Treat AI Output Like an Unsupported Journal Entry

A review of outputs cannot see a missing entry

A review of proposed entries could find an incorrect amount or period. It could not, by itself, establish that every expected obligation had been considered.

For a close owner, the key design question is: what independent expectation would make the missing accrual visible? Supplier activity, contracts, approved source populations and known billing changes can provide that expectation.

Why this matters in the close

Finance teams already expect journal entries to have a supportable basis. If an analyst proposes an accrual, a reviewer can inspect the relevant contract, invoices, calculation, account coding, and approval. A reconciliation should similarly connect balances and differences to underlying records.

AI can help prepare these outputs faster. It may propose an accrual, match thousands of transactions, flag exceptions, or draft variance explanations. But speed changes neither the accounting assertion nor the need for reliable support.

Consider three ways AI might enter the same close:

1. Drafting: AI suggests a potential accrual for an analyst to investigate. The suggestion is working material, not yet a booked amount or control conclusion.
2. Preparing: AI calculates an accrual and attaches supporting records for a designated reviewer. The review must establish whether the calculation and source population are appropriate.
3. Deciding or evidencing: AI posts the accrual, marks a reconciliation complete, or supplies the evidence used to sign off a control. The organization is now relying on the output, so the controls must address the output’s accuracy, completeness, authorization, and traceability.

The point of reliance matters more than whether an AI system was used. The COSO-commissioned publication draws a similar line: under its working definition, AI reliance occurs when management depends on AI outputs as part of the evidence supporting the design or operating effectiveness of an internal control over financial reporting (ICFR), and where reliance exists, the AI controls should meet the evidence standards expected for ICFR.

The overlooked distinction: accuracy versus completeness

An AI-generated result can be wrong in two different ways:

• Incorrect inclusion: It proposes a journal entry for the wrong amount, account, period, or entity, or matches transactions that do not belong together.
• Missing information: It fails to propose an accrual, omits transactions from the population, or overlooks an unresolved reconciliation item.

The first problem can sometimes be found by reviewing the proposed entry or match. The second may call for an independent completeness check, for example reconciling the input population to an approved source report, comparing expected suppliers to recorded accruals, or investigating significant account-level movements.

A reviewer cannot validate the completeness of the population by inspecting only the records selected by the AI. This is why a polished reconciliation report is not, by itself, sufficient evidence that all relevant items were considered.

The guidance behind the recommendation

PCAOB Auditing Standard AS 2401 directs auditors to design procedures to test the appropriateness of journal entries and other adjustments, because fraud often involves recording inappropriate or unauthorized entries. It lists characteristics that often identify inappropriate entries, among them entries to unrelated, unusual, or seldom-used accounts, entries made by individuals who typically do not make journal entries, and period-end or post-closing entries with little or no explanation. It also requires auditors to select entries from the general ledger and examine their support. This is an auditing standard governing auditors, not a direct rule that every company must treat every AI output as a journal entry.

PCAOB staff’s January 2025 Audit Focus: Journal Entries reminds auditors to understand the design of controls over both automated and manual journal entries, to consider fraud risks in entries processed through automated systems, and to test the completeness of the journal-entry population; failing to do so is among the common deficiencies it reports from inspections of firms that audit smaller public companies. It states the views of PCAOB staff and is not a rule, policy, or statement of the Board.

NIST’s Generative AI Profile identifies confabulation: the production of confidently stated but erroneous or false content. The COSO-commissioned publication says generative AI outputs are probabilistic and can be confidently wrong, and that controls should treat them as claims requiring validation rather than facts to accept by default. It calls for human corroboration in proportion to risk, says processes using generative AI should capture and store the information needed to understand and validate output, including prompts, inputs, outputs, source references, and model and configuration versions, to the level of explainability the control activities require, and lists periodic manual re-performance of reconciliations among its illustrative controls for automated transaction processing and reconciliation. It also states that it contains general information only. These sources support a risk-based validation approach; they do not establish a universal required evidence record for every AI response.

None of these sources prescribes a single evidence packet for generative-AI-supported accounting work. PCAOB staff lists guidance on what it now calls companies’ and auditors’ use of “super intelligence” among the topics it is considering; none has been issued. The UK audit regulator’s March 2026 guidance lists inconsistency among the kinds of error AI output can contain, alongside hallucinations, omissions, distortions and faulty reasoning. Keeping a prompt and configuration version alone does not guarantee reproducibility when models, data and workflows change.

The proposed control standard below is CoE | AI Cost & Governance analysis informed by those sources.

A practical support standard for AI-assisted close outputs

Before finance relies on an AI-generated amount, match, exception conclusion, or explanation in a material close process, it should be able to answer six questions.

This record should be proportionate to the risk. An exploratory suggestion does not warrant the same documentation as a material, automated entry or an ICFR control conclusion. A prompt log can help, but a saved prompt does not establish that an output is correct, complete, or reproducible.

A simple implementation example: supplier accruals

A close owner could apply the standard without introducing a new software platform:

Step 1: Define the decision. The AI may propose expense accruals for review; it may not post them without an authorized control step.

Step 2: Define the population. Establish approved supplier, contract, purchase-order, and invoice sources for the period. Reconcile source counts and totals where appropriate.

Step 3: Preserve the basis. Save the calculation logic, relevant assumptions, data references, workflow configuration, and output.

Step 4: Test both directions. Verify that sampled proposed accruals are supported (validity and accuracy) and that expected obligations were not omitted (completeness). Compare current activity to prior periods and known billing-cycle changes.

Step 5: Record the review. Resolve exceptions, document material judgments, and retain approval before posting or using the result as control evidence.

Warning signs that warrant closer review

• An AI reconciliation says “complete” but provides no independently checked source population.
• The tool proposes journal entries without retrievable underlying documents or a testable calculation.
• A result is approved solely because the narrative sounds plausible.
• A change to prompts, model configuration, retrieval logic, or source connections is not tracked.
• Exceptions vanish between periods without an explanation.
• Review is performed after posting, solely to reconstruct documentation for auditors.
• The same system generates the result and is treated as the only evidence that the result is correct.

None of these signs automatically proves the accounting is wrong. Each indicates a gap in the evidence or control design that should be investigated.

What finance leaders should do next

Start with a targeted inventory of AI-supported close activities. For each activity, identify where its output first becomes relied-upon information: a booked journal entry, a reconciliation sign-off, a variance conclusion, a control assertion, or disclosure support.

Then prioritize material workflows and establish minimum evidence requirements. In particular, ensure that the validation process can detect omissions, not just erroneous items that the AI happened to produce. Retain source references and approved calculations in existing close workpapers where practical; introduce separate records only where the present workflow does not provide adequate traceability.

Finally, agree with the controller, internal audit, and external auditor as appropriate on how the controls fit the organization’s reporting requirements and risk assessment.

Before the output becomes evidence

Finance should not accept an AI-generated result as support for itself. Before a material AI-assisted close output is booked, approved, or used as evidence, the organization needs traceable source information, a defensible method, an appropriate completeness check, and accountable review.

The most dangerous result may not be the wrong number the AI produced. It may be the number or exception it never produced at all.