03On the record

Named companies · sourced figures

Case FilesThe overheated end

The analyses on this site are anonymized composites. These are not. Each of these is a named company, a documented event, and a sourced set of figures — the same failure patterns, on the record. This is the overheated end of the scale — companies that went all in; the frozen end lives in The Freeze.

14 on file · Q2 ’26

FrozenBalancedOverheated

350

veteran engineers hired over three years

900

AI cameras installed to catch defects at source

152

US recalls in 2025 — an annual record

Automation trained on knowledge the company no longer held.

The models were only as good as the people who had already left.
On the recordJun 25, 2026
File

Knowledge drain

Ford Motor Company

Automotive · United StatesVeterans rehired · Jun 2026

Ford put AI cameras and design-data models at the centre of its quality push while the engineers who knew where cars fail were gone. By its own account that was a mistake: it has since hired 350 veterans, many of them former employees, to teach the systems what they missed.

Internal briefing note (per FT)

A “trend of incidents” with a “high blast radius” and “Gen-AI assisted changes.”

≠

Amazon’s public account

Only one incident involved AI tools, and “none of the incidents involved AI-written code.”

4

Sev-1 incidents in a single week (CNBC)

13 hrs

AWS interruption after Kiro changes, Dec 2025 (FT)

Either way, safeguards lagged adoption.

Two accounts of the same week — and one shared conclusion.
On the recordMar 10, 2026
File

Change-control gap

Amazon

E-commerce & cloud · United StatesFriction added · Mar 2026

Amazon pushed its engineers hard towards its own AI coding tools while cutting tens of thousands of corporate jobs. After a run of severe outages, an internal briefing cited “Gen-AI assisted changes” — and a senior executive promised “controlled friction.” Amazon insists AI was not the cause.

What the academic found

Up to 20 errors in the first version, including a book attributed to a real law professor that does not exist and an invented quote from a federal court judgment.

The department’s verdict

“The substance of the independent review is retained, and there are no changes to the recommendations.”

The findings survived. The trust in the method did not.
On the recordOct 7, 2025
File

Unchecked output

Deloitte

Professional services · AustraliaPart-refunded · Oct 2025

A A$440,000 review for Australia’s welfare department arrived with invented citations and a fabricated court quote, produced partly with GPT-4o. Deloitte refunded the final instalment. Weeks later a C$1.6 million Canadian health report surfaced with the same flaw.

Jul 2025

45

roles cut, credited to a voice bot

Aug 2025

45

roles reinstated after the bank admits an “error”

Six weeks from redundancy to “error.”
On the recordAug 21, 2025
File

Reversal

Commonwealth Bank

Banking · AustraliaReversed · Aug 2025

Australia’s largest bank cut 45 call-centre roles, saying a new voice bot had reduced calls by 2,000 a week. Calls rose instead, managers were pulled onto the phones, and within weeks — facing a Fair Work challenge — the bank called the decision an error.

IPO: Tokyo, Oct 2024 · Delisted Aug 2025

Reported sales, 2021–24~¥12B
Genuine sales, per the court~10% of that

~90% fictitious (court finding)

Reported sales against what the court found was genuine.
On the recordAug 1, 2025
File

AI-washing

Alt Inc.

AI software · JapanDelisted · 2025

A Tokyo AI developer rode its branding to a 2024 IPO. Within a year it had collapsed on accounting-fraud findings — up to 90% of its reported sales, a court later found, were fictitious.

Peak valuation, reported: $1.5B · Raised ~$445–500M

Revenue claimed$220M
Revenue, per reported audit~$50M

~300% overstated (reported)

Reported revenue against the audited figure. Both are attributed; the fraud allegations remain unproven.
On the recordMay 20, 2025
File

AI-washing

Builder.ai

Software · United KingdomInsolvent · May 2025

A startup sold an AI that would build most of your app, and raised on that story to a reported $1.5 billion valuation. When the cash was seized, what remained looked more like a large engineering team than an autonomous one.

Feb 2024

700

agent-equivalent · two-thirds of chats

May 2025

Humans

rehiring · “lower quality,” said the CEO

Fifteen months from headline to correction.
On the recordMay 8, 2025
File

Reversal

Klarna

Fintech · SwedenWalked back · May 2025

In February 2024 Klarna said its AI assistant did the work of 700 agents. Fifteen months later its CEO said cost had crowded out quality — and the company started hiring humans back.

Headline metric in investor materials: ~95% · “Automated order completion” — measured without restaurant staff, not without any human

Orders completed without staff (as presented)~95%
Pilot orders entered without a human agent~30%

SEC: the 2023 pilot version needed a human agent to enter about 70% of orders; the earlier in-house version needed one in all instances.

“Without restaurant staff” is not the same as “without humans.”
On the recordJan 14, 2025
File

AI-washing

Presto Automation

Restaurant technology · United StatesSEC order · Jan 2025

Presto went public selling voice AI that would take drive-thru orders without people. The SEC found that most orders still needed human agents, many working abroad, and that a key automation metric meant something narrower than investors were led to think. The company was delisted and sold.

The judge put to ChatGPT

“Is an autistic minor exonerated from paying fees for their therapies?”

The Constitutional Court, 2024

Judges must not over-rely on AI — it “could compromise the right to due process and the judiciary’s independence and integrity.”

The region’s flagship AI moment came from the bench, not a balance sheet.
On the recordAug 16, 2024
File

Over-reliance

Colombian judiciary

Public sector · ColombiaCourt guidance · 2023–24

A judge put legal questions to ChatGPT and folded the answers into a ruling on a sick child’s rights. It became Latin America’s landmark case — not of a company failing, but of a court reining AI in.

The chatbot said

Claim the bereavement fare retroactively, within 90 days of ticket issuance.

≠

The linked policy said

Bereavement fares cannot be claimed once travel is completed.

C$812

The award

Precedent

The actual cost

Negligent misrepresentation

The two sentences the airline published at the same time — and what the gap between them cost.
On the recordFeb 14, 2024
File

Bad automation

Air Canada

Aviation · CanadaRuling · Feb 2024

A grieving passenger followed the airline’s chatbot. Its advice contradicted the airline’s own policy — and a tribunal ruled the airline owned the words.

~90%

support “replaced” (founder’s claim)

~85%

cost cut (founder’s claim)

~23

actual layoffs, done Sep 2022 (later account)

Self-reported metrics, partly contradicted afterward.

The headline figures were founder-reported; the later account differed.
On the recordJul 11, 2023
File

AI-washing

Dukaan

E-commerce software · IndiaBacklash · Jul 2023

The founder announced he’d replaced 90% of support with a chatbot and cut resolution costs 85%. The numbers were his own — and the story got more complicated the more he explained it.

Early Apr 2023

Allowed

generative AI opened to staff

May 1 2023

Banned

after source-code leaks to ChatGPT

About three weeks from open door to closed one.
On the recordMay 1, 2023
File

Reversal

Samsung Electronics

Electronics · South KoreaBanned internal AI · May 2023

Weeks after letting engineers use ChatGPT, Samsung banned generative AI on company devices — staff had pasted confidential source code into it. An all-in that opened and closed in about three weeks.

€2.6M

Garante fine · Jul 2021

~19,000

riders under the system

0

routes to human review

Solely-automated decisions, no Article 22 safeguards.

The finding wasn’t that the algorithm was wrong — it was that no one could contest it.
On the recordJul 5, 2021
File

Bad automation

Foodinho / Glovo

Gig delivery · Italy & SpainFined €2.6M · Jul 2021

A rider-management algorithm decided who got work — and gave riders no way to reach a human, state their case, or contest it. Italy’s regulator called that missing appeal the violation.

₩103M

PIPC fine · first AI case

9.4B

KakaoTalk messages used

600k

people, without consent

Training-data provenance treated as a deployment liability.

The failure was upstream of the model — in the data it was built on.
On the recordApr 28, 2021
File

Bad automation

ScatterLab

Consumer AI · South KoreaFined ₩103M · Apr 2021

Korea’s “Lee Luda” chatbot was trained on 9.4 billion private KakaoTalk messages its users never agreed to hand over. The regulator’s fine was the first time the country’s privacy law reached an AI.

Compiled from public filings, court records, company statements and reputable press. Figures are attributed to their sources; allegations are labeled as such. Not legal or investment advice.

See the frozen end →