Key Takeaways
AI is moving beyond routine audit work, contributing questions, suggested procedures and draft conclusions.
The training challenge extends to future managers: how will they develop the judgement needed to challenge agents?
More sophisticated outputs could encourage greater deference, especially when reviewers perceive the technology as more capable than themselves.
Smaller teams may require more senior oversight, putting pressure on the traditional audit pyramid.
KPMG’s pilot does not establish better audit quality overall. Deeper questioning still needs to produce reliable evidence.
Much of the discussion about AI in auditing has focused on junior work: checking documents, extracting information and carrying out repetitive procedures. The concern has been that firms will automate the tasks through which new recruits learn the profession.
KPMG’s latest experiment pushes that debate further up the hierarchy.
According to the Financial Times, the firm deployed more than 100 AI agents alongside its conventional audit of an unnamed FTSE 250 company. The agents gathered and analysed information, documented their work and produced draft conclusions. They also suggested next steps.
KPMG’s UK head of audit, Cath Burnet, said that in some situations the technology asked better questions than the human team.
If agents can contribute reliably to deciding what deserves investigation and what the evidence means, they are entering work through which auditors develop the expertise needed to manage an engagement.
The profession has been asking how juniors will learn when AI performs their foundational work. It must now ask how auditors will learn to manage an engagement when agents increasingly investigate issues, suggest procedures and draft conclusions.
What KPMG actually tested
KPMG ran two audits alongside each other: its conventional audit and an experimental AI-assisted audit called Project Exponential.
An audit partner signed off the experimental work, but KPMG based its statutory audit opinion on the conventional audit. It did not rely on the AI-assisted exercise when issuing its audit report.
The Financial Times reports that KPMG built the agents in four months. Two partners and a director worked with senior managers and junior staff to build and run them, although the firm declined to disclose the team’s exact size.
Humans set the agents’ objectives, supplied information about the business and previous audits, reviewed their work and challenged their conclusions. In one area, a person intervened after an agent failed to match regulations to company documents.
Burnet nevertheless described instances in which the agents investigated issues more deeply:
“the AI was able to go so much deeper”
Cath Burnet, KPMG’s UK head of audit, describing several instances in the pilot to the Financial Times
Moving from processing information to shaping the audit
Audit work does not divide neatly into routine tasks for juniors and judgement for everyone else. Junior auditors exercise judgement, while senior staff also perform detailed technical work. AI was already contributing to analysis before this experiment.
Extracting information from a contract is one task. Suggesting that its terms require further investigation, identifying the evidence needed and proposing a conclusion take the technology into the reasoning that shapes an engagement.
Those activities touch work performed by experienced team members: directing enquiries, reviewing evidence and deciding when an explanation needs to be challenged.
Firms have often presented the automation of routine work as a way to free people for higher-value activities. If agents also contribute to those activities, the training problem moves with them. Auditors need opportunities to develop the judgement required to direct and review the work, even if they spend less time producing it.
KPMG’s pilot does not establish that AI can replace a manager or an engagement partner. It suggests that the division of work between agents and people may reach further into the hierarchy than a focus on junior jobs implies.
Who learns to challenge the agents?
Auditors prepare for managerial responsibility by planning work, reviewing what others have done, identifying weaknesses and resolving competing explanations. They learn when to pursue a discrepancy, when to seek specialist advice and when the evidence remains insufficient. If firms transfer more of those activities to agents, they will need to preserve opportunities for future managers to practise them.
Reviewing an agent’s conclusion may itself provide valuable experience. AI could expose trainees to more varied cases, make alternative approaches easier to compare and allow experienced auditors to spend more time coaching.
But those benefits depend on how firms organise the work. An employee who repeatedly accepts a plausible answer may gain less experience than one required to formulate an independent view, test the evidence and explain a disagreement.
The task of supervision also changes as the outputs become more sophisticated. Checking whether an agent extracted a figure correctly differs from judging whether its proposed audit approach addresses the relevant risks. A polished analysis can require considerable expertise to challenge.
Training for judgement, rather than tool use
The firms are already trying to address this problem. KPMG US has redesigned its audit internship training around fraud scenarios, incomplete information and critical thinking. PwC is introducing critical-thinking and judgement certification for more than 85,000 professionals, including simulations that require employees to challenge AI-generated conclusions.
These initiatives recognise that operating the technology and evaluating its work require different capabilities. Their effectiveness will depend on whether participants can apply what they learn to unfamiliar problems on live engagements.
In his examination of Arthur Andersen’s training model, William Englehaupt argues that firms must reconsider staffing, promotion and incentives alongside formal instruction. A training programme can encourage independent judgement, but the workplace must also give people the time, responsibility and support to exercise it.
As agents contribute to more demanding work, that challenge extends beyond preparing juniors. Firms will need to develop people who can assess an investigation, identify what it overlooked and challenge the conclusions it produced.
Does greater capability encourage greater deference?
Wayne Banks has warned about automation deference: the weakening of scepticism when reviewers become accustomed to receiving plausible, professionally presented AI output.
Moving AI into higher-level work could intensify that risk in two ways. The firm’s confidence in the technology may encourage employees to regard its conclusions as authoritative. At the same time, increasingly complex outputs may require more expertise and effort to challenge.
An auditor can check an extracted figure against a document. Evaluating an agent’s proposed explanation for an unusual transaction requires a broader understanding of the business, the accounting and the evidence. A convincing explanation may therefore be harder to verify than a routine output, even when both look equally polished.
Claims that agents ask better questions than human teams could reinforce that perceived authority. A less experienced reviewer might reasonably wonder whether their own objection reflects a genuine weakness or something the system understands better than they do.
As AI’s assigned role becomes more senior, firms must ensure that its perceived authority does not substitute for evidence.
Concerns about AI audits
In its 5 August submission to the PCAOB, PwC argued that the use of AI by companies and auditors should become the board’s foremost priority. The firm sought guidance on automation bias, explainability, traceability and the ability to reconstruct AI-generated work.
PwC specifically identified the need to assess the relevance and reliability of audit evidence, maintain professional scepticism even when an AI output appears plausible, and ensure that agentic workflows allow adequate human oversight and intervention.
These are the tests against which KPMG’s pilot needs to be assessed. More extensive analysis matters only if auditors can establish its relevance, evaluate its reliability and explain the conclusions they draw from it.
Pressure on the Big Four pyramid
The traditional audit pyramid combines two functions. It distributes work across different levels of experience, and it provides a route through which recruits become managers, directors and partners.
That structure also supports the economics of the partnership: a relatively small group of senior professionals directs and reviews work carried out by a much larger workforce.
If agents reduce production work, firms could operate with smaller teams. But if reviewing their conclusions requires a greater proportion of experienced staff, those teams may also become more senior. The effect on costs and partner leverage would depend on how much production work disappears and how much review remains.
The commercial risk is that firms reduce their staffing pipeline before they understand the sustained demand for human expertise.
The long-term challenge is how the firms sustain the expertise on which their new model depends. They may need fewer people to produce the work, but they will still need people capable of challenging it.
If agents increasingly perform the work through which auditors acquire that expertise, firms must explain how they will train the managers and partners expected to exercise it.
Related reading
KPMG Asked for the AI Dividend. Now Its Clients Want One Too: How KPMG’s demand for a cheaper audit gave clients an argument for sharing the savings from AI.
KPMG US Redesigns Audit Training as AI Takes Over Routine Testing: Why simulations, mentoring and critical thinking are becoming more important as AI changes the work available to junior auditors.
What Arthur Andersen’s Training Model Can Teach the Big Four About AI: William Englehaupt examines why preserving professional judgement may require changes to staffing, promotion and training.
PwC to Train 85,000 Staff in Critical Thinking as AI Reshapes Work: A programme designed to help professionals evaluate evidence, question assumptions and challenge AI-generated conclusions.
EY US to Keep Assurance Interns for Up to a Year as AI Reshapes Entry-Level Audit: How extended practical experience and coaching could help address the profession’s changing apprenticeship needs.
The Big Four’s AI Reckoning: A Business Model Under Pressure: The wider implications of AI for consulting fees, staffing structures and professional careers.
UK Audit Regulator Warns AI Could Threaten Auditor Independence: Where technology-generated insights can create ethical risks or cross the boundary between audit and advisory work.
AI Governance in Audit: Wayne Banks on What Firms Must Do to Protect Trust: Why apparently authoritative AI output can weaken scepticism, and the controls needed to preserve meaningful human oversight.
This article is part of the Big4News Investigations & Analysis series, which examines the structural forces shaping Deloitte, PwC, EY and KPMG.
About Claudine Cassar

I’m a corporate anthropologist and former Deloitte equity partner. I sold my technology business to Deloitte in 2016 and led the Malta Consulting team for five years. I am the founder and editor of Big4News, which provides independent, clear analysis of PwC, Deloitte, EY, and KPMG — free from corporate spin.
Find me on LinkedIn, X, Instagram, or my author website.
Feel free to reply to this newsletter — I read every reply.



