KPMG US has redesigned its audit internship programme, placing greater emphasis on professional judgement and critical thinking. This comes as the industry grapples with the challenge of training junior auditors now that artificial intelligence is set to take over repetitive work that once built their experience.
In an April 2026 speech, PCAOB board member George Botic observed that junior auditors traditionally developed judgement and scepticism while performing procedures under supervision. He questioned how auditors who have not performed those basic procedures will acquire the knowledge required to review work completed by AI.
The risk of automation deference
There are serious concerns that junior auditors with limited experience will not have the confidence to challenge AI and maintain professional scepticism when the technology delivers output that appears polished, coherent and technically credible.
Wayne Banks has described this as automation deference: the gradual weakening of scepticism that can occur when AI agents produce plausible and professionally presented output. An inexperienced auditor may accept an incorrect conclusion because it is presented confidently, fail to identify an unsupported assumption or overlook evidence that does not fit the system’s analysis. That would create a direct risk to audit quality.
Professional responsibility nevertheless remains with the human auditor, not the AI system. Firms must therefore put in place controls that require staff to verify sources, challenge assumptions, document their reasoning and escalate results that cannot be independently supported.
Without those safeguards, AI could gradually erode the professional scepticism on which audit quality depends.
What research found
These concerns are supported by a field study run jointly by KPMG and the University of Texas at Austin. The experiment involved 523 US-based early-career professionals working with AI agents on assignments resembling client work.
Researchers first established what an AI agent could produce without human involvement. They then compared that baseline with the work produced when employees used the same technology.
The participants fell into three groups:
AI Amplifiers, representing 50.1%, improved on the AI-only result.
AI Delegators, representing 25.8%, produced results broadly equivalent to the AI working alone.
AI Apprentices, representing 24.1%, performed below the AI baseline.
The Delegators present the clearest example of the deference problem. They scored comparatively poorly on foundational skills but did not produce the weakest final work because the AI was already capable of generating a competent-looking answer. They largely accepted its output with limited scrutiny and added little independent judgement of their own.
That creates a performance-assessment problem for audit firms. A polished deliverable may conceal whether an employee understood the work, tested its assumptions or merely accepted what the system produced.
The study suggests that firms will need to evaluate more than the final answer. They will also need to examine how staff framed the problem, applied professional knowledge, interrogated the evidence and improved—or rejected—the AI’s conclusions.
The Amplifiers did precisely that. They treated AI as a system requiring direction and supervision, grounding its work in relevant professional knowledge and refining the results over several rounds. Their value came not simply from using the technology, but from applying judgement to what it produced.
The Apprentices illustrate a different difficulty. They possessed relatively strong foundational skills and attempted to criticise the AI’s work, but their interventions sometimes focused on irrelevant issues or steered the system in an unhelpful direction.
Teaching scepticism therefore requires more than encouraging junior auditors to question AI. They must also learn which questions matter and how to determine whether the evidence supports the conclusion.
The KPMG Auditor Training Programme
KPMG’s revised programme attempts to provide alternative opportunities for early-career staff to practise scepticism, investigate anomalies and decide when an AI-generated conclusion requires further examination.
Nearly 1,000 audit and assurance interns attended an intensive programme at KPMG Lakehouse, the firm’s US training centre in Florida. The exercises included fraud-detection scenarios, problem-solving challenges and activities designed to test how participants respond when information is incomplete or potentially misleading.
The programme focuses on six areas: technical knowledge, critical thinking, agility, business acumen, interpersonal skills and professional behaviour. KPMG said these capabilities would become more important as auditors move away from repetitive testing and spend more time evaluating information generated by automated systems.
Can simulations replace experience?
KPMG’s programme is an attempt to teach deliberately what earlier generations of auditors acquired gradually through repeated work on live engagements.
Fraud scenarios and structured exercises can expose junior staff to ambiguity, require them to identify contradictory information and make them explain their reasoning. They cannot fully reproduce incomplete client records, difficult management teams, deadline pressure or the consequences of reaching the wrong conclusion during an actual audit.
KPMG’s own survey suggests that interns recognise the risk. Of 361 US winter interns surveyed, 43% identified over-reliance on AI limiting critical thinking and decision-making as their principal concern about the technology. They ranked hands-on project work and in-person mentoring as the strongest contributors to their readiness for full-time employment.
This helps explain why KPMG is combining AI-enabled work with simulations, mentoring and an immersive in-person programme rather than treating technological fluency as sufficient preparation.
The unresolved question is whether junior auditors can become effective reviewers of AI-generated work without first spending enough time performing the underlying procedures themselves.
EY US has since begun tackling the same problem from a different direction. Its new Career Residency will allow selected assurance interns to remain with the firm for another eight to 12 months while completing university, with explicit emphasis on professional judgement, critical thinking and scepticism.
KPMG’s experiment will therefore be watched beyond the firm. As AI removes more entry-level work, the profession will need to demonstrate that it has found another way to develop the technical understanding, judgement and professional scepticism on which audit quality depends.
Related Coverage
The Wall Street Journal: Inside KPMG’s AI-era audit training programme
Reports from KPMG Lakehouse and describes the exercises used to develop judgement, scepticism and fraud-detection skills.
KPMG US: Shaping Early-Career Success in the Age of AI
Explains the research underpinning the programme and the firm’s framework for evaluating human performance alongside AI agents.
Business Insider: KPMG revamps internships as AI changes junior work
Provides earlier reporting on the decision to reduce the emphasis on conventional technical instruction during the central internship programme.
KPMG Intern Pulse Survey
Sets out KPMG’s findings on how Generation Z interns expect AI and automation to affect their future roles.
Follow Big4News coverage of how AI is changing auditor skills and development → Audit Training & Talent
Want to stay up to date on all things Big Four around the world?
Check out the News section and subscribe to Big4News for weekly deep dives and briefings.
About the author: Claudine Cassar is the founder and editor of Big4News, covering audit, consulting, regulation and governance across Deloitte, PwC, EY and KPMG.
The ability and authority to challenge GenAI output is the crux of the systemic flaw I have been researching. This is different to automation bias; there a person is over-relying on the output because other pressures prevent proper analysis. You are describing exactly what I have called Judgement Displacement.
I like the idea of role-playing for juniors to learn what they need. (You could have GenAI design the training to incorporate the most important, useful, and common tasks. :)
A large multi-national corporate I visited recently have begun training their interns as leaders of virtual teams: several GenAI members are given complementary roles — fundamentally checking each other's work — while the human is responsible for the oversight and final product.