AI agents ask better questions than KPMG accountants in audit trial
AI audit agents reportedly asked more probing and comprehensive questions than KPMG’s human accountants in internal trials, reigniting debate over how far large language models can outperform professionals in rule-heavy, document-intensive work. Commenters note that much of modern auditing is constrained by misaligned incentives, junior staffing, and “check-the-box” processes, making it a natural target for automation — but warn that AI’s exhaustiveness, opacity, and monoculture risks could create new failure modes and avenues for fraud. Many expect hybrid workflows where AI handles most routine analysis while humans provide judgment, sign-off, and accountability, with significant implications for jobs, regulation, and who ultimately benefits from productivity gains.
Perceived quality of KPMG and Big Four audits
- Many argue the result reflects poorly on KPMG and large audit firms more than it flatters AI.
- Big Four audits are described as often perfunctory “check-the-box” or “whitewashing” exercises, with incentives to keep clients happy rather than probe deeply.
- Several commenters note the heavy use of overworked, inexperienced juniors overseen by a thin layer of seniors; quality for smaller or less strategic clients is seen as especially weak.
- Audits are also framed as something companies buy for signaling and legal protection, not necessarily for discovering every problem.
Where AI appears strong in auditing/accounting
- AI is seen as well-suited to rule-bound, document-heavy, search-space-defined tasks: tax exams, compliance checks, SOC2, FP&A, and policy/code cross-checks.
- Users report AI being more exhaustive and consistent than typical analysts or auditors, especially at remembering details, applying checklists, and surfacing anomalies.
- AI doesn’t tire, can parallelize work across agents, and can in principle move audits from sampling toward near-comprehensive analysis.
Limitations and risks of AI agents
- Commenters highlight hallucinations, mis-mapping of regulations to documents, “trained-in laziness” from token limits, and over-focus on esoteric issues while missing obvious ones.
- Concern that AI buries humans in verbose output, leading to “cognitive surrender” while still making basic errors.
- AI is seen as bad at judgment about materiality, incentives, and what is “worth digging into,” potentially being too rigid or nitpicky.
- Monoculture and exploitability risks: bad actors could pre-test fraud schemes against the auditing AI; targeted exploits seen as more worrying than universal crashes.
Audit practice, incentives, and materiality
- Experienced auditors are said to learn what matters: material vs trivial issues, and how to balance thoroughness, cost, and client relationship.
- Much of real audit work is described as negotiation, context-specific risk assessment, and understanding business realities, not just checklists.
- Legal liability remains with humans (e.g., CEOs signing filings), raising questions about trusting AI-only audits.
Labor, economics, and industry impact
- Many expect AI to displace large portions of junior white-collar work (accounting, consulting, coding), with humans moving toward oversight, training, and edge cases.
- Debate over who captures productivity gains: partners/shareholders vs AI vendors vs clients.
- Some see parallels to Excel/calculators as tools; others stress that unlike those, AI can autonomously run complex workflows.
Skepticism about the trial and reporting
- The trial is self-reported by KPMG; details are vague (“several instances”), with no quantitative comparison of misses vs catches, making the headline feel overstated.
- Some suspect marketing and regulatory signaling rather than a clear, rigorously demonstrated superiority.