Companies are using AI more than ever in their workflows, but not all uses are going as per plan.
KPMG has withdrawn a global report on agentic AI, titled “Total Experience: Redefining Excellence in the Age of Agentic AI,” after an investigation found that most of its citations were fabricated. The report, published in October last year, was meant to showcase how organisations around the world were putting autonomous AI systems to work. Instead, it has become a case study in what happens when a consulting firm uses AI to write about AI without bothering to check what the AI wrote.
The unraveling started with AI detection firm GPTZero, which ran the report through its system and found that out of 45 citations, only five pointed to real, verifiable sources. The remaining 40 were either invented outright, paraphrased from real papers with fabricated details bolted on, or worded so vaguely that they couldn’t be traced to anything at all. GPTZero researcher Paul Esau didn’t mince words about what that implies, saying the firm suspects no human at KPMG checked the citations, the claims, or the sources before the report went out.
The fabrications weren’t limited to footnotes. Several of the organisations named in the report as case studies pushed back once they read what KPMG had written about them. Transport for London disputed claims that it used AI agents to predict congestion and run the city’s transport network. NHS Greater Manchester said the report’s description of it using AI agents to organise patient records, automate referrals, and predict hospital readmissions didn’t align with what the trust actually does. UBS and Swiss Federal Railways raised similar objections, according to the Financial Times.
One example singled out by GPTZero involved Emirates. The report claimed the airline runs a chatbot called Sara that lets passengers talk to it and change their own flight bookings. Sara is real, but it launched in 2023 as a mobile assistant, and it was never built to alter bookings on its own. GPTZero’s read on this is that the errors look like the output of an AI research tool that kept inventing examples once it ran out of real ones to cite, a pattern its researchers call “vibe citing,” where a reference looks plausible at a glance but falls apart the moment anyone checks it.
There’s also a smaller, almost comic detail buried in the numbers. The report cited “KPMG research” claiming that 55% of chief executives ranked AI as their top investment priority. KPMG’s own CEO Outlook survey, published the same month, put that figure at 71%. The firm’s AI-assisted report couldn’t even get KPMG’s own data right.
KPMG has pulled the report from its websites and opened an internal investigation. A spokesperson said the firm takes the accuracy and integrity of its publications seriously and that employees may have breached internal guidelines on AI use, which require human oversight to validate content and verify sources independently.
This is the second time in recent months that a Big Four firm has had to retract research over AI-generated errors. EY pulled a loyalty-rewards study last month after it turned up with fake footnotes, also flagged by GPTZero. Deloitte had to refund part of a $440,000 contract with the Australian government earlier after a report it produced included a fabricated court quote and citations to academic papers that don’t exist.
The pattern extends past consulting. Courts have been catching lawyers submitting fake AI-generated citations for two years now, with fines running into the tens of thousands of dollars. Even peer-reviewed research hasn’t been spared, with hallucinated citations turning up in papers accepted at NeurIPS, one of the most selective conferences in AI.
What makes the KPMG episode worth paying attention to is reach. Reports from firms of this size get picked up by trade press, cited in other research, and increasingly absorbed into the training data and live answers of AI chatbots themselves. GPTZero has pointed out that a single unverified document can work its way into the answers ChatGPT or Gemini give users on a topic, long after the original report has been quietly taken down. KPMG can pull the PDF from its own website. Getting it out of everywhere else it has already spread is a different problem entirely.