AI governance for the finance function

AI governance in finance is not a sub-category of IT governance. It is its own discipline, with its own controls, its own documentation, and its own audit relationships. Treating it as a downstream consequence of the technology team's tooling decisions is the error that produces the unpleasant audit conversation.

The governance question, properly posed, is not "have we approved this tool?" It is "who owns the output, what evidence supports it, and what review preceded reliance?" The framework below is what answers those questions defensibly.

What "governance" should mean in finance

For the finance function, governance of an AI system has four practical components.

• Ownership. Every AI output that enters a financial process has a named human owner. The model produced the output. The human accepts it, rejects it, or modifies it. The accountability is human.
• Evidence. Every output is traceable to the inputs that produced it, the model version that ran, and the review that was applied. "The system says so" is not evidence.
• Review. Outputs are reviewed before action is taken on them. The review is proportional to the stakes — high-volume, low-individual-stakes outputs are reviewed by sampling and aggregate monitoring; low-volume, high-stakes outputs are reviewed individually.
• Disclosure. Where AI has been used materially in a financial output, the use is disclosed to the audit committee, the auditors, and where required, externally.

Each of these components is straightforward in concept. The work is in instantiating each in a function with hundreds or thousands of distinct outputs per period.

Model risk management, adapted from the banking framework

The banking sector has spent two decades building a mature model risk management framework — the discipline of documenting, validating, monitoring, and challenging quantitative models used in financial decisions. The framework is, with adaptation, the right starting point for AI governance in any finance function.

The core components transfer cleanly:

• Model inventory. A registered list of every AI model used in finance, with its owner, its purpose, its inputs, its outputs, and its risk classification.
• Validation. An independent assessment of whether the model performs as claimed, on data it will encounter in production, with documented findings and remediation.
• Monitoring. Ongoing measurement of model performance against expectations, with defined triggers for re-validation or withdrawal.
• Challenge. A function — separate from the model owner — that has the authority and the competence to challenge the model's outputs.

An AI model used to generate a forecast, a reconciliation, an anomaly flag, or a regulatory submission should be inside this framework. An AI model used to draft an email or summarise an internal document may not need to be — the consequence of error is lower. The framework should be proportionate, but it should exist.

Data provenance and the hallucination problem

Large language models can produce confident, well-formatted, plausible output that is incorrect. In a finance context — variance explanation, contract summary, tax analysis — a hallucination that is not caught becomes a misstatement.

The governance response has two layers.

First, constrain the model to operate on verified data. Where the use case allows, the model is given the source documents it should reason over and instructed to cite them. The output is checkable against the cited source. Outputs that lack citation, or whose citations do not support the claim, are flagged.

Second, structure the review process to assume occasional hallucination. The reviewer is not asked to confirm that the model is reliable in general. The reviewer is asked to confirm that this output, on this material, is supported by this evidence. The discipline scales because it does not require the reviewer to trust the model — only to verify the specific instance.

Access control as a finance discipline

AI systems trained on or given access to sensitive financial data create access control risks that are not the same as the conventional IT access risks. A model that has ingested board materials, M&A analysis, or unreported earnings estimates may, in some architectures, surface that information in response to queries from users who should not have access to it.

The access control framework for AI systems must be at least as rigorous as the framework for the underlying data. In practice this means:

• Per-user access controls applied at the AI interface, not just at the data store.
• Audit logging of who asked the AI what, and what answer it gave.
• Periodic review of the data the model has been trained on or given retrieval access to.
• A clear deletion path when sensitive data must be removed from the model's reach.

Audit readiness: the question the auditor will ask

The external auditor, increasingly, asks whether AI-generated outputs have been subject to human review before reliance is placed on them. The question is not whether AI was used. It is whether the finance function owns the output.

The defensible answer has three components: the governance framework that requires review, the evidence that review occurred on the specific outputs in scope, and the demonstration that the framework is applied consistently. Each component requires documentation that is contemporaneous, not reconstructed.

A finance function that can produce, on request, the model inventory, the review log for any sampled output, the incident history, and the model performance metrics will pass the audit conversation. A finance function that cannot will spend a long quarter explaining why.

The minimum viable governance pack

For a finance function beginning to deploy AI materially, a minimum viable governance pack has the following components. None is optional.

• A policy statement signed by the CFO and approved by the audit committee.
• A model inventory, maintained current.
• A validation report for each model in the inventory, refreshed at a defined cadence.
• A review log that records, for each in-scope output, the reviewer, the date, and the outcome.
• A monitoring dashboard for the model performance metrics that matter.
• An incident log, with root cause and remediation for each event.
• A periodic governance report to the audit committee.

None of this is unusual in the wider risk-management context. It is the standard discipline applied to a new model class, with attention to the specific failure modes AI introduces.

This piece sits inside the CFO in AI framework. See also agentic FP&A and the continuous forecast and audit readiness in an AI-enabled finance function. Lorna writes from practice at IMPT. The verified page records what is and isn't published here.