KPMG just pulled a report off the internet because parts of it were not true. The report was titled, with no apparent irony, "Redefining excellence in the age of agentic AI." The excellence did not survive contact with the companies it described.
The firm published it in late 2025. This month it came down, after organizations named in it said the claims about their own AI use were false or misleading. The detection group GPTZero traced the errors to AI hallucinations, fabricated statements about real companies, presented as fact in a document carrying one of the most trusted names in professional services.
The Names Make It Worse
This was not a typo in a footnote. UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London all confirmed that what the report said about their AI adoption was either wrong or deceptive. These are not small clients. They are the kind of institutions whose logos a consultancy puts on a slide to prove it understands the market.
A report about how brilliantly the world is adopting agentic AI was itself produced with AI that invented facts about that adoption. The medium undermined the message in the most public way possible. And KPMG is not alone. EY recently withdrew a loyalty rewards report that contained fake footnotes and hallucinated content. Two of the largest advisory firms on earth shipped AI slop in their own deliverables inside the same stretch.
KPMG's response was the corporate liturgy you would expect. The firm expects its people to follow guidelines on responsible AI use, "including human oversight to validate content and verify independent sources." Which is another way of admitting that this time, nobody did.
The Real Cost Is Not the Retraction
Pulling a report is embarrassing. It is not the expensive part. The expensive part is what it does to the firm's core product, which is trust. You do not hire KPMG for labor. You hire it so that when its name is on a number, you can stop checking the number. That is the entire value proposition, and a hallucinated report puts a crack in it.
I have written that being mentioned by AI is not the same as being believed, and this is the supply-side version of the same truth. When the audience learns that your content might be machine-invented, every claim you make inherits a discount. The reader now has to verify what they used to take on your authority. You have handed your own credibility back to them to re-check.
This is the bill for using AI without a verification layer. The productivity is real. A report that took six weeks now takes six days. But if the time you saved goes entirely to producing more, and none of it goes to checking what you produced, you are not faster. You are wrong faster, at a larger scale, with your name attached.
Trust is asymmetric, which is what makes this so dangerous. It is built over years of being right and spent in a single afternoon of being publicly wrong. A consultancy can publish a hundred accurate reports and be known for the one it had to pull. The audience does not average your accuracy. It remembers your worst miss, especially when household-name institutions line up to confirm it.
The Discipline That Separates Winners From Headlines
The lesson is not to keep AI out of your deliverables. That ship has sailed, and it was the right ship to board. The lesson is that AI moved the bottleneck. The scarce, valuable, billable skill is no longer producing the draft. It is judging whether the draft is true.
That means a verification step that is owned, not assumed. Someone whose name is on the accuracy of every external claim, the same way a fact-checker once sat between a journalist and print. Marketing firms already learned this when some got fined for AI theater, dressing up capabilities they could not stand behind. The deliverable is the new front line for the same problem.
Build the check into the workflow, not the apology. Before anything external ships, every statistic, company claim, and citation gets traced to a real source by a human who is accountable for it. This is slower. It is also the only thing standing between your brand and a retraction notice with four blue-chip names confirming you made things up.
In practice that is a rule no external document gets to violate. Every number traces to a primary source a named person checked. Every company claim is confirmed with the company or dropped. Anything the model asserts that you cannot verify does not ship. It is unglamorous, and it is slow by exactly the amount KPMG now wishes it had been.
The firms that win the AI era will not be the ones that generate the most. They will be the ones you can still believe without checking. KPMG just reminded everyone how quickly that asset burns, and how little it takes to light the match.