KPMG sold a report on agentic AI adoption, then pulled it after key examples appeared to be wrong. For a firm advising companies on responsible AI, that is not a small embarrassment.
KPMG has removed a report on agentic AI after researchers and named organisations challenged claims that looked like the very thing the report was meant to help businesses avoid: AI hallucinations dressed up as authority.
The report, Redefining excellence in the age of agentic AI, was released in October and described how major organisations were supposedly using AI agents in finance, healthcare and public transport. According to the Financial Times, the document made false or misleading claims about UBS, the UK National Health Service, Swiss Federal Railways and Transport for London. GPTZero first identified many of the problems, and the FT said it verified them with the organisations involved.
This is the awkward part. KPMG is not a student handing in a rushed essay. It is one of the Big Four accounting and consulting firms, and it is selling advice to companies trying to decide how much trust they can place in AI systems. When a report about AI excellence cannot keep its own examples straight, the problem is not only reputational. It gives clients a very practical reason to ask what quality control actually means inside the firms now pitching themselves as AI guides.
The UBS claim was especially sharp because it sounded plausible. KPMG said the Swiss bank had integrated AI agents across investment advisory, risk management and compliance monitoring, inside a composable platform co-developed with Microsoft. UBS told the FT the assertions were factually incorrect.
Swiss Federal Railways faced a similar problem. The KPMG report said the rail operator offered AI agents that helped users plan, book and optimise journeys using preferences, real-time conditions and carbon impact. The railway told the FT that was not accurate. Transport for London was described as using AI agents to predict and manage congestion, personalise commuter updates and coordinate multimodal transport. TfL said the claim was misleading.
The NHS example shows how these errors can grow from something real into something much grander. KPMG cited NHS Greater Manchester as using AI agents to predict hospital readmissions, triage patients and automate referrals. The FT reported that the claim appeared to have been based on a release about an AI tool for lung cancer, with no mention of agentic AI doing those tasks. That is how a narrow health technology announcement becomes a sweeping case study if nobody checks the last mile.
TechRadar, citing GPTZero, reported that the KPMG document contained 45 citations and that only five accurately pointed to real sources. The rest were described as fabricated, distorted or misleading. GPTZero has called this pattern vibe citing, meaning references that look serious enough to pass a quick glance but collapse when someone follows the trail.
Consulting cannot outsource trust
KPMG International told the FT it takes the accuracy and integrity of published content seriously and had removed the report from websites while it investigated how it was published. The firm also said it expects staff to follow responsible AI guidelines, including human oversight to validate content and verify independent sources.
That statement is sensible, but it also describes the minimum. Human oversight is only useful if the human actually checks the source, reads the document and asks whether the claim says what the report says it says. A link that exists is not the same thing as evidence. A brand name in a footnote is not verification.
The episode lands harder because KPMG is not alone. The FT reported last month that EY retracted a study after GPTZero found fake footnotes and other errors. Deloitte had to revise a report for a Canadian provincial government after fake academic citations were found. Sullivan & Cromwell apologised to a New York court in April after a filing contained AI-related legal inaccuracies. These are not fringe publishers. They are institutions that businesses and governments pay precisely because they are supposed to reduce risk.
For startups, there is a useful warning here. Big firms often set the language that later becomes procurement policy, boardroom consensus and investor shorthand. If a KPMG report says agentic AI is already being used in a certain way by UBS or TfL, that claim can travel through pitch decks and strategy memos before anyone notices the foundation is weak. GPTZero chief executive Edward Tian told the FT that error-riddled publications by Big Four firms can poison the well of information, especially when they are cited by other publications and then absorbed into AI systems.
This is where the story becomes bigger than one pulled report. The AI market is already crowded with confident claims about agents, autonomy and enterprise adoption. Some of those claims are real. Many are pilots, narrow tools, or ordinary automation with a new label on top. The difference matters because companies are spending real budgets on systems they may not fully understand, and vendors are under pressure to show that everyone else is already moving.
KPMG pulling the report was the right step. It was also late. The real test now is whether professional services firms treat AI-assisted publishing as a controlled process rather than a faster way to produce thought leadership. Clients should ask to see the audit trail, not just the slide deck. If the people selling responsible AI cannot prove where their own claims came from, their advice deserves a much colder reading.
Also read: Anduril wants export controls to catch up with drone warfare • AI data centers are becoming a 2026 campaign liability • Bill Gates warns Washington is changing the rules for tech
