KPMG pulled an AI report after 40 of its 45 citations failed a check
Only 5 of 45 citations in a KPMG report on AI agents matched their sources, GPTZero found. KPMG removed it; marketing now vets reports with a checker.
At a glance
- Company
- KPMG International · Professional services · Size not stated
- Team and use case
- Marketing · Content production
- Tools
- AI tool used on the report: not disclosed, Hawk (in-house hallucination checker, added afterwards)
- Stage
- Paused or rolled back, since Published October 2025; removed June 2026
- Results
- Citations that pointed to the source they named: 5 of 45 (GPTZero's review)
- Factual claims false, unsupported or misattributed: Roughly half (GPTZero's allegation)
- Time the report was public before removal: About 8 months
- Report status: Published → Removed, under review (June 2026)
- Cost
- Not disclosed. No figure for the report, the review or the Hawk checker.
- People
- Authors and approvers are not named and no action against staff is reported. KPMG's guidelines required human verification of sources. Marketing now vets thought leadership with Hawk; who signs off after it is not disclosed.
- Evidence
- Independently measured
An academic study, an auditor, a regulator filing or another party with no stake in the result.
What this leaves out: GPTZero has no tie to KPMG but sells hallucination detection. The Decoder says the Financial Times verified the errors; I could not read that report. The Register and Fortune describe the 40 failed citations differently. KPMG has not confirmed AI was used.
KPMG removed a report on AI agents from its websites in June 2026 after a review found that only five of its 45 citations pointed to the source they named, [The Register reported on 12 June 2026](https://www.theregister.com/ai-and-ml/2026/06/12/kpmgs-ai-report-turns-into-a-demo-of-ai-hallucinations/5255029). The review was done by GPTZero, a company that sells AI-detection software. KPMG International published the report, "Total Experience: Redefining Excellence in the Age of Agentic AI", in October 2025, so it was public for about eight months.
What the check found
- 40 of 45 citations (89%) did not match. GPTZero said they ranged from mangled and misleading to partially fabricated or too vague to verify, according to The Register. [Fortune, on 22 September 2026](https://fortune.com/2026/09/22/kpmg-ai-reorg-todd-lohr-steve-case-retirement/), went further and said 40 "were later found to be fabricated".
- Roughly half of the report's factual claims were false, unsupported or attributed to the wrong source, GPTZero alleges.
- Case studies of AI agents at UBS, Swiss Federal Railways and Transport for London cited sources that did not support them, The Register reported. [The Decoder](https://the-decoder.com/kpmg-fabricated-ai-case-studies-in-a-report-designed-to-sell-clients-on-ai-adoption/) adds the UK's National Health Service, and says the Financial Times verified the errors and that every organisation named disputed what the report said about it.
- The report said the airline Emirates had a mobile chatbot called Sara that could change flights. GPTZero says Sara is a robot assistant introduced in 2023 that cannot change bookings.
- It disagreed with KPMG's own data: it said 55% of chief executives rank AI as their top investment priority, while KPMG's 2025 CEO Outlook, published the same month, said 71%.
What KPMG said and what it changed
A KPMG spokesperson told The Register: "The report has been removed and we are reviewing the circumstances surrounding its publication. 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."
Three months later Todd Lohr, who leads the US firm's new Client Technology & Innovation group, told Fortune that KPMG had recently launched an AI tool called Hawk that checks its content for AI hallucinations, meaning statements or references a model has made up, and that it covers most KPMG content. A spokesperson told Fortune that Hawk is mainly used by the marketing department to vet thought leadership pieces. "At the end of the day, it's your output. You need to own the work product," Lohr said.
The cost, the people and the catch
No source I could read says which AI tool produced the report, and KPMG has not said that one did. The Decoder suggests careless use of AI search as the likely cause; KPMG's statement points to its AI guidelines without confirming it. Those guidelines already required a human to verify sources, so either the check did not happen or it did not work, and KPMG has not said which.
KPMG has published no cost, no outcome of its review, nothing on who wrote or approved the report, and nothing on whether clients relied on it. It has given no error rate for Hawk and has not said who signs off after Hawk has run. GPTZero is not a neutral party in one respect: it sells hallucination detection and has $30 million in annual recurring revenue, [Fortune reported on 2 October 2026](https://fortune.com/2026/10/02/hallucination-detection-gptzero-edward-tian-alex-cui/).
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
What to do before your next report goes out
- Marketing: for each report, white paper or case study, name one person who opens every citation and confirms the page says what your text claims. Record their name and the date with the sign-off.
- Marketing: check every claim about a named company against that company's own published page, or ask the company. KPMG's errors were about other organisations, and they disputed them.
- Legal and compliance: if your AI policy says a human reviews AI-assisted content, ask for the evidence of that review on the last three pieces published. KPMG had the policy.
- Run the same check on reports already on your website, starting with the most downloaded. KPMG's was public for about eight months before an outsider checked it.
Take this to work
Forward this
KPMG removed a published report on AI agents in June 2026 after AI-detection firm GPTZero found that only 5 of its 45 citations pointed to the source they named, The Register reported. KPMG already had a policy requiring human checks of sources, so a policy alone is not enough: every report needs a named person who opens each citation before it is published.
Questions to ask
- Who opens each citation in our reports before publication, and where is that recorded?
- Which of our published reports and case studies were drafted or researched with an AI tool, and have they been rechecked since?
- How do we confirm a claim about a named customer or company: from its own page, or from a summary?
- If we buy or build a checker, what share of bad citations does it miss, and who signs off after it runs?
- What is our process for withdrawing and correcting a report, and who tells the clients who downloaded it?
The arithmetic
Share of citations that fail = citations that do not match their named source ÷ citations in the document. Review time = citations in the document × minutes to open and check each one.
KPMG's report: 45 citations, 5 matched, so 40 ÷ 45 = 89% failed on GPTZero's count. No source says how long the check took or what it cost, so the review-time sum needs your own minutes per citation.
Questions this article answers
What was wrong with KPMG's agentic AI report?
GPTZero found that only 5 of the 45 citations in "Total Experience: Redefining Excellence in the Age of Agentic AI" pointed to the source they named, and alleged that roughly half its factual claims were false, unsupported or misattributed, The Register reported on 12 June 2026.
Did KPMG withdraw the report?
Yes. A KPMG spokesperson told The Register in June 2026 that the report had been removed and that the firm was reviewing how it came to be published. No outcome of that review appears in the sources cited here.
Did KPMG use AI to write the report?
KPMG has not said so. Its statement refers to its guidelines on responsible use of AI, and The Decoder suggests careless use of AI search as the likely cause, but no cited source names a tool.
What is KPMG's Hawk tool?
An in-house AI tool that checks KPMG content for AI hallucinations. Fortune reported on 22 September 2026 that it covers most KPMG content and is mainly used by the marketing department to vet thought leadership pieces; no error rate has been published.
Apollo plans to slow hiring growth from 12-15% to 3-5%, citing AI
Apollo's co-president says AI lets it cut staff growth to 3-5% a year, but hiring has not slowed yet. Test your own 2027 plan role by role.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.