The Big Four sell AI governance. Their own reports cannot pass it.

A report about agentic AI that nobody verified

In October 2025, KPMG published a report titled "Total Experience: Redefining Excellence in the Age of Agentic AI." In June 2026 it quietly pulled the report after the Financial Times and the AI-detection company GPTZero went through its sources. GPTZero checked all 45 citations. Only 5 pointed cleanly to a real source that said what the report claimed. Twenty-eight were paraphrased, fused, or invented outright, and twelve were too vague to verify at all. Around half of the factual claims the citations were supposed to support turned out to be false or misattributed.

The errors are not subtle. The report described case studies of agentic AI at UBS, Swiss Federal Railways, and Transport for London that do not appear to have happened as written. It described Emirates' assistant "Sara" as a mobile chatbot that can change your flight, when Sara is a physical robot introduced in 2023 that does no such thing. It cited a 2019 press release as evidence of agentic AI in production, which is impossible, because commercial agentic AI did not exist in 2019. And in the same document, the report claimed 55 percent of CEOs rank AI as their top investment priority while KPMG's own CEO Outlook, published the same month, put the number at 71 percent. The firm contradicted itself across two of its own October publications.

Here is the point, and it is the same point we made about AI slop, now proven by the firms that sell the cure. The failure was not that AI wrote the report. The failure was that nobody verified the output before it shipped. Verification is the product. The model never was.

This is not a detector story, and the difference is the whole lesson

It is worth being precise about what actually caught KPMG, because the tool that caught it is widely misunderstood, and the misunderstanding is doing real damage in places that matter more than a consulting report.

GPTZero is best known for a statistical classifier: paste in text, and it returns a probability that a machine wrote it, based on how predictable the prose reads. That tool is the one teachers run on student essays. It is also genuinely unreliable for that job. Independent testing has put its false-positive rate as high as 18 percent, and around 8 percent even on short passages. Detection degrades 15 to 20 percent the moment text is lightly paraphrased, and the tool disproportionately flags people who write in a second language as "AI." Johns Hopkins, among others, tells faculty not to treat a detector score as proof of anything on its own. As a machine that renders a verdict on whether a person cheated, it is the wrong tool, and the studies that made you skeptical are correct.

But that classifier is not what found the fabrications in the KPMG report. GPTZero did something completely different and completely boring: they took each of the 45 citations and checked whether the cited source actually exists and actually says what the report claims. That is not a probability. You click the link, the paper is there or it is not. They coined a name for what they found, "vibe citing," references that look credible until someone checks them, and concluded that the pattern was "likely the result of an AI research tool over-complying with a request to find examples of agentic AI in the wild."

Hold those two things side by side, because the contrast is the entire argument:

• Statistical detection guesses whether AI wrote something. It is probabilistic, it has no ground truth, and it punishes the wrong people. This is the tool a teacher points at a kid instead of teaching the kid.
• Source verification checks whether a claim is true. It is deterministic, it has ground truth, and it is exactly the step KPMG skipped. This is the tool nobody pointed at a 45-citation report before it went to clients.

The lesson is not "use AI detectors." It is that the cheap, checkable, unglamorous act of verifying outputs against reality is the thing that separates work you can trust from work that embarrasses you. One side of the story shows a verification step misused as a substitute for judgment. The other shows the absence of any verification step at all.

A fixable process gap, not an indictment of the model

It would be easy to read the KPMG story as proof that AI is not ready, or to read the classroom story as proof that students cannot be trusted. Both readings miss the mechanism.

The model did exactly what large language models do. Asked to find examples of agentic AI in the wild, it produced confident, well-formatted examples, some real, some stitched together, some invented. That is a known property of the tool, not a malfunction. The error was downstream: a publishing pipeline at a global firm that ships AI-governance advice had no step that clicked the links before the document went out under its name. KPMG's own response named the gap precisely. The firm said it expects its people to follow guidelines on responsible AI use, "including human oversight to validate content and verify independent sources." The policy existed. The verification step did not run.

This is not a uniquely KPMG problem, and we are not above it. Deloitte's Australian practice partially refunded a roughly 440,000 Australian dollar government contract in 2025 over a report with the same class of AI-introduced errors. The legal profession now has a running list of court filings with citations to cases that do not exist. The common thread across all of them is not a bad model. It is a missing checkpoint between a confident output and a consequential decision.

How we build the verification step in

The whole value of an AI deployment lives in the layer most people skip: the part that checks the model before anyone acts on it. None of it is exotic. It is the discipline of treating an AI output as a draft to be verified, never a fact to be forwarded. Here is what that looks like in the engagements we run.

Ground every claim in a real source, mechanically. When an AI produces a citation, a figure, or a named example, the system checks it against the actual source before it reaches a human, the same source-verification move that caught KPMG, run automatically instead of after the fact. An unverifiable claim gets flagged, not shipped. A model that cannot ground an answer says so rather than inventing one.

Keep the receipts. Every output traces back to the data and the steps that produced it, so "where did this number come from" has a provenance trail as its answer, not a shrug. This is the difference between catching a fabricated citation in review and discovering it after a client did.

Put a human on the decisions that matter. Routine, recoverable outputs can flow. Anything that goes out under your name, into a contract, a filing, a board deck, or a client deliverable, passes a human checkpoint with the receipts attached, so the reviewer is checking sources, not vibes.

Measure trust, not volume. The metric that matters is not how much AI you used or how much text it produced. It is how often its outputs survived verification. An AI that generates a hundred confident pages, half of them wrong, is not a productivity win. It is a liability with good formatting.

That is the same no-slop discipline we wrote about in our earlier Field Note, Slop is a deployment failure, not a model failure. The KPMG report is that argument's proof: the failure was never the model's intelligence. It was the absence of the system around it.

What this article is not

This is not a knock on KPMG or Deloitte. They are large, serious firms that got caught by a failure mode every organization using AI is exposed to right now, including ours if we drop our own discipline. Naming it is more useful than pretending we are immune.

This is not a claim that AI detectors work, or that schools should run more of them. The opposite. The statistical classifiers are unreliable for judging people, and pointing one at a student is a substitute for teaching, not a form of it. The case for teaching students to use AI well, and to verify what it gives them, is stronger after this story, not weaker.

And this is not a claim that you should never use AI to research or draft. We use it every day. The point is narrower: an AI output is a draft until it is verified, and building the verification step is the actual work.

One-sentence takeaway

The firms selling AI governance just proved the thesis by skipping their own: the model was never the product, verification is, and the cheap deterministic act of checking outputs against reality is the difference between AI you can ship and AI that pulls your report.