A Pulled KPMG AI Report Puts Its Trust Pitch on Trial

If KPMG can't verify an AI report about AI, why should clients accept its AI governance advice without seeing the controls?

That is the question this pulled report raises. Not whether generative AI hallucinates. Everyone serious already knows it does. The sharper issue is whether firms that sell trust, assurance, and AI strategy can prove their own AI-assisted work is being checked before it reaches clients, executives, journalists, and search engines.

KPMG removed a report titled “Redefining excellence in the age of agentic AI” after multiple organizations said its claims about their AI usage were untrue or misleading, according to TechCrunch. The report was published in October 2025. Research group GPTZero identified inaccuracies and told the Financial Times they appeared to stem from AI hallucinations.

That makes this more than a publishing mishap. It is an accidental audit of the AI consulting business itself.

Why does a hallucinated AI report hit harder when KPMG is the publisher?

Because KPMG is not a random content shop chasing search traffic. It is a professional services firm whose commercial pitch rests on rigor.

Clients pay firms like KPMG for discipline: reviewed claims, defensible sourcing, clean methodology, and risk controls that survive scrutiny. So when a report about agentic AI appears to contain false or misleading claims about AI adoption, the reputational damage comes from the mismatch. The firm selling judgment appears to have published material that needed more judgment.

TechCrunch reported that UBS, the UK’s National Health Service, Swiss Federal Railways, and Transport for London told the FT that the report’s claims about their AI usage were either untrue or misleading. Those are not minor formatting errors. They go to the core of what the report was supposed to document: who is using AI, how, and with what level of maturity.

KPMG’s public response was the right first move.

“We expect all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content and verify independent sources,” a KPMG spokesperson said.

That statement is also the problem. If those guidelines existed, the open question is whether they failed, were ignored, or were too weak to catch the issue.

Was the scandal the use of AI, or the absence of proof?

Using AI to assist research or drafting is not the scandal. Publishing unverified claims is.

Agentic AI, in this context, refers to AI systems marketed as able to take more autonomous steps toward a task than a simple chatbot response. That topic already attracts inflated claims. A report about it needs stricter evidence, not looser sourcing.

The Register reported that GPTZero claimed only five of the report’s 45 citations correctly matched their sources, while the rest were “mangled and misleading,” partially fabricated, or too vague to verify. GPTZero described the pattern as “vibe citing”, a neat term for a familiar failure: references that look plausible until someone actually checks them.

That is the ugly lesson. Hallucinations are not rare glitches from alien machinery. They are predictable outputs from systems optimized to generate convincing language. If the workflow rewards speed and polish, the machine will deliver both, whether the underlying claim is sound or not.

A serious AI report needs boring controls:

• Source tracing: Every factual claim should map to a live, relevant source.
• Citation checks: Titles, authors, dates, and links must match the cited material.
• Quote verification: Direct claims from named organizations need confirmation.
• Data validation: Statistics should reconcile with the source cited.
• Human accountability: A named person or team must own final factual accuracy.

None of that is exotic. It is the minimum.

What happens when confident AI research outruns verified evidence?

Bad reports do not stay neatly inside PDF files.

City A.M. reported that GPTZero found 40 out of 45 citation titles in KPMG’s report were fake, and that about half of the factual claims supported by those citations appeared false or misattributed. It also reported that flawed statistics and claims from the KPMG report had been recycled by industry publications and a Czech newspaper, and were being cited directly by ChatGPT and Gemini.

That is how an error hardens into business reality. A false claim gets published. A secondary outlet repeats it. A model ingests or retrieves it. A future user sees it in a generated answer and treats it as corroboration. The original mistake gains authority by repetition.

This is the same credibility problem that runs through other AI deployment risks XOOMAR tracks, even when the operational details differ. Our coverage of AI Power Crunch Pulls GM and Ford Into Energy Storage dealt with physical infrastructure strain, not hallucinated research. But the governance lesson rhymes: AI adoption creates second-order risks that companies cannot manage with slogans.

The consulting industry’s incentive problem is obvious. AI thought leadership rewards speed, authority, and confident packaging. Careful validation slows the machine down. But reports on AI adoption can shape boardroom decisions, procurement choices, risk policies, and budgets. Accuracy is not a newsroom nicety here. It is part of the product.

Does pulling the report settle the matter?

No. It only starts the accountability process.

KPMG deserves credit for removing the report while it investigates. Leaving questionable material online would have been worse. A quiet takedown, though, does not answer the questions clients should now be asking.

• Process: Who reviewed the report before publication?
• Tools: Was generative AI used to draft, summarize, source, or edit the report?
• Controls: Were citations independently checked?
• Ownership: Who signed off on the final version?
• Correction: Will KPMG publish a clear note explaining what was wrong?

The strongest counterargument is that organizations make mistakes, and the responsible move is to correct them. Fair. But that defense is weaker when the organization sells governance around the very technology involved in the failure.

TechCrunch also noted that EY withdrew a report last month on loyalty rewards programs that appeared to include fake footnotes and AI hallucinations. The Register pointed to Deloitte refunding the Australian government after AI-generated content slipped into a taxpayer-funded report. The pattern matters because each case chips away at the same assumption: that brand reputation alone guarantees verification.

It doesn’t.

What should companies demand before the next AI report reaches the boardroom?

Companies should not ban AI from research. That would be lazy governance dressed up as caution. AI can help draft, summarize, cluster information, and accelerate early-stage analysis.

But it cannot borrow institutional credibility without earning it.

Any advisory firm, publisher, or corporate team producing AI-assisted research that could influence decisions should adopt clear rules:

This is also where internal AI governance needs to connect with operational planning. As XOOMAR noted in AI Power Crunch Pulls GM and Ford Into Energy Storage, AI’s consequences do not stop at the model layer. They show up in infrastructure, capital allocation, and risk management. In KPMG’s case, they showed up in trust.

The practical takeaway is blunt: don’t accept an AI report because the logo is famous. Ask how it was made. Ask who checked it. Ask whether the citations were verified one by one.

The firms telling everyone else to govern AI now have a simpler task in front of them: prove they can govern their own work.

Impact Analysis

• KPMG’s credibility depends on rigorous verification, making alleged AI hallucinations in its own AI report especially damaging.
• The incident raises questions about whether AI consulting firms are applying adequate controls to their own AI-assisted work.
• Organizations named in AI adoption reports may face reputational risk when unsupported or misleading claims spread through major publications.