A Big Four firm published a flagship report on agentic AI. Then it quietly took the report off its own websites. The reason is one of the clearest teaching moments any leader will get this year, and it has almost nothing to do with whether AI is ready for your business.
In October 2025, KPMG International released a report called Total Experience: Redefining Excellence in the Age of Agentic AI. By mid-June 2026, the report was gone from KPMG's sites. Named companies had started disputing what the report said about them. An AI-detection firm had gone through it line by line. And the verdict was hard to argue with.
Of the 45 citations in the report, only 5 pointed to real, uncorrupted sources.
Source: GPTZero, 2026What happened
The detection company GPTZero investigated the report and manually checked every citation. They traced the errors back to AI hallucinations. In plain terms, a language model appears to have written large parts of a report about AI, and no human checked the model's work before it went out the door.
GPTZero gave the pattern a name: "vibe citing." The references look convincing. Real-sounding titles. Plausible authors. Fragments stitched together so they read like genuine scholarship. They hold up right until someone clicks the link and finds the source does not say what the report claims, or does not exist at all.
The specific cases are worth seeing, because they show how confident a wrong citation can look.
- Emirates. The report described a mobile chatbot named "Sara" that could change flight bookings. The real Sara is a robot assistant the airline introduced in 2023, with no power to alter a booking.
- Verbund. The case study mixed up the Austrian utility with a Swedish startup called Qurrent, then invented details about household AI agents managing smart appliances and EV charging.
- JR East. The supporting citation was a 2019 press release, written before commercial agentic AI existed, used to claim the railway runs AI agents for travel recommendations.
- Toyota, UBS, KPN, NHS Greater Manchester, Transport for London. Each had a source that either contained no mention of AI agents, described no agentic behavior, or referred to human employees rather than software.
Several of the organizations named in the report, including UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London, told reporters the report's claims about their AI use were untrue or misleading.
There was one more tell. The report stated that 55% of CEOs rank AI as a top investment priority. KPMG's own 2025 Global CEO Outlook, a survey of more than 1,300 CEOs released the same month, put the number at 71%, up from 64% the year before. The firm's flagship AI report contradicted the firm's flagship CEO survey, and nobody caught it before publishing.
KPMG removed the report and said it is reviewing how it got published. Its statement is the most useful sentence in this whole story:
"We expect all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content and verify independent sources."
Read that again. The company that got burned named the exact step that would have prevented it.
The lesson is the opposite of "AI is not ready"
Here is where most people draw the wrong conclusion. They read this and think the takeaway is that AI cannot be trusted with work. That reading is comfortable, and it lets a hesitant leader off the hook. It is also wrong.
The tool did its job. It drafted fast. It produced clean prose and a long reference list. What was missing was a person between the draft and the publish button.
Think about who got hurt here. Not the firms that avoid AI. The firm with every resource on earth, using AI without an oversight step. That is the pattern across these stories. KPMG is not an outlier. Last October, Deloitte's Australian practice agreed to partially refund a roughly AU$440,000 (about US$290,000) government report that carried AI-generated errors, including academic papers that did not exist and a quote from a court judgment that was never said. Same root cause. A capable tool, no verification layer, work shipped under the firm's name.
The companies getting embarrassed are not the ones experimenting with AI. They are the ones who skipped the part where a human reads it.
What this means for the work you own
If you run a team, a department, or a small business, you do not need to become technical to apply this. You need one rule, and you need it enforced.
AI can draft. A person verifies every fact and every source before anything leaves the building under your name.
That is the whole policy. It sounds obvious. KPMG and Deloitte both skipped it, so clearly it is not obvious enough to survive deadline pressure on its own. The verification step is what stands between useful output and a public liability with your name on it.
A few practical pieces of that, in language your team can act on this week:
- Name an owner for every AI-assisted deliverable. One person is accountable for checking it, by name, before it ships. Not "the team." A person.
- Treat every citation as a claim, not a fact. If the AI produced a source, someone opens the link and confirms it says what the draft says. A reference that nobody clicked is a reference nobody can stand behind.
- Cross-check numbers against what you already know. KPMG's 55% sat next to its own 71%. A thirty-second sanity check against existing data would have caught it.
- Make the human step visible. When verification is a named, expected part of the process, it survives a tight deadline. When it lives in someone's head as "I'll glance at it," it disappears the first busy week.
In the 7 Levels of AI Proficiency, this is not an advanced skill. It sits around Level 3, the Critical Thinker, the point where a person stops trusting the output and starts checking it. The firms in this story have the budget to operate several levels higher. They got caught because proficiency is about the workflow a person runs, not the tools they can afford.
A next step you can take today
Pick one thing your team is about to publish with AI's help. A report, a proposal, a client email, a deck. Before it goes out, ask one question: who personally checked every fact and clicked every source in this?
If the answer is a name, you have an oversight layer. If the answer is a shrug, you just found the most expensive risk in your AI rollout, and the cheapest one to fix.
Related reading: Level 3: The Lieutenant (Critical Thinker).
Sources
- KPMG's Disappearing Report: A Case Study in AI Hallucinations (GPTZero)
- KPMG pulls report on AI usage due to apparent hallucinations (TechCrunch)
- KPMG's AI report turns into a demo of AI hallucinations (The Register)
- Global CEOs double down on AI and talent (KPMG 2025 Global CEO Outlook)
- Deloitte to partially refund Australian government for AI errors (The Register)
- Deloitte will refund the Australian government after using AI (Fortune)
Frequently Asked Questions
Did a client pay for the KPMG report?
There is no reporting that a client commissioned or paid for this specific report. It appears to be self-published thought leadership. No refund has been reported. The Deloitte case is the paid example, where an Australian government report was partially refunded.
Does this mean AI writing tools are unreliable for business?
The failure here was a missing human review step, not the tool. The same AI that produced the bad citations can produce useful drafts when a person verifies the facts and sources before publishing. The reliability comes from the workflow around the tool.
What is "vibe citing"?
It is GPTZero's term for AI-generated references that look real, with plausible titles and authors, but do not support the claim or do not exist when you check them. They pass a quick glance and fail an actual click.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.