The Consultant Hallucinated the Case Study

The Consultant Hallucinated the Case Study

There are business scandals that require a long explanation, and then there are scandals that arrive fully gift-wrapped.

KPMG, one of the world’s largest professional services firms, published a report about agentic AI. The report was supposed to help explain how organizations are using AI to improve customer experience. Instead, according to reporting by the Financial Times and TechCrunch and a detailed investigation by GPTZero, it became something much more efficient: a live demonstration of why people should not trust polished AI material without checking the facts.

The report, titled “Total Experience: Redefining Excellence in the Age of Agentic AI,” was published in October 2025. GPTZero later examined its citations and claims. The findings were not a minor typo. GPTZero said that of the report’s 45 citations, only five accurately pointed to real sources. It said 28 citations had paraphrased titles or fake components attached to real sources, while 12 were too vague or flawed to verify confidently. It also said that around half of the claims supported by those citations appeared to be fake or misattributed.

That would already be bad for a student paper. It is more awkward when the author is a global consulting firm selling credibility, diligence, and transformation advice to executives.

Then came the part that makes the story almost too clean for satire. TechCrunch reported that KPMG pulled the report after multiple organizations named in it said the claims about their AI usage were untrue. UBS, the U.K.’s National Health Service, Swiss Federal Railways, and Transport for London all reportedly told the Financial Times that claims about their use of AI were either untrue or misleading. A KPMG spokesperson said the firm had removed the report while it reviewed the circumstances around publication and expected its people to follow responsible AI guidelines, including human oversight and source verification.

That statement is sensible. It is also exactly the kind of statement a firm usually gives after the thing it advises other people to control has escaped from its own building.

The Evidence Looked Finished

The danger with AI hallucinations has never been that they look ridiculous. The danger is that they look professionally formatted.

A fake citation does not usually arrive wearing a clown wig. It arrives with a title, a date, a publisher, and the calm confidence of a document that has passed through a template. It sits in an endnote. It borrows the authority of a brand. It gives the reader just enough comfort to keep moving.

GPTZero described the pattern as “vibe citing.” The phrase is irritatingly perfect. It captures the new academic and corporate ritual of producing references that feel like citations without doing the basic job of citations. They gesture toward evidence. They imitate the furniture of verification. They are not necessarily clean inventions from nowhere. Some appear to be warped versions of real sources, with titles paraphrased, authors confused, dates shifted, or claims attached to materials that do not actually support them.

That makes the problem harder to catch and easier to excuse. A fully invented source is fraud in a fake mustache.

A mangled source can be defended as a formatting issue, a workflow error, or an unfortunate compression of research notes. The surface looks close enough to reality to survive casual inspection.

That is where the real story lives. AI did not simply make up a few facts. It appears to have produced a version of business reality that was plausible enough to publish.

The Case Studies Were the Product

Consulting reports are not neutral containers of information. They are credibility machines.

A report like this does several jobs at once. It tells executives that a trend is real. It tells buyers that competitors are moving. It creates examples that can be used in presentations. It gives sales teams a language for urgency. It turns a vague market mood into a set of apparently documented facts.

That is why the disputed case studies are so damaging. If the report had merely misstated a statistic, the embarrassment would be narrower. But case studies are not decorative. They are the emotional proof of transformation. They make an abstract promise feel operational. They allow a board member, procurement lead, or chief digital officer to think, “Someone like us is already doing this.”

According to GPTZero’s analysis and subsequent coverage, several examples allegedly stretched or misrepresented what organizations were actually doing with AI. The Register highlighted GPTZero’s discussion of an Emirates example, where the report allegedly described a mobile chatbot named Sara that could converse with passengers and change flights. GPTZero said Sara was a robot assistant introduced by Emirates in 2023, not a chatbot, and lacked the flight-changing capability described.

This is the kind of error that sounds small until one remembers how these documents are used.

A nonexistent capability becomes a benchmark. A misdescribed experiment becomes proof of adoption. A customer-service robot becomes an agentic AI system. The market advances, at least on paper.

Paper is where many transformations begin.

The Most Expensive Version of “Trust Me”

The consulting industry runs on a strange bargain. Clients know they are buying frameworks, interpretation, and borrowed confidence. Consultants know they are selling structure in situations where the client is often uncertain, politically exposed, or moving too slowly. Everyone understands the ritual, but the ritual depends on a basic floor of reliability.

A report does not need to be perfect to be useful. It does need to show that someone took evidence seriously.

That floor is now wobbling. The KPMG episode follows other professional-services embarrassments involving AI-generated or AI-assisted errors. EY recently withdrew a report on loyalty rewards programs after GPTZero identified apparent hallucinations and fake footnotes. Deloitte previously had to address AI-related errors in a report for the Australian government. The pattern is no longer a quirky warning about lazy students. It is moving through firms that are paid to reduce risk for other people.

This is why the story has such comic force. The people selling AI readiness are now giving us readiness failures. The people telling clients to establish governance, controls, review layers, and source validation are discovering that those words do not verify citations by themselves.

There is a wonderful bleakness to that. The AI transformation industry has spent years telling companies that adoption must accelerate. Now the same industry is being reminded that acceleration without verification can simply move nonsense faster.

The Report Became a Procurement Hazard

Enterprise buyers should pay attention to the mechanics of the failure. A hallucinated consulting report is not just a reputational problem for the firm that published it. It can become an input into downstream decisions. A fake case study can be quoted by another report. A distorted statistic can appear in a deck. A misleading example can be repeated by a vendor. A claim that began as a hallucination can be normalized through repetition until it starts looking like industry consensus.

City A.M. reported that flawed statistics and claims from the KPMG report had already been recycled by industry publications and were surfacing in responses from large language models.

That is the contamination problem. Once synthetic evidence enters the public knowledge stream, it can be reabsorbed by other systems and redistributed with new confidence.

For procurement teams, this creates an ugly question. When a vendor or advisor brings a market report into the room, what exactly is being evaluated? The insight? The brand? The data lineage? The diligence behind the examples? The answer used to be fuzzy because the risk felt manageable. In an AI-saturated research environment, fuzziness becomes expensive.

The new rule should be uncomfortable but simple. If a report is used to justify enterprise AI spending, the evidence behind the report deserves audit treatment. Not a quick glance. Not a trust-based review. Not a decorative source list. The claims, citations, case studies, and named-organization examples need to survive contact with reality.

A consulting logo is not a verification protocol.

Agentic AI Did Not Need This Help

There is a deeper irony here. Agentic AI is already one of the most inflated phrases in the market. It can mean a meaningful software system with tool access, task planning, memory, constraints, and execution authority. It can also mean a chatbot with a new label and a sales deck.

That ambiguity is useful for marketing and dangerous for governance. Companies want the excitement of agents without the burden of explaining what autonomy actually exists. Vendors want to describe “agentic” systems in ways that make conventional automation sound newly alive. Advisors want to show that the market has moved from experimentation to operational adoption.

A report full of shaky case studies does not merely make factual errors. It pushes the market toward premature certainty. It tells executives that agentic AI is already everywhere, already proven, already delivering.

That pressure can distort procurement. It can make cautious leaders look slow. It can reward theatrical adoption over controlled implementation.

The KPMG situation is therefore not only a story about citations. It is a story about evidence inflation. The market wants proof that AI transformation is working. AI can help manufacture the smell of proof. Consulting templates can make that smell presentable. Then everyone gathers around the document and pretends the smoke is a signal.

Human Oversight Is Not a Magic Phrase

Every AI governance failure eventually reaches the phrase “human oversight.” It sounds responsible. It is easy to say. It fits nicely into policy documents and public statements.

The problem is that human oversight is not a ritual blessing. It is work.

Someone has to open the source. Someone has to check whether the title exists. Someone has to confirm whether the cited document supports the claim. Someone has to contact the organization if the report attributes a significant AI deployment to it. Someone has to notice when a 2019 railway press release is being used to imply agentic AI adoption before the phrase had become common market language. Someone has to ask whether a customer-service robot has been promoted into a chatbot with powers it does not have.

That work is boring, which is why AI is so tempting. AI makes the boring part appear done. It fills the page, finds plausible references, smooths the language, and delivers a document that looks ready for review. The more finished it looks, the less likely a busy professional may be to treat it as raw material.

That is the trap. AI does not always fail by producing obvious garbage. Sometimes it fails by producing something that looks like the last draft.

The Industry Now Needs Proof of Proof

The lesson for enterprises is not that consulting reports are useless. The lesson is that authority now needs chain of custody.

When a report says a company adopted an AI system, there should be a traceable basis for that claim. When a citation appears, it should lead to the source it names. When a statistic is attributed to internal research, it should match the underlying publication. When a case study is used to show market maturity, the named organization should recognize the description.

This will make reports slower to produce. Good. Some things should be slower.

The AI market has been drowning in claims about productivity, transformation, disruption, acceleration, and autonomous workflows. Much of that language already floats above operational evidence. If leading advisory firms allow hallucinated citations and disputed case studies into that environment, they do not merely embarrass themselves. They lower the quality of the market’s shared reality.

That is a serious problem hiding inside a very funny story.

The Mirror Worked

The funniest version of this story is obvious. A major consultancy published an AI report that reportedly hallucinated evidence about AI. It then had to pull the report and remind everyone that its people are expected to verify sources.

That joke writes itself.

The less funny version is more useful. AI is not just changing how companies produce content. It is changing how business claims acquire authority. It can turn weak research into confident prose. It can turn partial evidence into apparent certainty. It can turn the market’s desire for proof into a supply of proof-shaped objects.

KPMG did not invent that problem. It merely became the cleanest example of it.

The report was supposed to explain the age of agentic AI. Instead, it explained the age of synthetic confidence. The future arrived with citations. Most of them apparently needed checking.