When the consultants cite ghosts: KPMG's hallucinated report and what it tells us about the AI-industrial complex

On 14 June 2026, The Indian Express reported that KPMG had formally retracted an agentic-AI study after outside researchers identified hallucinated citations and fabricated references inside it. The study was pitched as evidence that AI agents were already reshaping white-collar work; on inspection, a non-trivial share of the named papers, authors and journals did not exist. KPMG acknowledged the failure and pulled the report. The substance of the original argument — that autonomous AI systems are measurably altering corporate workflows — is a live and serious question. The methodology used to advance it was not.

This is the second time in twelve months that a major professional-services firm has had to walk back a flagship AI publication after the citations failed basic verification. The pattern matters more than any single retraction. A consulting industry that sells rigour to boards, regulators and the courts is now demonstrating, in public, that its own internal quality control cannot keep up with the tools it is recommending to clients. The question is no longer whether AI hallucinates. It is whether the institutions that monetise AI advice have any independent capacity to notice when the machines they champion are making things up.

The mechanics of a quiet retraction

Agentic AI — systems designed to plan, browse and act on a user's behalf with limited supervision — has become the most fashionable line item in enterprise software roadmaps. The promise is straightforward: replace repetitive knowledge work with software that closes its own browser tabs. Vendors and their consulting partners have been racing to produce quantitative evidence that the shift is already underway. KPMG's study was part of that wave, positioned as a survey of deployments across industries.

According to The Indian Express, the trouble surfaced when independent researchers attempted to follow the study's footnotes. Several led to journal articles that did not exist; others pointed to authors with no publication record in the named field. The Indian Express reported the retraction on 14 June 2026 at 07:52 UTC, citing the firm's own acknowledgement. KPMG has not, as of that report, publicly identified which external researchers raised the flag or which internal review process had originally signed the work off.

The structural problem is that the marginal cost of producing AI-assisted text has collapsed, and the marginal cost of verifying it has not. A junior associate with access to a competent model can generate a thirty-page report in an afternoon. The same afternoon used to produce, at most, a five-page memo. Verification, however, still requires reading the cited material, contacting the cited authors, or at minimum running the references through a working database — none of which the model itself can do reliably when it is the source of the citations in the first place.

Why this is not just a KPMG problem

The temptation is to treat this as an embarrassing episode for one firm and move on. That misses the point. KPMG is one arm of a four-firm oligopoly that audits the books of nearly every listed company in the major markets and consults to a large share of the rest. The credibility of capital markets rests, in practice, on the assumption that these firms can read a number and tell the difference between it and a hallucination. The agentic-AI study is the first widely-reported instance in which a Big Four firm has been caught publishing hallucinated material under its own brand. It will not be the last if the underlying economics do not change.

The deeper issue is epistemic capture. Consulting firms are simultaneously the producers, the certifiers and the consumers of corporate AI advice. They sell the strategy. They sell the implementation. They increasingly sell the audit of the implementation. When the same firm writes the report, grades the report and bills the client for the report, the corrective mechanism that is supposed to catch a hallucination — somebody else's eyes — is structurally absent. The Indian Express reporting indicates that the flag came from outside the firm entirely. The internal review did not catch it.

What the agents will not tell you

There is a related temptation to treat the incident as evidence that agentic AI itself is overhyped, and to use the retraction as a stick to beat the vendors. That is a category error. The retraction tells us something about the evidence base for the claims being made on behalf of agentic AI. It does not tell us whether the underlying technology works in production, where it is being measured against operational metrics rather than literature reviews. The two questions — does the technology perform, and is the marketing of the technology honest — are separable, and conflating them suits everyone except the buyers.

This is also where the Global South angle bites. India's outsourcing and IT-services industry is one of the world's largest markets for agentic-AI deployment, and Indian firms have been early adopters of these systems for client-facing work in finance, customer support and software testing. The Indian Express's prompt coverage of the KPMG retraction is itself a signal that domestic commentators are alert to the credibility risk. If a marquee study underpinning the agentic thesis turns out to be partly fictional, the consequences for firms in Bangalore, Hyderabad and Pune that have already re-priced their labour around that thesis are non-trivial. Boards that have approved headcount cuts in anticipation of agentic productivity gains may want to see the underlying evidence — and, for the first time, may find that there is less of it than the slide deck suggested.

The stakes, and the limits of what is known

The cleanest reading of the episode is also the most uncomfortable: the consulting industry is now operating a two-tier system of evidence. Externally, it demands audited numbers from its clients. Internally, it publishes numbers that have not been audited by anyone, in studies that have not been peer-reviewed, in fields where the supply of genuine domain expertise is thin. The KPMG retraction is the first time a major firm has had to admit, in public, that its own product failed the standard it imposes on others. It will not be the last, because the economics that produced it have not changed.

What remains genuinely uncertain is whether the episode marks a turning point. The Indian Express's reporting names KPMG and the fact of retraction; it does not, in the material available, identify the individual authors, the precise number of fabricated citations, or whether the firm plans to reissue the study. It also does not name the external researchers who caught the errors, which makes independent corroboration harder. A fuller accounting — the kind a parliamentary committee or a securities regulator might demand — would name the human authors, list the failed references, and describe the internal sign-off chain. Until then, the right working assumption is that the public version of events is the optimistic one.

The bigger story is the one this publication finds most worth telling: the institutions that sell certainty to the rest of the economy are themselves, in the new world of generative tooling, increasingly uncertain. The KPMG report was not wrong about a market. It was wrong about a fact. That is a more dangerous class of error, because it travels further before it is caught.

Desk note: The Indian Express wire led this story; Monexus treated it as the trigger and resisted the urge to amplify it into a sweeping verdict on the entire agentic-AI category. The line we held is the line the evidence supports — that one firm's evidence base failed, and that the wider profession's incentives are not yet aligned to catch the next one.