The business died; the valuation tool survived. Full case in the roundup
After his grandmother died, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares and was told a partial refund could be claimed after travel. The airline’s policy said otherwise. On February 14, 2024, the British Columbia Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation (Moffatt v. Air Canada, 2024 BCCRT 149). It rejected the argument that the chatbot was a separate legal entity. The total owed was C$812.02.
What it does not show: binding precedent, a hallucination finding, or an order to “honour the policy.” It is a small-claims tribunal decision under British Columbia law, decided on a thin record. Read the full case
The allegation is not that the software malfunctioned. It is that it worked as designed, using landlords’ confidential data to nudge them toward similar prices. RealPage denies wrongdoing. The Justice Department settled in November 2025 with no damages and no admission, and a federal judge approved that deal on May 19, 2026. State claims continue, and on October 1, 2026 a judge said the federal settlements did not address the core claims the states are pursuing.
What it does not show: a finding of wrongdoing. This is the entry most likely to change, so it is rechecked on the day each update is published. Full case in the roundup
A 237-page report for Australia’s Department of Employment and Workplace Relations contained a fabricated quote from a federal court judgment and references to research papers that do not exist. The revised version disclosed that a generative AI tool (Azure OpenAI GPT-4o) was used. Deloitte agreed to repay the final installment of its AU$440,000 contract.
What it does not show: that Deloitte blamed AI. It did not publicly attribute the errors to the tool. Full case in the roundup
McDonald’s tested IBM-linked automated order taking at more than 100 US restaurants and told franchisees it would be switched off by July 26, 2024. McDonald’s never published accuracy data or an official reason. In June 2021 its CEO said the system was about 85% accurate and handled about 80% of orders. A 2022 analyst report put accuracy in the low 80% range against a 95%-plus target. In June 2026 McDonald’s began testing a new Google-powered system at five US restaurants.
What it does not show: that viral videos killed the pilot, or that the new system is better by a measured margin. The “above 90%” figure for the new system is not defined as accuracy or as orders completed without human help. Read the full case
A court or tribunal ruled: Air Canada, Robodebt (and RealPage, where litigation continues)
A regulator penalized or fined: Knight Capital, Dutch childcare benefits
A business or project ended: Zillow Offers, Amazon recruiting tool, McDonald’s and IBM pilot
A client was refunded: Deloitte Australia
Controls and human oversight: Knight Capital, Robodebt
Inputs and training data: Dutch childcare benefits, Amazon recruiting tool
Forecasting a market that moved: Zillow Offers
Generated falsehoods: Air Canada chatbot, Deloitte Australia
Performance against a stated target: McDonald’s and IBM drive-thru
Market conduct: RealPage
Most retellings of these cases go wrong on one figure. These are the ones to check before you quote anything.
| Case | The loose version | What the sources support |
|---|---|---|
| Knight Capital | “Lost $440 million” | About $440 million by the company’s count; over $460 million by the SEC’s |
| Robodebt | “A$1.8 billion” as the full cost | Two separate settlements: A$1.8 billion (mostly refunds and wiped debts) and about A$475 million in new compensation |
| Air Canada | “Ordered to honour its policy” | Damages for negligent misrepresentation; the decision did not order a policy change |
| Amazon | “Used to reject women” | Built, tested and abandoned; Amazon says recruiters never used it to evaluate candidates |
| Deloitte | “Blamed AI for the errors” | It disclosed AI use in the revised report but did not publicly attribute the errors to it |
| McDonald’s | “The new system is over 90% accurate” | A company-stated figure whose metric is undefined; no independent measurement published |
These patterns are analysis, not reported fact, so read them as arguments.
The failure is usually organizational. Knight lacked deployment controls. Robodebt removed the human check. Deloitte did not catch invented references before delivery.
Humans are removed first and restored last. The cheapest safeguard, a person who can say “this looks wrong,” is the one most often engineered out.
Liability stays with whoever deploys the tool. Air Canada tried to hand responsibility to its chatbot and lost. Deloitte’s partial refund is the commercial version of the same principle.
Scale turns small errors into balance-sheet events. Speed and volume are the point of automation, and also why mistakes compound.
Technology often did what it was built to do. The decisions that failed were human ones: what data to use, what to automate, what to skip testing, and who was allowed to object.
This hub covers documented failures because that is where the record is most detailed and most often misreported. Many “AI case study” pages elsewhere are written by vendors, and we hold success claims to the same standard as failures: a filing, a regulator, a court, or two independent outlets. For guidance on pilots, returns and rollout, see our Enterprise AI Implementation guide. For the legal landscape around these cases, see AI Regulation.
What is an AI case study? A documented account of an AI or algorithmic system in real use: what it was meant to do, what went wrong or right, what it cost, and which claims about it are supported by evidence.
Are all nine cases really AI? No. Knight Capital was a software deployment failure and Robodebt was rule-based automation. RealPage involves algorithmic coordination, where the harm alleged is market conduct. The page labels each case by tier.
Can a company be held responsible for its chatbot’s mistakes? In Moffatt v. Air Canada, a Canadian tribunal said yes and rejected the argument that a chatbot is a separate legal entity. It was a tribunal ruling, not a binding court precedent.
Was Zillow’s failure caused by AI? Zillow’s pricing technology is described as ML-based, but its own explanation emphasizes forecasting uncertainty and operational limits. “The algorithm was solely to blame” is not supported by its filings.
Is McDonald’s still using drive-thru AI? It ended the IBM test in 2024 and, as of mid-2026, is testing a new Google-powered system at five US restaurants.
How do you verify these cases, and how often are they updated? Core facts must rest on a filing, a regulator, a tribunal or court decision, or at least two independent outlets, and every claim carries a confidence label. Developing cases are rechecked before each update. See our Methodology.
We publish corrections openly and show what changed. Send the claim, the page it appears on, and your source through our contact page. Case suggestions need a primary source: a filing, decision, regulator notice, or two independent reports.
Not legal or financial advice. These summaries describe public records and reporting as of the review date above.
The short answer: no. The “95% of AI pilots fail” line did not come from a measured failure rate of pilots. It came from a funnel chart in a small, non-peer-reviewed report that headlines and the report’s own summary misread. Read correctly, the same chart says that about one in four companies that ran a … Read more
In June 2023 a federal judge fined two New York lawyers $5,000 for a brief full of cases ChatGPT had made up. At the prices verified for this article, $5,000 buys about 67,000 runs of a ten-page document summary on the priciest OpenAI or Anthropic tier, and about 6.7 million runs on the cheapest. Most … Read more
By Tom Morgan · Last updated 22 October 2026 · General information, not legal or financial advice The short answer. The cheapest failures in this file are the most instructive. A Canadian tribunal order of C$812.02 against Air Canada and a partial refund by Deloitte Australia needed no technical sophistication to repeat: nobody checked the … Read more