A Big Four Firm’s AI Report Flipped Over

Only 5 out of 45 citations were accurate. The problem did not come from some random content account, but from an AI report by KPMG.

Many non-specialist readers in China may not be familiar with the name KPMG. Put simply, it is one of the Big Four accounting firms, alongside PwC, Deloitte, and EY. Do not let the phrase “accounting firm” narrow your impression. These firms do much more than audits. They also provide enterprise consulting, digital transformation, risk management, and, in recent years, AI transformation reports.

That is what makes this case awkward: an institution that makes money from professional judgment published a report about how companies use AI, only for the report’s own citations and examples to appear to have been hit by AI hallucinations.

According to GPTZero’s investigation, the report, titled “Total Experience: Redefining Excellence in the Age of Agentic AI,” had 45 citations. Only 5 accurately pointed to real sources. Another 28 involved rewritten titles, fabricated components, or similar issues, while the remaining 12 were too vague to clearly verify.

This is not a typo-level problem.

If a WeChat article gets a company name wrong, it is embarrassing at most. But if an externally published professional report misstates or distorts other organizations’ AI use cases, the problem becomes larger: readers, media outlets, clients, executives, and later citations may all continue spreading the error.

TechCrunch reported that UBS, the UK’s NHS, Swiss Federal Railways, and Transport for London told the Financial Times that the report’s claims about their AI usage were untrue or misleading. A KPMG spokesperson said the firm had removed the report from its websites while conducting an investigation.

However, KPMG’s public website still shows the Global Customer Experience Excellence 2025-2026 page. That is a useful reminder: we should not inflate “a report was removed” into “it has been completely taken down worldwide,” otherwise we would fall into the same trap ourselves.

The real issue is not exactly how much AI KPMG used. It is the step companies are most tempted to cut when using AI: verification.

Many companies now talk about AI productivity gains by calculating “how much writing time was saved.” Reports, proposals, bids, training materials, public posts, customer-service scripts—AI can indeed make all of these much faster.

But in professional content, the expensive part was never merely writing the words.

The expensive part is checking whether every citation goes back to a real source, whether every case has been confirmed by the organization involved, whether every number has a source, and whether every sentence like “Company X has adopted technology Y” can stand up to scrutiny. AI raises drafting speed, but it does not automatically complete the chain of responsibility.

Many companies have not priced in this cost.

In the past, a serious report involved research, writing, review, legal, brand, and communications checks. It was slow, but people roughly knew which step belonged to whom. After AI enters the workflow, the dangerous part is that the process looks unchanged while output suddenly accelerates, so people start assuming that “if it reads smoothly, it is probably fine.”

It is not.

One thing AI is very good at is writing uncertain content as if it were certain fact. It can produce a plausible title, plausible citations, plausible corporate cases, and even a professional tone. Without going back to the source, it is hard to spot the problem at first glance.

That is why this case is not only about consulting firms.

A technical manager at a small company asking AI to write API documentation, an operations person using AI to write product copy, a founder asking AI to summarize competitors, or a salesperson generating client proposals are all facing the same question: is this content merely usable-looking, or has it actually been verified for external use?

The gap between those two states is not writing skill. It is responsibility.

I have become more conservative about enterprise AI productivity claims. Not because AI should not be used, but because we cannot only ask how many person-days were saved. We also need to ask: was part of the saved writing time spent on verification? If not, the cost has merely been moved from production to the aftermath of a mistake.

Four kinds of content should not be reviewed only at the language level.

First, content that names specific companies or institutions. Second, content that cites data, reports, papers, or policies. Third, content that will be read by clients, bosses, or investors. Fourth, content that may affect contracts, advertising, compliance, or brand reputation.

In these cases, AI can be an assistant, but not the accountable owner.

The irony of KPMG’s case is that it was not just any report. It was a report about agentic AI and how companies use AI to transform customer experience. KPMG’s own page still says customer expectations are accelerating and companies need proactive, predictive experiences to respond.

That may be true.

But the more advanced those words sound, the more they need the dumbest safeguard: open the source, read the original, and ask whether the case actually supports the sentence.

Only 5 out of 45 citations were accurate. The number does not tell us “AI is useless.” It tells us that AI can write things that look too real.

Next time a company says AI improves productivity, do not rush to envy it.

Ask first: who verifies it, and who is responsible if it is wrong?

I am Lao Hua. Follow me for more on AI, tools, and information gaps.