On 6 October 2025, Deloitte and Anthropic announced the largest Claude enterprise deployment to date, putting the model in front of about 470,000 Deloitte staff. The same day, news broke that Deloitte would refund the final instalment of its fee on a roughly AU$440,000 government report, after a University of Sydney researcher found fabricated references and an invented quote from a Federal Court judgment inside it.

One firm, one day, moving in both directions at once. It was buying AI at the scale of half a million seats and handing money back because AI had written something no one checked. That is the whole question about the Big Four in a single news cycle, and the honest answer is not the one the headlines reached for.

I have watched these firms from the buying side. For eleven years I ran distressed corporate debt inside a major bank, where I helped build its in-house forensic team. So I read those two announcements the way a client does. Neither one says the firm is finished. Together they say which part of it is.

AI will not remove the big consulting firms from the market. It is removing the structure that built them: the wide base of junior analysts whose billable hours funded the climb to partner. The clients stay. The revenue records stay. What goes is the pyramid, the machine that has manufactured every partner these firms have.

What actually failed was verification

This audience already knows the anecdote. The part that matters sits underneath it. The 237-page Australia report, published on 4 July 2025 for the Department of Employment and Workplace Relations, cited academic papers that do not exist and quoted a Federal Court judgment that was never written. Deloitte reissued it on 3 October 2025 with an appendix disclosing that it had used Azure OpenAI, and repaid the last part of the fee.

Weeks later the pattern repeated. A $1.6 million health-workforce plan for the Canadian province of Newfoundland and Labrador, 526 pages, cited at least four papers that do not exist and attributed studies to researchers who never wrote them. Deloitte said it “firmly stands behind the recommendations” and that AI had produced only a limited number of the citations.

The public incidents so far cluster on Deloitte. No comparable scandal has surfaced at EY or KPMG, and it would be dishonest to imply one has. But the failure is not really Deloitte’s. It is what unsupervised model output does inside any professional document, and the legal profession is the leading indicator. Damien Charlotin’s public database of AI-fabricated citations in court filings grew from about 200 cases to about 1,598 by June 2026, and the penalties have climbed from a single $5,000 fine in 2023 to six-figure sanctions. What a client pays a Big Four firm for is the confidence that a report has already been checked. In both Deloitte cases, no one had checked it.

Clients buy expertise and cover

Demand did not move. Deloitte closed the same fiscal year with record global revenue of $70.5 billion, the first professional-services firm past that mark. The scandal and the record sit in the same twelve months, and there is no contradiction between them. Clients buy two things from these firms, and AI touches neither. The first is decades of accumulated expertise. The second is cover: a decision backed by a report under a prestigious logo can be defended later, in front of a board or a court, in a way the same decision reached in-house cannot. The industry says it without embarrassment. Nobody ever got fired for hiring McKinsey.

Which is why the cash at stake is almost beside the point. The Australia refund came to AU$97,587, under a quarter of the contract. The incident cost Deloitte very little. What it puts at risk is the asset the fee actually buys: the client’s freedom not to check the work.

Every firm has picked its lab

The firms are not waiting for this to happen to them. Each of the majors has committed to a frontier lab, and several to two.

FirmAI lab or platformAnnounced
DeloitteAnthropic Claude, rollout to about 470,000 employeesOctober 2025
PwCOpenAI, first ChatGPT Enterprise reseller, 100,000+ seats, plus an expanded Anthropic allianceMay 2024 and May 2026
EYMicrosoft Azure OpenAI, with NVIDIASeptember 2023 and March 2025
KPMGMicrosoft, $2 billion over five years, with Google CloudJuly 2023 and April 2025
McKinsey, BCG, Accenture, CapgeminiOpenAI Frontier AlliancesFebruary 2026

Whether those alliances are a defence or a slow handover of the client relationship to the labs is the subject of a separate article in this series.

The cuts land on the base

The spending coincides with a contraction, and the contraction is concentrated exactly where AI is strongest: the entry level. UK Big Four graduate intakes fell 6% to 29% from 2023 to 2025, the deepest cut at KPMG, which the firms attribute to AI taking over routine junior work and to offshoring. In the United States, an internal PwC plan reported by Business Insider projects 32% to 39% fewer entry-level audit hires by 2028, and PwC’s AI assurance leader, Jenn Kosar, told the outlet that new hires would be doing “manager-level work within three years.”

Read that as a concession. The training that once took years of routine files now has to happen without them. A partner is the survivor of a decade of selection from a wide junior base, and thinning the base thins the selection. It works like football: the fewer children play the game, the fewer world-class players a country eventually produces, and the lag runs long enough that nobody notices the shortage until it is too late to fix.

Cut entry intake by a quarter now and the partner class of 2040 is short by at least a quarter. The profession has started to call this the apprenticeship problem. Kosar’s three-year track is the first live attempt to rebuild the funnel, and nobody yet knows whether it produces partners.

The spreadsheet is the precedent, and its limit

The historical answer to the panic deserves stating at full strength. When the spreadsheet automated the previous generation of routine financial work, the United States lost about 400,000 bookkeeping and accounting-clerk jobs and added about 600,000 higher-order accounting jobs, by NPR Planet Money’s count. The routine layer genuinely disappeared and the profession grew by doing harder work. The clerks were never rehired, and nobody mourns the job.

The precedent has one limit, and the limit is the whole argument. The spreadsheet left the thinking to the junior sitting on top of it. The reasoning stayed human, and the tool only did the arithmetic. A language model drafts the reasoning itself, quickly and with total confidence, which is why the training problem is new. The junior who once learned by producing the first pass now competes with a machine that produces it, which is what the two Deloitte reports actually show: a first pass with no experienced hand left to check it.

The model cannot replace the human yet

I can say what this looks like from inside a practice that runs on the tools. At Interiu we use AI every day in financial analysis, due diligence and OSINT. This is first-hand experience rather than a study, so I will not dress it up as a statistic. The model cannot replace the human yet. Left unsupervised, it fails in the exact way the two Deloitte reports failed: polished prose sourced to material that is not there. An experienced reviewer catches it in minutes. A junior who has never done the work without the machine may not catch it at all. So the work does not disappear. It changes shape. Fewer people do routine collection and reconciliation. The ones who remain supervise the model and verify its output, and that verification becomes a deliverable a client will pay for on its own.

The strongest objection is that the firms will not be AI’s victims at all. They will capture the AI advisory market themselves, and disruption will be their largest sale in a decade. The numbers back it. AI-related work already reached about a quarter of BCG’s 2025 revenue. Accenture’s generative-AI bookings roughly doubled, from $3.0 billion in fiscal 2024 to $5.9 billion in fiscal 2025. EY says its AI-related consulting grew about 30% in a year while total revenue grew about 4%. A new billable line, AI assurance, auditing other companies’ AI systems, is being built to sell.

All of it is true, and none of it saves the pyramid. Capturing AI revenue and preserving the model are different claims. The new revenue is implementation work, shaped like Accenture’s, billed by outcome and senior expertise rather than by the hour on a wide junior bench. It replaces the analyst-hour arbitrage the pyramid ran on rather than restoring it. Consulting already survived the move to cloud by reselling it, and the pyramid came through intact because juniors still did the reasoning on top of the new systems. This wave differs in one way that matters: the tool does the first draft of the reasoning. A firm can win the whole AI market and still be unable to fill the bench the old way, because the work the bench used to do is the work being automated.

What to demand before you sign

This has an edge you can use now, before AI settles anything. If you are commissioning a report, a private-equity partner buying diligence or a board buying a strategy review, run four questions over the engagement before you sign it. They return one of two readings: a firm selling named senior judgment it will stand behind, or a junior bench billed at partner rates with unverified model output underneath. A vague answer is itself the finding.

  1. Named senior time. Ask the engagement letter to name who does the work and at what seniority, rather than quote a blended rate over an anonymous team. It passes if you can point to specific senior hours. It fails if the answer is a bench you cannot see, billed near the top rate. This is the layer the firms are cutting, so the question is fair.
  2. AI use disclosed in writing. Ask where AI was used to produce the deliverable, and by whom. In the Australia case the disclosure appeared only after the fabrications were caught, and governments now want disclosure written into the contract. It passes with a clause before signature. It fails on silence or “we follow our internal policy.”
  3. Sources traced by a named human. Ask that every citation, figure and quote be checked against its source by a person who signs for it. Both Deloitte reports failed on exactly this, and the court filings show it is not rare. It passes if a named reviewer is accountable for the sources. It fails on “our quality process covers that.”
  4. Liability you can read. Ask what the firm will actually defend and what the liability cap in the engagement letter really is. Deloitte “firmly stands behind the recommendations” of a report with invented citations, so find out what standing behind it is worth before you rely on it. It passes with a liability position in writing. It fails with a logo and a handshake.

The firms worth their fee answer all four without flinching. The ones selling the old pyramid at the old price give you a brochure sentence, and that is your answer.

Go back to 6 October 2025. In one day Deloitte put Claude in front of 470,000 people and paid a government back for a report a rival model had written and no one had read. Both moves are rational, and both point the same way. The firm is spending heavily to survive the technology, and the part of itself it is spending to replace is the pyramid that made it. The firms will still be here in 2040. The base that staffed them will not. If you are working out what your own practice looks like after the pyramid, that is the work we do: what we build for advisory services.

The short version

  • AI will not remove the big consulting firms from the market. It removes the base of the pyramid, the junior-analyst layer whose billable hours funded the path to partner.
  • On 6 October 2025 Deloitte announced the rollout of Claude to about 470,000 staff and, the same day, agreed to refund part of its fee on a roughly AU$440,000 Australian government report that contained fabricated references (CNBC, Fortune).
  • Public AI-slop incidents in Big Four reports so far cluster on Deloitte, which had two in 2025. The mechanism is industry-wide: court filings with AI-fabricated citations rose from about 200 to about 1,598 in a year (Charlotin database via HAQQ, 2026).
  • UK Big Four graduate intakes fell 6% to 29% from 2023 to 2025 and an internal PwC plan projects 32% to 39% fewer entry-level audit hires by 2028, while Deloitte posted record FY2025 revenue of $70.5 billion (City AM on FT figures, Business Insider, Deloitte).
  • Before commissioning a consulting report now, a buyer can demand named senior time, AI use disclosed in writing, every source traced by an accountable human, and a readable liability position. A vague answer to any of the four is itself the finding.

Sources

  1. CNBC, 6 October 2025, Deloitte and Anthropic announce the largest Claude enterprise deployment: https://www.cnbc.com/2025/10/06/anthropic-deloitte-enterprise-ai.html
  2. TechCrunch, 6 October 2025, the deployment and the refund on the same day: https://techcrunch.com/2025/10/06/deloitte-goes-all-in-on-ai-despite-having-to-issue-a-hefty-refund-for-use-of-ai
  3. Fortune, 7 October 2025, the Deloitte Australia report and partial refund: https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund
  4. Australian Greens media release, October 2025, the AU$97,587 refund figure: https://greens.org.au/news/media-release/greens-slam-deloittes-unethical-behaviour-and-measly-refund-over-ai-report
  5. CBC News, November 2025, fabricated citations in the Newfoundland and Labrador health-workforce plan: https://www.cbc.ca/news/canada/newfoundland-labrador/nl-deloitte-citations-9.6990216
  6. HAQQ analysis of the AI Hallucination Cases database (Damien Charlotin), June 2026: https://www.haqq.ai/blog/ai-legal-hallucination-audit
  7. NPR, 3 April 2026, penalties for AI-fabricated citations in the legal system: https://www.npr.org/2026/04/03/nx-s1-5761454/penalties-stack-up-ai-spreads-through-legal-system
  8. Deloitte Global press release, October 2025, FY2025 revenue of $70.5 billion: https://www.deloitte.com/global/en/about/press-room/global-revenue-announcement.html
  9. City AM, on FT-published firm figures, UK Big Four graduate intake cuts 2023 to 2025: https://www.cityam.com/big-four-slash-graduate-jobs-as-ai-takes-on-entry-level-work/
  10. Business Insider via Yahoo Finance, August 2025, PwC US entry-level hiring plan and Jenn Kosar quotes: https://finance.yahoo.com/news/pwc-cutting-number-grads-hires-192755746.html
  11. Accounting Today, on AI, training and the accounting talent pipeline: https://www.accountingtoday.com/opinion/how-ai-will-reshape-talent-training-and-pipelines-in-accounting
  12. NPR Planet Money, Episode 606 “Spreadsheets!”, February 2015: https://www.npr.org/transcripts/389027988
  13. Yahoo Finance, BCG reports about 25 percent of 2025 revenue from AI-related work: https://finance.yahoo.com/sectors/technology/articles/bcg-reports-25-revenue-ai-133415166.html
  14. Accenture, fourth-quarter and full-year fiscal 2025 results, generative-AI bookings: https://newsroom.accenture.com/content/4q-full-fy25-earnings/accenture-reports-fourth-quarter-and-full-year-fiscal-2025-results.pdf
  15. Business Insider via AOL, 2025, EY AI-related consulting growth: https://www.aol.com/articles/big-four-giant-ey-ai-113040002.html
  16. consultancy.uk, the Big Four weighing new AI-assurance offerings: https://www.consultancy.uk/news/40424/big-four-weighing-up-new-ai-audit-offerings
  17. Accounting Times, on the Deloitte case and calls for AI disclosure in government contracts: https://www.accountingtimes.com.au/technology/deloitte-ai-debacle-underscores-need-for-government-contract-transparency-ocp-says
  18. CNBC, 29 May 2024, PwC becomes OpenAI’s first reseller and largest enterprise customer: https://www.cnbc.com/2024/05/29/pwc-to-become-openais-first-reseller-and-largest-enterprise-user.html
  19. PwC press release, May 2026, expanded alliance with Anthropic: https://www.pwc.com/us/en/about-us/newsroom/press-releases/anthropic-pwc-expand-alliance-agentic-enterprise.html
  20. EY press release, September 2023, EY.ai launch on Microsoft Azure OpenAI: https://www.ey.com/en_gl/newsroom/2023/09/ey-announces-launch-of-artificial-intelligence-platform-ey-ai-following-us-1-4b-investment
  21. EY press release, 18 March 2025, EY.ai Agentic Platform with NVIDIA: https://www.ey.com/en_us/newsroom/2025/03/ey-launching-ey-ai-agentic-platform-created-with-nvidia-ai-to-drive-multi-sector-transformation-starting-with-tax-risk-and-finance-domains
  22. Axios, 11 July 2023, KPMG’s five-year Microsoft cloud and AI commitment: https://www.axios.com/2023/07/11/kpmg-microsoft-ai
  23. Google Cloud press corner, 9 April 2025, KPMG expands its Google Cloud AI alliance: https://www.googlecloudpresscorner.com/2025-04-09-KPMG-Expands-AI-Alliance-with-Google-Cloud-to-Support-Industry-Adoption-of-Agentspace-and-Deliver-Multi-Agent-AI-Solutions
  24. Fortune, 23 February 2026, OpenAI’s Frontier Alliances with McKinsey, BCG, Accenture and Capgemini: https://fortune.com/2026/02/23/openai-partners-with-mckinsey-bcg-accenture-and-capgemini-to-push-its-frontier-ai-agent-platform/