When AI Hallucinates: Lessons from KPMG's Pulled Report

In mid-June 2026, KPMG removed one of its own thought-leadership reports from its websites. The report, titled "Redefining excellence in the age of agentic AI," had been published in October 2025 and was meant to showcase how leading organisations were adopting so-called agentic AI — AI that can take actions, not just answer questions. It was built around a series of real-world case studies.

The problem, as first reported by the Financial Times, was that several of those case studies did not hold up. KPMG pulled the report around June 12-13, 2026 and opened an internal investigation into how it was produced.

For a firm whose entire value rests on trust and accuracy, quietly deleting a published report is not a small step. It points to something more serious than a typo or an outdated figure.

Some of the world's best-known institutions were named in the report — including UBS, the UK's National Health Service (NHS), Swiss Federal Railways, and Transport for London (TfL). When the Financial Times approached them, they said the report's descriptions of how they were using AI were fabricated, untrue, or misleading. In other words, the report described things that simply had not happened.

The AI-detection firm GPTZero then ran a forensic review of the report's sources, and the result was striking. Of the report's 45 citations, only 5 correctly pointed to the source they claimed to cite. The remaining roughly 40 were mangled, misleading, partially fabricated, or too vague to verify.

That pattern — confident-sounding references that fall apart the moment you check them — is the classic signature of AI hallucination. The footnotes looked authoritative, but a large share of them led nowhere real.

The uncomfortable irony is hard to miss: a report about the responsible use of agentic AI appears to have itself been substantially AI-fabricated. And it was not an isolated event. EY recently withdrew a report of its own that contained fake footnotes and AI hallucinations.

This is the core risk that businesses keep underestimating. Modern AI does not signal uncertainty the way a nervous junior analyst would. It produces fluent, well-structured, professional-sounding output whether or not the underlying facts are true. A fabricated case study reads exactly like a real one.

A KPMG spokesperson said the firm expects "all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content and verify independent sources." That is the right principle — but the episode shows how easily even a sophisticated organisation can publish AI output that nobody fully verified.

It is tempting to read these stories as a warning to simply use AI less, or to wait for a smarter model that hallucinates less often. But that misses the real lesson. The reports failed not because the AI was not clever enough, but because it was not grounded in real, verifiable, traceable sources. When AI has nothing solid to draw from, it fills the gaps with plausible-sounding inventions.

The difference between a trustworthy AI answer and a hallucinated one usually comes down to a single question: can you trace it back to a real record? A claim about a customer, a service interaction, or a financial figure is only as reliable as the data behind it — and only checkable if that data actually exists in a system you can point to.

For care and NDIS providers especially, where AI might one day summarise client records, flag funding issues, or draft reports, an invented "fact" is not just embarrassing — it can be a compliance and care-quality risk. The safeguard is not blind trust in the model. It is grounding every answer in real, governed data.

The KPMG and EY incidents are a clear signal for any organisation rushing to adopt AI: the foundation matters more than the tool. AI that is wired into your real, governed business data can be checked, traced, and corrected. AI that is left to improvise will confidently make things up — and look polished while doing it.

This is exactly why Intellova's approach starts with the data layer, not the model. By unifying the information scattered across your CRM, accounting, and other systems into one governed, traceable database, you give AI and analytics a single source of truth to draw from — where every figure can be tied back to a real record rather than a confident guess.

Before you ask AI to write your next report, ask a simpler question: is it answering from your real data, or inventing its own? A unified, AI-ready data foundation is what turns that question from a gamble into a guarantee.