Internal Control Over Generative AI, Applied to the Close

Why this matters to controllers

Most close processes already have defined responsibilities: data preparation, reconciliations, journal-entry review, exception handling, approval and retention of support. AI can change one or more of these steps even when the final approval remains with a person.

Three practical risks follow:

• Evidence risk: The AI-generated result looks plausible, but the reviewer cannot establish which source population or calculations support it.
• Change risk: A prompt, model, connector or matching rule changes between reporting periods, affecting the output without a corresponding review of the control design.
• Oversight risk: A reviewer approves the result without a sufficiently specific test of completeness, accuracy or exceptions.

These are recognizable internal-control problems, expressed through a new technology. An AI inventory that sits outside the close-control documentation will not, by itself, demonstrate that these risks are addressed.

What COSO actually changed, and what it did not

COSO’s 2013 Internal Control — Integrated Framework sets out five integrated components and 17 principles of internal control. The COSO-commissioned publication translates that framework into practices for AI applications. It is guidance, not a new regulation or a replacement for the original framework: it states that it contains general information only and that its views may differ from those of regulators, and the 2013 framework states that no other COSO guidance takes precedence over it. It also focuses on risks unique to or significantly amplified by generative AI and does not repeat established IT general controls, such as segregation of duties in ERP systems, which it says remain necessary.

For SEC registrants subject to the relevant provisions, management evaluates the effectiveness of internal control over financial reporting (ICFR) as of each fiscal year end using a suitable, recognized control framework. Management also evaluates any change in ICFR during each fiscal quarter (each fiscal year, for a foreign private issuer) that has materially affected, or is reasonably likely to materially affect, ICFR. Introducing AI should therefore prompt an assessment of whether a financial-reporting process or control has changed. Not every AI implementation is automatically a material ICFR change. The nature and significance of the use determine the required response.

The following four practices are a CoE | AI Cost & Governance implementation interpretation of the guidance, not a checklist the publication mandates; where a practice draws on the publication’s own text, that is stated.

Four controls to put into practice

1. Know where AI participates in the close

Start with the actual workflow rather than a company-wide list of approved applications. Identify each AI use that influences reconciliations, accruals, journals, eliminations, management review, or financial disclosures.

The publication makes this the second step of its implementation roadmap: an inventory of all active and planned generative AI use cases, including periodic scans or surveys to detect “shadow AI” use, with each use classified by capability type, assigned an owner and documented. Its eight capability types range from ingestion and transformation to posting, judgment and monitoring.

For each use, record the business purpose, process owner, system or data sources, type of output, whether the output is relied on, and its level of authority. Distinguish an assistant that drafts commentary from an agent that posts entries: they do not create the same risk.

Practical test: Could the controller identify every AI-dependent step in a material close process and the person accountable for it?

2. Establish what evidence makes AI output usable

A reconciliation proposed by AI is not supported merely because the tool labels it complete. Finance needs access to the source population and a way to test the result.

Evidence requirements should address source and extract identifiers, completeness of inputs, method or transformation, exceptions, output version and review performed. Retain prompts and configuration details where they materially affect reproducibility or control operation. These records may be held in existing close workpapers rather than a new AI-only repository.

The publication is specific on this point. Under principle 13, on relevant, quality information, it says processes using generative AI should capture and store the information needed to understand and validate output, including prompts, inputs, outputs, source references, model and configuration versions, and any confidence scores, to the level of explainability the control activities require, because each can affect control conclusions. Drawing on what it describes as recent industry discussion, it adopts a working definition of AI reliance: management depending on AI outputs as part of the evidence that an ICFR control is designed or operating effectively. Where reliance exists, it says AI controls should meet the same evidence standards expected for ICFR, including a documented prompt, configuration and model version, a clear sampling rationale with exception resolution, and retention of sufficient evidence.

Practical test: Can another qualified reviewer trace an AI-proposed reconciliation or accrual back to independent source records, including items the tool did not match or process?

3. Govern changes to the AI process

A recurring accrual process would not normally permit uncontrolled changes to its underlying logic. An AI-supported process should receive comparable attention when changes could affect reporting results.

Define who may update prompts, connectors, model configurations, source mappings and automated actions. Record material changes, test their effect and obtain approval before production use where warranted. Finance should own the financial-reporting requirements; IT, security and platform teams may operate the technical controls.

The publication treats prompts, system prompts, retrieval connectors and transformation rules as governed configurations with version history, approval workflows and rollback plans. Among its common control activities, it lists separating the ability to configure AI settings from the authority to approve or review outputs. In close terms, we read this as: the person who tunes the accrual prompt should not be the person who signs off the accrual.

Practical test: If the system produced a different result this month, could the team tell whether financial activity changed, the input changed or the AI configuration changed?

4. Monitor performance and investigate exceptions

AI-supported controls need more than initial implementation testing. Monitor whether expected source populations arrive, matching or exception rates change, significant outputs differ from independent benchmarks, and unusual results receive timely investigation.

Thresholds should be risk-based. A tolerance appropriate for drafting management commentary may be unacceptable for automated journal posting. Some processes need transaction-level deterministic checks, even when an AI model contributes to the workflow.

The publication anticipates this shift: because generative AI is probabilistic rather than deterministic, it says the definition of a control failure may need to evolve, with organizations relying on several tolerance measures rather than a single pass or fail.

Practical test: What specific result would trigger investigation, who investigates, and what evidence shows the issue was resolved?

Where conventional controls need adaptation

An unchanged prompt is not always the same as an unchanged process

Generative AI output can vary with model behavior, retrieved information, system configuration and other context: the publication describes outputs as probabilistic and capable of being confidently wrong. 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. Retaining the prompt alone may not explain a result. Finance may need to preserve input versions and other execution context, and validate outputs through independent rules or source-based checks.

Human sign-off is not sufficient on its own

A named reviewer is valuable only when the review is designed to detect specific errors. “Reviewed AI output” is not the same as checking input completeness, accounting treatment, threshold breaches or unexplained changes.

Shared responsibility is not the same as unclear responsibility

Finance should remain accountable for financial-reporting assertions and acceptance criteria. IT, cybersecurity, risk and model-governance teams may own identity management, logging, platform configurations and monitoring infrastructure. The control design must identify these dependencies clearly.

A practical implementation sequence

First, identify reliance. List AI uses in the close and separate experimental assistance from outputs relied on for bookings, reconciliations, management reviews or disclosures.

Next, map the existing controls. For each relied-upon output, identify its current control owner, source evidence, review procedure, system dependencies and gaps introduced by AI.

Then, define and test new requirements. Establish completeness checks, source traceability, permission boundaries, configuration approvals, monitoring thresholds and evidence retention proportional to the risk.

Finally, evaluate and revisit. Determine whether the implementation changes ICFR, document the assessment as appropriate, and reassess after material process or configuration changes. Coordinate with internal audit and external auditors where relevant.

Where practice and guidance stand

The COSO-commissioned publication provides a structured approach for adapting established internal-control principles to generative AI. In Gartner’s 2025 survey of 183 CFOs and senior finance leaders, 59% said their finance function used AI; among those that had implemented it, accounts payable automation and error and anomaly detection were among the most common uses. In April 2026, PCAOB staff also reported that audit committee chairs are increasingly seeking their auditors’ views on AI’s potential impacts on financial reporting and controls.

How external auditors will approach AI used in financial reporting is not yet settled: 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. PCAOB staff said in March 2025 that its target team’s 2024 inspections included reviewing auditors’ assessments of companies’ use of AI, and that a report on that work would follow; none had appeared on the PCAOB’s staff publications page by October 2026.

Show that the changed close controls still work

AI does not remove the need for internal control. It changes where the underlying risks can arise and what evidence a control needs to work.

The most useful question for a controller is not “Do we govern AI?” It is “Which close controls now depend on AI, and can we demonstrate that those controls still work?”