The Big Four's AI Problem: Every Number Needs a Receipt

An AI-generated number is not a fact. It is a claim, and in finance every material claim needs evidence.

That distinction became difficult to ignore in July 2026, when researchers flagged false, misleading, or unverifiable citations in four PwC Middle East thought-leadership reports. The Financial Times independently checked examples from the investigation. PwC said it was updating a limited number of supporting citations and reiterated that its people are expected to follow its quality-control processes.

The wording matters. The reports contained problems that researchers described as characteristic of AI hallucination, but PwC did not publicly confirm that AI authored the reports. The episode involved thought leadership, not an audit opinion or a client's financial statements. Even with those limits, it is a useful warning for every finance team introducing AI into reporting, forecasting, reconciliation, and decision support.

What happens when an AI-generated financial insight looks convincing but cannot be traced back to the books?

Finance has a verification problem, not an adoption problem

AI is already useful in finance. It can draft variance explanations, summarize management reports, classify documents, identify unusual movements, and help a reviewer find the transactions most likely to need attention. The opportunity is real. So is the control gap.

A global survey of 1,600 finance professionals found that 93% were concerned about the integrity and verifiability of AI-generated insights. At the same time, more than 60% said their teams had increased their use of real-time operational data during the previous two years. The findings, published by ACCA and Chartered Accountants Australia and New Zealand, describe a finance function processing more information, more quickly, without always having the governance or skills needed to verify the result.

That combination creates a specific risk. In marketing, an AI error might produce an awkward sentence. In accounting, a plausible error can misclassify an expense, distort a forecast, overstate profit, or influence a decision involving real money. Fluency is not evidence, and confidence is not a control.

The ledger must remain the source of truth

Large language models are designed to generate plausible responses. Accounting systems are designed to preserve transactions and balances. Those are fundamentally different responsibilities.

AI can interpret financial data, but it should not become an alternative source of financial truth. Revenue, expenses, assets, liabilities, equity, and cash must still trace to their underlying ledger entries. Derived measures should use one canonical definition that can be filtered and drilled back to transaction grain.

If an AI assistant says gross profit declined by 12%, a reviewer should be able to inspect:

• The transactions included in the calculation.
• The current and comparison periods.
• The account and report-section mappings applied.
• The formula, polarity, and sign rules used.
• Any entity, department, class, currency, or other filters.
• The reviewer and approval status.

Without that evidence, "gross profit declined by 12%" is not yet a financial insight. It is an unsupported sentence containing a number.

This is also why a trustworthy AI finance architecture needs evidence attached at ingestion rather than reconstructed after an answer is generated. Our reference architecture for trustworthy AI finance systems explains how source evidence should travel with financial objects through calculation, review, and approval.

Five controls every AI-generated financial insight needs

1. Source-backed numbers

Every material number should link to its source data. A reviewer must be able to move from an executive summary to the report, from the report to an account balance, and from that balance to the individual transactions.

The source reference should travel with the number. Asking the system to search for support after it has generated an answer is weaker because the retrieved evidence may not be the evidence that influenced the original output. AI should make drill-down faster, not make provenance optional.

2. Deterministic calculations

AI can explain a calculation, but established financial measures should be computed using defined formulas. Net income should not change because a prompt was worded differently. A balance sheet should balance. Debits and credits must remain equal.

Use AI for interpretation, planning, and communication. Use deterministic logic for calculations, accounting sign rules, period handling, and report rollups. If a financial measure exists in several dashboards or reports, each surface should consume the same canonical calculation rather than maintain a separate version that can drift.

3. Automatic reconciliation

AI-generated reporting should be compared with the accounting source before anyone acts on it. The system should identify missing accounts, unexpected differences, sign errors, incomplete periods, duplicate records, and totals that do not reconcile.

Reconciliation turns confidence from a feeling into a testable result. A useful output does not merely say that a report passed. It shows the source total, calculated total, difference, tolerance, and any exceptions requiring review.

The same principle applies to close controls more broadly. In our guide to audit-ready controls, we explain how evidence and review can be built into the workflow instead of assembled in a panic after period end.

4. Human approval for consequential actions

AI can prepare a draft, flag an exception, or recommend an investigation. It should not silently publish a board report, change an account mapping, post a journal entry, or file a return.

A qualified person must remain accountable for actions that materially affect the books. In a survey of 100 middle-market CFOs, two-thirds described human oversight of agentic AI as extremely or very important for accuracy, according to the Journal of Accountancy. Human review should be a defined control with an owner, evidence, and threshold, not a vague promise that someone remains "in the loop."

5. An immutable audit trail

Finance teams need to know what the AI proposed, which data it used, which rules it applied, what a reviewer changed, who approved the final result, and when each event occurred.

This history should be append-only. The original recommendation must not disappear when someone edits the final output. A complete record supports accountability, makes errors reproducible, and provides the feedback needed to improve future recommendations without hiding earlier mistakes.

The right role for AI in accounting

The choice is not between fully autonomous AI and no AI at all. The better model is a clear division of responsibility:

AI proposes. Financial systems calculate. Accountants approve. Audit trails remember.

This model lets finance teams analyze more data and investigate exceptions faster without weakening the controls that make financial information credible. It also gives AI the work it is good at: organizing context, surfacing patterns, drafting explanations, and guiding a reviewer toward the evidence that matters.

Leading finance organizations appear to be reaching the same conclusion. KPMG's 2026 survey of more than 1,000 senior finance leaders argues that as AI influences financial information more directly, independent assurance becomes more important for validating data integrity, model reliability, cyber resilience, and the accuracy of AI-driven insights. Read KPMG's summary of the research.

A practical test before trusting an AI-generated report

Before using an AI-generated financial insight, ask five questions:

1. Can I trace every material number to its underlying transactions?
2. Was the number calculated using a defined, repeatable formula?
3. Does the result reconcile with the accounting source?
4. Has a responsible person reviewed the assumptions and exceptions?
5. Can I see a permanent record of what the AI did?

If the answer to any question is no, the output may still be useful as a draft or investigative lead. It is not ready to support a financial decision.

Frequently asked questions

Can finance teams trust AI-generated reports?

Yes, when material numbers trace to source transactions, calculations are deterministic, outputs reconcile to the accounting system, consequential actions require human approval, and every change is recorded in an audit trail. Trust should come from these controls, not from how polished the answer sounds.

What should an AI audit trail contain?

An AI audit trail should record the source data used, calculation or rule applied, AI proposal, reviewer edits, approval decision, reviewer identity, timestamp, and reconciliation result. The history should be immutable and append-only.

Should AI calculate financial measures?

AI can help assemble a query or explain a result, but established financial measures should be calculated using defined, repeatable formulas at transaction or ledger grain. The same inputs and rules should produce the same number regardless of how a user phrases the request.

AI should accelerate trust, not replace it

The recent problems found in major professional reports are not an argument against using AI. They are an argument against accepting polished output without inspecting the evidence behind it.

Finance teams do not need AI that merely sounds confident. They need systems that can show their work, reconcile to the books, preserve review decisions, and make every material number traceable.

That is the standard FinBoard is built around: live financial reporting and analysis with drill-down, reconciliation, human review, and an auditable path back to the underlying data. Explore FinBoard to see how AI-assisted finance can move faster without turning trust into guesswork.

Because in accounting, every number needs a receipt.

This article is general information and does not provide accounting, audit, tax, legal, or investment advice.