Beyond the Introduction: Governance, Not Gadgets

Our general primer — see AI in Accounting & Audit — explains what AI does in finance. This guide is the technical companion: how to control it. Assume the tools are already in the building (they are — expense coding, reconciliation matching, draft flux commentary, contract extraction). The question is no longer whether finance uses AI, but whether each use is risk-assessed, approved, monitored, and auditable. If the answer to any of those is no, the efficiency gain is an unmeasured control deficiency.

How Finance Departments Actually Use AI Today

Inventory reality before governing theory. The common deployed uses, ranked by statement risk:

  1. Transaction coding and matching (expenses, AP, bank rules): high volume, low individual materiality, errors aggregate — see Balance Sheet Reconciliation for where the aggregates surface.
  2. Reconciliation assistance and exception prediction: proposes matches and flags anomalies; risk concentrates in auto-cleared items nobody reviews.
  3. Draft narratives (flux commentary, disclosure drafts, board variance notes): low direct-booking risk, high reliance risk — plausible prose that reviewers sign without verifying.
  4. Contract and document extraction (lease terms, covenants, revenue clauses): errors propagate into recognition and measurement judgments.
  5. Forecasting and anomaly scoring: influences provisions, impairments, and going-concern inputs — model risk with audit implications.

Every use in production needs an entry in the approved-use inventory with its risk tier, data classification, owner, and approval path. Shadow AI — staff pasting GL data into public chatbots — is the first finding; address it in policy before anything else.

Risk Taxonomy: Name Each Risk to Control Each Risk

Risk What goes wrong Statement impact
Accuracy / hallucination Confident fabrication of figures, citations, standards Misstatement; false audit evidence
Data leakage Confidential figures used as training data or exposed via prompts Breach; regulatory exposure — see NIST CSF for Finance
Bias / inconsistency Same facts, different answers across runs or counterparties Uneven provisions, selective flagging
Model drift Silent degradation after retraining or data shifts Error rates rising unnoticed
Over-reliance Reviewers rubber-stamp plausible output; skills atrophy Control operation fails in substance
Accountability gaps Nobody owns the output when it is wrong Deficiency in control environment
Vendor lock / opacity Black-box updates, unlogged changes, data residency shifts Loss of audit trail; compliance breach

Score each deployed use against this table on likelihood × impact. Anything scoring high on accuracy and touching the statements needs the full control stack below; low-risk drafting assistance needs proportionally lighter controls. Tiering is what keeps governance credible instead of bureaucratic.

Internal Controls for AI-Assisted Accounting

Map AI controls onto COSO rather than inventing a parallel universe — see COSO Framework:

  • Control environment: approved-use inventory, named owners, acceptable-use policy signed by finance staff, training on AI limitations and escalation.
  • Risk assessment: tiered scoring per use case, re-performed on model changes, vendor updates, or new data feeds.
  • Control activities: input validation (AI works only on approved, reconciled source data), restricted action rights (AI proposes, humans post), and automated pre-issuance checks on AI outputs.
  • Information and communication: versioned model and prompt registry; every output traceable to model version, inputs, and reviewer.
  • Monitoring: error-rate and override tracking with thresholds that trigger retraining, restriction, or decommissioning.

The key architectural decision: AI proposes, humans dispose. No AI agent posts directly to the GL, releases payments, or files returns without human approval inside the system — enforced by access rights, not by policy language.

Human Oversight and Approval Procedures

Calibrate review to risk tier:

  • Tier 1 (low): drafting and summarization — reviewer verifies against sources before use; sampling suffices for high-volume coding with quality gates.
  • Tier 2 (medium): matching, extraction, anomaly flags — 100% review of exceptions and overrides, sampling of auto-cleared items, with reviewer independence (not the person who configured the rules).
  • Tier 3 (high): anything feeding recognition, measurement, provisions, or disclosures — independent re-performance on a risk basis, documented skepticism, and sign-off identifying the human accountable for the number.

Rotate reviewers periodically and track override rates: a reviewer who never overrides is either perfectly served or not reviewing. Spot-check with seeded test cases quarterly — known-error inputs that must be caught — to prove the oversight operates rather than exists.

Audit Trails and Documentation

If it is not logged, it did not happen — and auditors will treat AI-assisted balances accordingly. The minimum trail per material AI touchpoint: input data snapshot (or immutable reference), model and version, prompt or configuration version, raw output, human review actions with timestamps, and final posted values with journal references. Retain model-change logs from vendors contractually (demand them in procurement — see below) and file the trail with the close binder per Month-End Close so walkthroughs start from evidence, not explanations. Coordinate expectations with internal audit early using the Internal Audit Framework.

Vendor Risk Assessment for AI Tools

Every AI vendor touching finance data gets assessed before contracting and re-assessed on material updates:

  • Data handling: training-data use, retention, sub-processors, residency, and deletion — in the contract, not the sales deck. Public-model tools that retain prompts are disqualified for confidential finance data by default.
  • Change control: advance notice of model updates, version pinning options, and rollback rights during close-critical periods.
  • Security posture: SOC 2 or equivalent, access controls, encryption, incident notification SLAs — mapped to NIST CSF expectations.
  • Explainability and support: can the vendor show what the model did and why, and will they support an audit inquiry with documentation?
  • Exit plan: data export in usable form, no hostage formats, transition assistance priced in advance.

High-risk EU deployments add the regulatory layer: classify the use under the EU AI Act and assign its obligations (documentation, human oversight, logging) to named owners before go-live.

Monitoring Models in Production

Deploy metrics with teeth:

  • Error and correction rate per use case, trended monthly; threshold breaches trigger root-cause analysis and possible suspension.
  • Override and rejection rates by reviewer; outliers investigated for either poor tuning or poor reviewing.
  • Drift indicators: match-rate changes, score-distribution shifts, or complaint spikes after vendor updates — with a freeze-and-assess rule during close week.
  • Coverage metrics: percentage of in-scope volume actually processed through the governed path vs. manual bypasses that evade controls.

Report the dashboard to the controller monthly and to the audit committee quarterly — evidence-pack discipline per AI Governance Evidence Pack keeps the reporting consistent and the committee’s oversight demonstrable.

Building the Finance AI Governance Policy

Structure the policy as operable rules, not principles wallpaper:

  1. Scope and inventory: covered tools and uses, tiering criteria, and the approval path for adding a use.
  2. Data rules: classification tiers, prohibited destinations for confidential data, and approved-tool lists per tier.
  3. Oversight matrix: review and approval requirements per tier with named roles and evidence standards.
  4. Trail and retention: logging requirements, storage locations, and retention periods aligned to audit needs.
  5. Vendor requirements: assessment cadence, contractual minimums, and update-notification obligations.
  6. Monitoring and incidents: metrics, thresholds, suspension triggers, and the incident playbook (contain, assess statement impact, remediate, disclose as required).
  7. Training: role-based onboarding plus annual refreshers, with shadow-AI prohibitions explicit and enforced.

CFO Governance Checklist

  • Complete inventory of AI uses in finance with risk tiers, owners, and data classifications — including shadow tools discovered by survey, not assumed away.
  • Human-oversight matrix published with tier-calibrated review, evidence, and accountability for every Tier 2/3 use.
  • AI action rights restricted in systems: propose-only access, human posting, enforced by configuration.
  • Audit-trail standard implemented for all material AI touchpoints and filed with the close binder.
  • Vendor assessments current with contractual data, change-control, and exit protections; EU uses classified under the AI Act.
  • Monitoring dashboard live with thresholds, suspension triggers, and quarterly audit-committee reporting.
  • Incident playbook tested once (tabletop) before the first real failure forces an improvisation.

FAQ

Can AI outputs serve as audit evidence?

Only with the full trail: inputs, version, output, and documented human review. Standalone AI-generated analysis without human verification and sourcing is not competent evidence — auditors will ask who verified it and how, and “the model” is not an acceptable answer.

How do we handle staff using public AI tools for finance work?

Prohibit confidential finance data in non-approved tools explicitly, provide approved alternatives that actually work, monitor for violations proportionately, and enforce consistently. Policy without a usable alternative just drives the behavior underground.

Should small finance teams adopt lighter AI governance?

Lighter, not absent: tier ruthlessly so the one or two high-risk uses get real controls (human approval, trails, vendor terms) while low-risk drafting gets simple verification rules. A two-page policy that is followed beats a twenty-page one that is not.

Next Steps

  1. Run the use-case inventory this month — survey the team, check expense and browser histories proportionately, and tier everything found against the risk table.
  2. Implement propose-only system rights and the oversight matrix for Tier 2/3 uses before the next close; these two controls eliminate the largest sources of unaudited AI impact.
  3. Stand up the monitoring dashboard and book the first quarterly audit-committee slot — governance that is never reported is governance that never happened.