KPMG Retracts ‘Agentic AI’ Report After Hallucinations Create Fake Corporate Case Studies
Table of Contents
The Irony of the ‘Agentic AI’ Failure
In a twist of situational irony that has sent ripples through the professional services sector, KPMG has officially retracted its high-profile research paper, “Redefining excellence in the age of agentic AI.” The report, intended to showcase the cutting edge of autonomous AI agents in corporate environments, was pulled from the firm’s digital platforms after several of the organizations cited as success stories claimed they had no idea the report existed, let alone that they were employing the specific AI strategies described.
- The Incident: KPMG retracted a major report on agentic AI after organizations including UBS, the NHS, and Transport for London denied the report’s claims.
- The Cause: Technical analysis by GPTZero suggests the report suffered from AI hallucinations, likely because LLMs were used to draft the research without sufficient human verification.
- Industry Pattern: This follows a similar incident at EY, where a report on loyalty rewards was withdrawn due to fabricated footnotes.
- The Risk: The event highlights a critical failure in ‘Human-in-the-Loop’ (HITL) verification processes within Big Four accounting firms.
The fallout began when GPTZero, a leading AI-detection and research group, identified glaring inaccuracies within the October 2025 publication. According to reports from the Financial Times, the inaccuracies were not simple typos or outdated statistics, but fundamental fabrications of corporate behavior. Essentially, the AI used to write the report about AI invented a reality where global institutions were adopting agentic workflows more aggressively than they actually were.
Who Was Impacted by the Fabrications?
The scope of the hallucinations was extensive, touching several high-profile public and private entities. When approached by the Financial Times, UBS, the UK’s National Health Service (NHS), Swiss Federal Railways, and Transport for London (TfL) all confirmed that the claims attributed to them were either entirely untrue or significantly misleading.
For these organizations, the error is more than an embarrassment; it is a matter of corporate governance. For a public entity like the NHS or TfL, being cited as a pioneer in a specific AI implementation can trigger regulatory scrutiny or public expectation of services that do not yet exist. For a financial giant like UBS, accuracy in reporting AI adoption is tied to risk management and compliance standards.
Understanding the Technical Glitch: What are AI Hallucinations?
To understand how a firm as prestigious as KPMG could publish fabricated data, one must understand the nature of Large Language Models (LLMs). AI hallucinations occur when a model generates text that is grammatically correct and stylistically confident but factually incorrect. This happens because LLMs are probabilistic, not deterministic; they predict the next most likely token rather than querying a database of verified facts.
In the case of the “Agentic AI” report, the model likely suffered from confabulation. When asked to provide examples of how “agentic AI”—AI that can plan, use tools, and execute multi-step goals autonomously—is being used, the model may have synthesized general industry trends and applied them to well-known brands (like UBS or the NHS) to create a plausible-sounding but fake case study.
Agentic AI vs. Generative AI
The report specifically focused on Agentic AI. Unlike standard Generative AI (which responds to a prompt), Agentic AI is designed to act as an agent. It can break a complex goal into sub-tasks, use external APIs, and self-correct. By claiming these agents were already operational in the NHS or Swiss Federal Railways, the report was attempting to signal a shift from “AI as a chatbot” to “AI as a colleague.” The failure to verify these claims suggests a reliance on the AI’s perceived authority over actual primary source research.
A Pattern of Professional Services Failures
KPMG is not alone in this struggle. The incident echoes a recent failure at EY (Ernst & Young), which was forced to withdraw a report on loyalty rewards programs. In that instance, the document contained fake footnotes—a classic hallmark of LLM hallucinations where the AI creates a citation that looks legitimate (correct journal name, plausible title) but does not actually exist in any library.
This trend suggests a systemic issue within the “Big Four” and other global consultancies. These firms are under immense pressure to lead the AI narrative, often prioritizing speed-to-market for “thought leadership” over rigorous traditional auditing. When the goal is to be the first to define “Agentic AI,” the temptation to use the very technology being analyzed to write the report is high.
What This Means for the AI Industry
The KPMG retraction serves as a cautionary tale for the broader enterprise AI rollout. It exposes three critical vulnerabilities in the current corporate AI strategy:
1. The Illusion of Competence
LLMs are designed to be persuasive, not truthful. The professional, authoritative tone of a model like GPT-4 or Claude can lull human editors into a false sense of security. When a report looks and sounds like a KPMG whitepaper, the human reviewer may skip the tedious process of emailing the NHS to confirm a specific project’s status, assuming the AI “did the research.”
2. The Failure of ‘Human-in-the-Loop’ (HITL)
KPMG’s spokesperson stated, “We expect all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content.” The fact that the report was published proves that the “human” in the loop was either absent or performing a surface-level review. For HITL to work, the human must act as an adversarial editor, not a passive proofreader.
3. The Risk to Brand Equity
For a firm whose entire product is Trust, publishing hallucinations is a brand disaster. Audit firms sell accuracy. If a firm cannot be trusted to verify the case studies in its own marketing materials, clients may question the rigor of their financial audits or strategic consulting.
How to Prevent Hallucinations in Corporate Research
To avoid the “KPMG trap,” organizations must move away from simple prompt-and-publish workflows. Experts suggest the following framework for AI-assisted research:
- RAG (Retrieval-Augmented Generation): Instead of letting the AI rely on its internal weights, firms should use RAG to force the model to cite specific, uploaded documents. If the information isn’t in the provided source, the AI should be instructed to say “I don’t know.”
- Cross-Verification Protocols: Every corporate claim must be verified by a primary source. If the AI claims UBS is using a specific agent, a human must obtain a confirmation email or a public press release from UBS.
- Temperature Control: Reducing the “temperature” of an LLM makes its output more deterministic and less creative, which is essential for factual reporting.
- AI Detection as a Safety Net: Using tools like GPTZero to scan final drafts for synthetic patterns can alert editors to sections that may have been overly “dreamed up” by the model.
FAQ: AI Hallucinations and Corporate Reporting
Why did KPMG use AI to write a report about AI?
Many firms use AI to synthesize large amounts of data, structure outlines, or draft initial versions of reports to increase efficiency. However, when used for factual reporting without rigorous verification, it leads to the types of errors seen in the “Agentic AI” report.
What exactly is ‘Agentic AI’?
Agentic AI refers to AI systems that can operate with a degree of autonomy. Unlike a chatbot that just answers questions, an AI agent can take actions—such as booking a flight, updating a CRM, or conducting a multi-step market analysis—without needing a human to prompt every single step.
Is GPTZero an official auditor?
GPTZero is a leading AI detection company that specializes in identifying AI-generated text. While not a formal auditor of professional services firms, their technical analysis provided the evidence that the KPMG report contained synthetic inaccuracies.
Does this mean I can’t trust AI-generated business reports?
You can trust them if they are transparent about their methodology and provide verifiable primary sources (links to official PDFs, press releases, or interviews). If a report makes bold claims without direct citations, it should be treated with skepticism.
How does this compare to the EY report failure?
Both the KPMG and EY incidents involved “synthetic” data—where the AI fabricated facts or citations to fit a narrative. This suggests that the pressure to produce “AI thought leadership” is currently outpacing the implementation of safety protocols in the consulting industry.
The Verdict on Responsible AI
The retraction of the “Redefining excellence in the age of agentic AI” report is a humbling moment for the professional services industry. It proves that no matter how advanced the tool, the fundamental requirement of journalism and auditing remains the same: verify everything.
As we move further into 2026, the distinction between “AI-assisted” and “AI-generated” content will become the primary benchmark for credibility. Firms that prioritize transparency and rigorous human auditing will maintain their authority, while those that treat LLMs as infallible research assistants risk their reputations on a hallucination.
