One of the world’s biggest consulting firms quietly removed a flagship report about AI. The reason is awkward. The report itself was built with AI, and the AI invented client stories that never happened. The companies named in it are not happy.
The Gist
- The KPMG AI report cited UBS, the NHS, and two transport agencies that all denied the claims
- The fabrications were spotted by an AI detection group, then verified by the Financial Times
- EY pulled a similar AI-assisted report only weeks earlier, so this is becoming a pattern
Have ChatGPT Recap This Article
ChatGPTWhat KPMG did and why it got caught
KPMG is one of the four largest accounting and consulting firms in the world. They are known as the Big Four, alongside EY, Deloitte and PwC. Companies pay them millions to advise on big decisions, including how to roll out AI inside their own walls.
In October 2025, KPMG published a flagship report titled “Redefining excellence in the age of agentic AI“. The word “agentic” just means AI that can do tasks on its own, not only answer questions. The KPMG AI report was a sales tool to convince clients this is ready for prime time.
To make the case, the report pointed to real organizations supposedly using agentic AI in production. UBS, the UK National Health Service, Swiss Federal Railways, and Transport for London were all named. The problem is simple. None of them did what KPMG claimed.
All four organizations went public to deny the use cases or to call them misleading. KPMG removed the report and opened an internal investigation. Their statement says they expect their teams to verify content and check sources before publishing. That clearly did not happen here.
Why AI invents things that sound real
The fabricated stories did not appear by accident. They were spotted by GPTZero, a research group that builds tools to detect AI-generated text. Once GPTZero flagged the inaccuracies, the Financial Times ran the verification with each named organization. The trail leads back to AI hallucinations.
An AI hallucination is when a model invents a fact that looks correct but is not. It happens because models like ChatGPT or Claude are trained to produce text that sounds right, not text that is verified. The output reads like a confident report, even when the underlying facts do not exist.
The KPMG AI report is the kind of document where hallucinations are dangerous. A typo in a chat assistant is easy to laugh off. A fake case study about NHS using AI to do X, signed by a Big Four firm, gets quoted by trade press, by other reports, and by strategy teams inside companies. The error spreads.
This is not the first time. The previous month, EY had to pull a report on loyalty programs for the same reason. Made-up footnotes, citations pointing to sources that never covered the topic, surveys that did not exist. Two Big Four firms tripping over the same problem in weeks tells you the workflow itself is broken, not just one unlucky writer.
The pattern that emerges is awkward for the consulting industry. Firms that sell AI expertise are using AI to write their own publications. The AI hallucinates. The human review misses it. The hallucination is then published under the official logo of a globally trusted brand.
Keep learning on AI Noobies:
- Count Anything: the AI that counts anything you point at
- Google sues “Outsider Enterprise”, a $1.9B AI scam ring
- OpenAI Ohio data center: 10 GW with Nvidia behind it
What this means for anyone using AI
In the short term, the takeaway for a normal user is simple. If a Big Four firm with thousands of analysts can publish a report with invented client stories, your AI assistant can do the same in your own work. The KPMG AI report is a reminder that output that sounds confident is not the same as output that is correct.
The risk hits hardest when AI is used to research, summarize, or cite. If you ask a chatbot to find statistics, to quote a study, or to name companies using a tool, you should expect that some of the names, numbers, or links will be wrong. Verification is back as a basic habit.
In the medium term, AI products are evolving to push back on this exact problem. Some assistants now show their sources next to each claim. Others refuse to invent a quote when they cannot find it. Expect more of this kind of design over the coming months, especially in business tools where mistakes get expensive.
For consulting clients, this episode is also a signal. The shiny PDFs full of percentages and case studies that have been circulating since 2024 must now be read with sharper eyes. A reasonable buyer is going to ask, on every claim, whether a human checked the source. That is the new normal.
The good news is that the same AI tools that hallucinate can be used to catch hallucinations. GPTZero did the work here, but the technique is reproducible. If you write reports, blog posts, or briefs with AI help, run a final pass with a different tool that questions the citations. Two minutes of double check beats a public retraction.
Follow the story on AI Noobies.
About the Author
