In October 2025, KPMG published a global study on agentic AI, built in part around named case studies of companies putting AI agents into production. In 2026 the firm withdrew it, after UBS, NHS Greater Manchester, Swiss Federal Railways, and Transport for London separately disputed what the report said about their own use of AI.
An independent forensic review of the report's sourcing, run by the detection firm GPTZero, found that of forty-five citations supporting the report's claims, five pointed cleanly to a real, intact source. The remainder were paraphrased into something the original did not say, attached to the wrong document, or impossible to trace at all. In one case, a blog post about Transport for London was attributed to Transport for London itself, the model having confused the subject of an article with its author.
What the review found
Source: GPTZero forensic review, as reported by the Financial Times and others, June 2026.
The standard reading of this episode is that a consulting firm used AI to research AI, and the AI made things up. That is true, and it is also where most of the commentary has stopped. Treating this as one failure rather than two is a mistake, because it leaves the more dangerous version of the problem completely unaddressed.
Two ways a research pipeline can be wrong
The first failure mode is invention. It appears when a model is set loose on an open-ended task — find examples of companies doing something, with no fixed, bounded set of permitted sources to draw from. With nothing constraining it to a specific, retrievable document, the model is not actually retrieving anything. It is producing an object with the shape of a citation: a plausible title, a plausible author, a plausible publication, assembled from pattern rather than from a source that exists.
This is, in a specific sense, the easier failure to live with. It produces an artifact — a title, a quote, a named source — that either exists or does not. Anyone with twenty minutes and a search engine can run down forty-five citations and find out which five survive contact with reality. That is exactly what happened here. The failure was real and the consequences were public, but it was also, eventually, checkable from the outside, by people who had never seen KPMG's internal process and didn't need to.
The failure that leaves nothing to check
The second failure mode shows up in the opposite kind of system — one built specifically to prevent the first problem, by constraining the model to operate only over a fixed, defined, pre-vetted set of documents: an internal evidence base, an audited filing set, a contracted research corpus. This is the direction most serious enterprise AI deployments are already moving in, for exactly this reason. If a model can only cite something that demonstrably exists inside a known corpus, inventing a source from nothing becomes very difficult.
It does not become difficult to leave something out. A retrieval step can surface eight of the ten relevant documents and never return the two that would have changed the conclusion. A summarisation pass can quote every sentence it keeps with perfect fidelity and simply not keep the one clause that mattered. None of the citations in that output are fake. Every one is real, attributable, and accurate. What's wrong is not on the page at all, and there is no fabricated name or garbled title for a reviewer to trip over — because the question that would catch it is not "is this citation real," it is "is this everything that should have been here," and almost nothing in a standard review process is built to ask that second question.
Why this matters for where review actually gets built
KPMG's failure, now public and now precedent, will make citation-checking and source verification the obvious next line item on enterprise AI governance checklists. It should. But it would be a mistake to treat that as the finished governance programme, because citation-checking only catches the failure mode that produces a checkable citation in the first place. It does nothing for a system that was specifically engineered to avoid inventing sources, and quietly, undetectably, leaves things out instead.
Catching omission requires a different design question, asked earlier in the pipeline: before generation happens, what does complete coverage of this question actually require, and was that coverage achieved at retrieval. That is a harder problem to specify than "does the citation exist," because someone has to define, in advance, what a complete answer would have needed to include — a judgment that requires real domain expertise in the subject of the query, made before the model runs, not a fact-check applied to what it already produced.
None of this is a claim about what happened inside KPMG's own process specifically. The firm's internal workflow has not been made public, and the most that can be said with confidence is what the published report shows and what the disputing organisations have stated on the record. The argument here doesn't depend on knowing KPMG's internal process. It depends only on a structural fact that holds regardless of what produced this particular report: hallucination and omission are different failures, they require different controls, and the case everyone is currently discussing happens to be the more visible, more easily checked, and — for that reason — less dangerous of the two.
The organisations now adding a "verify your AI's citations" step to their process will have fixed the failure that already got caught in public. The harder work, and the one with no news cycle attached to it, is building a coverage check into the step before generation starts — the only point in the pipeline where an omission can still be seen, because by the time the output is finished, the thing missing from it has already stopped looking like a gap.