Auditors Now Use AI on Your Books. A £6 Million Fine Shows the Audit Trail Still Decides.

Audit season planning is under way at most calendar-year companies, and this year the auditor across the table is likely to arrive with new tools. The large firms are now running generative AI over the documents finance teams supply, looking for gaps and summarizing what the audit committee has already been told. It is tempting to read that as good news for the close: faster fieldwork, fewer requests, a lighter January. Two stories from this week suggest a different lesson. AI changes how evidence gets reviewed. It does not change whether the evidence is there, or whether a questionable balance gets challenged.

What AI Is Actually Doing Inside the Audit

PwC has been the most open about it. In an interview with CFO Dive published October 7, Shawn Panson, PwC's U.S. assurance and transformation leader, said the firm's auditors began using its enterprise generative AI tools about a year ago. The tools flag missing information in supporting documents, summarize audit committee communications so issues surface earlier, and help auditors document a client's business processes before walkthroughs. The final audit report looks the same, Panson said, because the firm is still bound by the same regulation and professional standards.

The time savings are real but small. Asked whether three months of work had been compressed into one, Panson described the change as "more of a shift and less dramatic," and said larger savings should come in later years as the technology and adoption mature. For a controller, the practical implication is narrower than the headlines suggest. A tool that flags missing support will flag it faster and more consistently than a first-year associate working down a request list. Incomplete reconciliations, undocumented journal entries and accruals with no clear basis are exactly what it is built to find.

The £6 Million Reminder That Skepticism Is the Job

The other story is a reminder of what tools do not supply. On October 8 the Financial Reporting Council sanctioned Deloitte over its statutory audits of Go-Ahead Group for FY16 to FY20. The penalty of £6,050,000 was reduced from £11 million, with a 10% discount for exceptional co-operation and 35% for admissions and early disposal. Deloitte also received a severe reprimand, a declaration that five years of audit reports did not meet the relevant requirements, and an order to prepare a root cause analysis for its FRC supervisor.

The underlying accounting was not exotic. According to the FRC, Go-Ahead's LSER rail subsidiary retained erroneous overpayments from the Department for Transport, released £2.4 million to profit in FY16 and accrued a further £27 million from FY16 to FY20. The Department later imposed a £23.5 million penalty on LSER. A second subsidiary, LM, released £5.6 million of accruals to profit in FY20 and worded the related note in a way the FRC said concealed it. A third, GABY, booked an onerous contract provision of €8.1 million for FY20 that was restated to €49.5 million the following year. The FRC found Deloitte failed to challenge these matters sufficiently, apply professional skepticism or evaluate fraud risk indicators. Penrose Foss, the FRC's Executive Counsel, called it "a highly concerning pattern of failure" to apply sufficient scrutiny.

Every one of those balances had a paper trail: a payment received, an accrual released, a provision estimated. The failure was not a missing document. It was that nobody with authority asked why the money was being kept, and whether the note explaining it told the full story. That is a judgment call no summarization tool makes, and it is the same call a controller faces every time a reserve is released to hit a number.

Where Regulators Are Pushing Hardest

The FRC's corporate reporting reviewers point to the same pressure points. Its annual review of corporate reporting, published September 29, said the share of reviews leading to substantive enquiry letters fell for the second consecutive year, a genuine improvement. The areas that still drew the most substantive questions were cash flow statements, financial instruments, impairment of assets, fair value measurement and revenue. "Good reporting is about more than compliance," said Anthony Barrett, the FRC's Executive Director of Supervision.

Several of those areas are estimates, and estimates are where AI-assisted review and human skepticism meet. A model can confirm that an impairment memo exists and that it references the right forecast. Whether the forecast is credible is a question the auditor still has to ask, and the finance team still has to answer with evidence that ties back to the ledger and the operating plan.

Finance leaders know the data underneath is the weak link. Deloitte's Finance Trends 2027 survey of 1,434 finance leaders at companies with $1 billion or more in revenue found that 43% plan to strengthen data governance and lineage for auditability, 49% plan to use AI or automation for routine compliance tasks, and 32% plan to improve alignment between financial and operating reporting. Among respondents whose AI ambitions exceed current capabilities, 41% cited employees' lack of trust in AI output quality as a top barrier. If the people inside finance do not yet trust the output, an auditor running its own models over the same data will not either, unless every number can be traced to its source.

The Playbook Before Year-End Fieldwork

• Ask your auditor what its AI tools do: Find out which procedures now use generative AI, what documents they ingest, and what kinds of gaps they flag, so the team can close those gaps before the request list arrives.
• Reconcile continuously, not at year-end: Balances that are reconciled and reviewed monthly, with support attached, are the cheapest evidence to produce and the hardest for an automated review to fault.
• Document every accrual release and reserve change: The Go-Ahead case turned on releases to profit. Require a written rationale and approver for any release above a set threshold, and keep it with the journal entry.
• Pressure-test estimates against the plan: Impairment, fair value and onerous contract provisions should tie to the same forecasts FP&A uses. A provision that moves from €8.1 million to €49.5 million in a year is a forecasting failure as much as an accounting one.
• Read the notes as a skeptic would: Have someone outside the preparing team check whether each significant disclosure explains the economics plainly, not just technically.
• Map data lineage for automated compliance: Where tax and compliance calculations are automated, keep a clear trail from source transaction to filed number so both internal reviewers and auditors can follow it.