Everything About AI
NextFin News - The world's four biggest audit firms are pouring billions into artificial intelligence and declaring that machines are making audits better, faster and cheaper. The evidence is seductive: the Big Four's combined deficiency rate in U.S. regulator inspections fell to 8% in 2025 from 26% three years earlier, even as Deloitte committed more than $3 billion to generative AI, PwC spent nearly $1.5 billion scaling its AI capabilities, EY grew AI-related revenue 30% in a single year, and KPMG pledged $2 billion on AI and cloud with Microsoft. But the same technology promising to fix audit quality is quietly dismantling the apprenticeship model that has produced competent audit partners for a century — and that is a risk no inspection statistic yet captures.
The enthusiasm is not irrational. Audit has always been trapped by sampling: you cannot test every journal entry, so you test a sample and hope it is representative. AI lets firms test entire populations at a fraction of the time. Yet the profession is discovering that the tool that makes audits more thorough may also make auditors harder to replace — and that the two outcomes are not as compatible as the press releases suggest.
The Numbers Behind the Rally
The scale of the bet is visible in the firms' own results. For the fiscal year ended mid-2025, the four firms that audit about 80% of the market capitalization of U.S. public companies reported combined revenue of roughly $220 billion. Deloitte led at $70.5 billion, up 4.8% in local currency. PwC reached $56.9 billion but grew only 2.7%, the slowest of the group. EY posted $53.2 billion, up 4.0%. KPMG, the smallest, hit $39.8 billion, up 5.1%.
Each firm framed that growth as AI-driven. Deloitte's global CEO Joe Ucuzoglu told investors that clients are placing trust in the firm "to navigate an unprecedented level of complexity and change," adding that Deloitte is "uniquely positioned to help businesses, governments, and markets capture the full potential of AI." The spending is concentrated and public: Deloitte has allocated more than $3 billion to generative AI through fiscal 2030; PwC accelerated a nearly $1.5 billion investment to scale AI across its network, including a global "AI factory" and an enterprise command centre it calls agent OS; EY invests more than $1 billion a year in AI-first platforms — 1,000 AI agents and more than 100 internal applications — and says AI-related revenue rose 30% year on year; KPMG is spending $2 billion over five years with Microsoft on AI and cloud, expecting roughly $12 billion of additional revenue.
On the core promise — better audits — the early data is strikingly favourable. The Public Company Accounting Oversight Board's inspection reports for 2025, published in August 2026, showed the Big Four's combined Part I.A deficiency rate at 8%, down from 20% in 2024 and 26% in 2023. Deloitte and EY each recorded 5%, PwC 9%, KPMG 13%. EY called its 5% rate the best in firm history and credited a $1 billion investment in technology and talent. Deloitte described its approach as "human-led, AI-powered."
Yet two shadows sit under the rally. In October 2025, Deloitte Australia agreed to partially refund a A$440,000 contract after an academic found fabricated citations and a made-up court quote in a welfare-compliance report the firm produced with generative AI. Around the same time, PwC's UK chief acknowledged cutting roughly 200 entry-level jobs, citing AI's ability to perform work once given to interns and trainees. The question is not whether AI can draft a memo. It is whether the firms can automate the routine without hollowing out the judgment that routine once built.
The Efficiency Dividend Is Real — And Already Priced In
The first-order case for AI in audit is almost too clean to argue with. Testing 100% of invoices, payments and contract clauses instead of a sample is not an incremental improvement; it is a change in the unit economics of assurance. But this is where the market's enthusiasm runs ahead of the mechanism. Efficiency gains in audit do not automatically become profit or quality. Audit is largely a fixed-fee business: when a procedure that took 100 hours takes 20, the client does not pay five times as much. The savings either flow to the client as lower fees or to the firm as margin — and in a competitive market for mandates, the pressure tilts toward the former.
The firms' own growth numbers show the tension. Despite the AI fanfare, revenue growth across the Big Four was modest, between 2.7% and 5.1%. AI is not yet a growth engine in audit; it is a margin defence. PwC's 2.7% — the weakest of the four — is the clearest signal that AI spending has not yet translated into top-line momentum.
There is also a second-order distributional effect the cheerleading omits. The firms that can afford billion-dollar AI programmes are the same four that already audit about 80% of U.S. listed market capitalization; non-Big Four U.S. firms combined audit less than 2%. AI investment widens that moat. The technology meant to democratize audit capability may instead entrench an oligopoly — a concentration risk that regulators have barely begun to price.
The Apprenticeship Problem No Dashboard Shows
The deeper tension is human capital. Audit quality is not produced only by procedures; it is produced by people who learned to spot anomalies by doing them. For generations, the profession's quality pyramid was built from the bottom up: juniors vouched invoices and traced transactions, and in that repetitive work they internalized what a real exception looks like. AI now does the vouching.
The hiring data shows the model is already bending. Reported UK graduate intake for 2022 to 2024 fell across all four firms — KPMG by 29%, Deloitte by 18%, EY by 11%, PwC by 6%. An internal planning document reported in 2025 projected tax and assurance associate hiring at PwC down 32% between 2025 and 2028, with audit new hires down 39%. PwC's UK chief described a "watch and wait" posture on entry-level hiring while AI reshapes work.
Firms argue this is not a loss but an acceleration: new hires will supervise AI systems and reach manager-level responsibility within three years instead of five. That may be true for the survivors. But it changes the selection and training function of the junior years. Hire one-third fewer graduates and you have one-third fewer data points from which to identify future partners. Train them to review machine output rather than handle raw transactions and they develop a different kind of judgment — one tuned to catching machine error, not to understanding a client's business from the ground up.
This is the cyclical-versus-structural fault line, and it deserves a clean call. The technology shift is structural: 100% testing is not going back to sampling, and the efficiency gain is permanent. But the talent response contains a cyclical correction that is almost certain to come. Firms that cut too deep will discover, with a lag of three to five years, that they lack enough experienced reviewers to sign off on the very AI outputs they deployed. The pipeline squeeze is not a permanent feature; it is a timing mismatch between a fast technology curve and a slow human-capital curve. The firms that over-cut entry-level roles will have to rehire — at a premium — before the decade is out.
There is a compounding effect. Audit is a licensure business: signing an opinion requires credentials and experience that cannot be compressed. If the cohort that would have become 2029's managers is 30% smaller, the 2032 partner class shrinks too. The shortage, when it arrives, will not be at the bottom of the pyramid where AI can substitute. It will be at the top, where the signature still carries personal liability — and where no agent can stand in.
Hallucination Is a New Failure Mode
The Deloitte Australia episode is the cleanest proof that AI introduces a new failure mode rather than simply eliminating old ones. A A$440,000 report for the Australian government, produced with generative AI, contained fabricated citations and a fabricated judicial quote — including a misspelled judge's name — before a University of Sydney academic flagged it. Deloitte refunded part of the fee and reissued the report.
This matters for audit specifically because of what PCAOB Board Member Christina Ho emphasized in September 2025: regardless of how sophisticated the tool, the human auditor remains accountable for the opinion. She argued that regulators should write standards that accelerate responsible adoption rather than wait for failures.
A technology-neutral approach is akin to being an anchor that weighs down innovation. Doing nothing is not a viable option for the PCAOB.
FINRA's 2026 oversight report, published in December 2025, added a standalone generative-AI section naming hallucinations and bias as risks firms must test for and govern. The regulatory message is consistent: the tool can err, but the signature cannot.
The mechanism of the risk is subtle. Traditional audit failure is usually a failure of effort — someone did not look hard enough. AI-era failure can be a failure of skepticism — someone looked at a plausible, confident, wrong answer and accepted it. That is a harder failure to catch, because the output looks more professional, not less. It also inverts the economics of review: checking an AI answer can be as hard as producing it, which means the "efficiency" gain evaporates precisely where the risk is highest.
The Counter-Thesis — Maybe the Enthusiasm Is Justified
The strongest argument against the skeptics is the inspection data itself, and it deserves full weight. The Big Four's combined Part I.A deficiency rate fell from 26% to 8% in three years while AI adoption accelerated. EY's rate dropped from 37% in 2023 to 5% in 2025. These are not small movements, and they are measured by the regulator whose entire mandate is to catch audit failure. If AI were degrading quality, it is odd that measured quality is improving so sharply.
There is a coherent mechanism here too. AI does not replace partner judgment; it replaces the mechanical work that consumed the hours partners would rather have spent on risk assessment. A senior auditor freed from sampling can spend more time on the estimates and judgments where real misstatements hide — fair value, impairments, revenue recognition. In that reading, AI does not hollow out judgment; it upgrades the mix of work toward judgment. The Deloitte Australia failure, on this view, was a consulting report, not an audit, and reflects poor governance on one engagement rather than a systemic flaw in AI-assisted assurance.
This counter-thesis is credible, and it is why the claim here is not "AI will wreck audits." The claim is narrower and, in a way, more uncomfortable: the technology is improving measured audit quality even as it degrades the profession's ability to reproduce itself. Both can be true at once. The inspection numbers validate the tool. The hiring numbers warn about the institution.
What To Watch
The base case is that AI continues to improve measured audit quality while compressing fees and concentrating market share among the firms that can afford the investment. The Big Four's moat widens; smaller firms struggle to keep pace; audit committees get better data but face a narrower choice of auditors.
The upside case is that AI finally delivers the long promise of continuous auditing — real-time assurance instead of an annual rear-view mirror — and that the profession uses the freed capacity to cover risk areas it has historically under-audited. In that world, the apprenticeship problem gets solved by redesigning early careers around AI supervision rather than transaction vouching, and the talent pipeline adapts.
The downside case is the one the inspection data cannot yet see. If firms cut entry-level hiring too aggressively, the reviewer shortage arrives with a lag just as AI-generated work becomes harder to verify. A single high-profile audit failure traceable to AI overreliance — a hallucinated confirmation, a biased model, an unchecked agent — could trigger a regulatory backlash that slows adoption and raises costs for everyone.
The falsifying signal for the "AI improves audit quality" thesis is concrete: if the Big Four's combined Part I.A deficiency rate rises for two consecutive annual PCAOB cycles while AI usage expands — back above 15% — the quality narrative breaks. Watch the August inspection releases. The short-term read is bullish on the firms' margins and their pricing power with clients who want AI-enabled assurance. The medium-term read is mixed: fee compression and a talent squeeze will offset some of the efficiency gain. The long-term read turns on a single institutional question — can the profession reinvent apprenticeship fast enough?
AI is exciting audit firms because it promises to make audits better. It should excite them a little less, because it also threatens to make auditors rarer. The machines are getting better at auditing. The question is whether the profession can keep getting better at being auditors.
Explore more exclusive insights at nextfin.ai.
