KPMG Published a Report on AI's Benefits. It Was Reportedly Full of AI Hallucinations.

What the Report Said

In October 2025, KPMG published a report called "Total Experience: Redefining Excellence in the Age of Agentic AI." The audience was exactly who you would expect: business leaders, operations executives, and digital transformation teams trying to figure out how aggressively to bet on AI agents — the kind of software that doesn't just answer questions but actually executes tasks, manages workflows, and makes decisions without human sign-off at every step.

The report made the case that agentic AI was already delivering measurable results across industries. It pointed to case studies and cited external sources to back that argument up. For a firm the size of KPMG — one of the four largest professional services organizations in the world, with a client list that includes major financial institutions, government agencies, and Fortune 500 companies — publishing on this topic carried real weight. When KPMG tells clients the agentic AI wave is here and the adoption numbers prove it, those clients pay attention. They share the report internally. They cite it in strategy presentations. They use it to justify budget decisions.

That institutional credibility is precisely what made what came next so significant. The report was not a blog post or a think piece. It was positioned as research — with citations, with sourced examples, with the kind of evidentiary scaffolding that signals "we checked this."

What GPTZero Found

GPTZero published its forensic review on June 12, 2026, the same day the Financial Times and The Register broke the story. The finding was specific and quantifiable: of the 45 citations in KPMG's report, only 5 accurately pointed to real sources. The remaining 40 were, in GPTZero's assessment, some combination of fabricated, mangled, or simply unverifiable — citations that looked legitimate on the page but fell apart the moment anyone tried to trace them back to an actual document or data set.

That 5-out-of-45 figure is the number worth sitting with. This was not a report where a few footnotes got sloppy or a secondary source was cited imprecisely. More than 88 percent of the citations failed basic verification. The case studies used to illustrate AI adoption across industries — the examples a business executive would read and use to benchmark their own organization — were built on a foundation that GPTZero's review describes as misleading, partially fabricated, or too vague to confirm.

The Financial Times coverage framed the core problem clearly: the report exaggerated AI adoption by businesses, using examples that could not be traced to anything real. The Register covered the same findings. Neither outlet reported any official response from KPMG at the time of publication.

The Irony Is Hard to Ignore

KPMG is not a firm that stumbled into the AI risk conversation last year. For years, consulting firms at this level have been the ones standing in front of enterprise clients explaining why AI hallucinations are a board-level concern — why you cannot deploy a language model in a high-stakes workflow without verification layers, human review, and governance frameworks. That has been the service. That has been the billable engagement. "Here is how AI can go wrong, and here is how we help you prevent it."

The October 2025 report inverts that entirely. A firm whose consultants routinely advise clients on AI risk controls shipped a document where more than 88 percent of citations failed basic verification. The case studies executives were meant to benchmark their organizations against were, according to GPTZero's review, built from fabricated or unverifiable sources.

That is not an editorial lapse or a footnoting error. It is precisely the failure mode KPMG has been paid to help clients avoid.

No editorializing is needed here. The facts land on their own. The firm that warns you about hallucinations published a hallucinated report. The report was about AI. The irony is not subtle, and it is not lost on the practitioners who have been citing guidance like this in their own strategy decks.

What This Changes for Practitioners

The practical takeaway is not "stop using AI for research." It is "stop skipping the verification step that would have caught this."

GPTZero's review found 40 of 45 citations either fabricated, mangled, or unverifiable. A single fact-checker with a browser and two hours would have surfaced that problem before the report went public. That is the human-in-the-loop review that did not happen. Not a sophisticated governance framework. Not a proprietary hallucination detection system. A person checking whether the citations pointed to real things.

For practitioners who rely on third-party research to brief leadership, justify budget decisions, or inform AI strategy, the verification question now has to be explicit. When a report cites a case study, trace it. When it offers an adoption statistic, find the source. When the sourcing is vague, treat the claim as unverified — regardless of the logo on the cover.

This applies to your own AI-assisted work too. If you are using a language model to draft research summaries, pull supporting data, or generate citations, the output is a starting draft, not a finished deliverable. The model does not know when it is wrong. You have to be the check.

KPMG's report did not fail because AI was involved. It failed because no one apparently verified what the AI produced before it went out the door.