AI lets auditors scan full ledgers—but cannot absorb their accountability

In an analysis column, the Financial Times’ John Gapper describes how the Big Four are embedding AI agents in audits: filtering transactions for fraud and errors, checking junior auditors’ work and looking across millions of entries. KPMG’s UK audit technology chief compares the change to filtering an entire river instead of dipping in a bucket.

That could improve audit quality because traditional sampling necessarily misses transactions. A model can flag unusual payments or accounting patterns across the whole ledger, leaving humans to investigate exceptions. The benefit is not that AI “does the audit,” but that it changes where scarce professional attention is applied.

The risks arrive at the same point. The UK Financial Reporting Council says audit partners remain responsible for conclusions and cannot blame an AI system for mistakes. Firms need to explain why evidence supports a judgment even when an opaque model produced the alert. Junior staff may also lose the repetitive work through which they learn how accounts fail, creating a deskilling problem. Commercializing insights found by audit models is constrained by independence rules governing non-audit services.

AI can remove the statistical limitation that forced auditors to inspect samples, but it cannot remove the evidentiary and legal burden of deciding what an anomaly means. The real quality test is whether firms measure better detection and preserve skeptical human judgment—not how many transactions a model touches.