Writing Strong AI Audit Findings

An AI audit finding can sound sophisticated and still be useless. Statements such as “AI governance is inadequate” or “the model presents hallucination risk” may be technically plausible, but they do not tell management which expectation was missed, what evidence proves the gap, why it exists, what business decision is exposed, or what needs to change.

The problem is rarely a shortage of AI terminology. It is a broken chain of audit logic.

A strong AI audit finding turns evidence into a decision. It connects an applicable criterion to a verified condition, investigates cause, translates the deviation into risk or effect, and leads to proportionate action. AI changes the subject matter; it does not replace the fundamentals of sound audit work.

The finding logic does not change because the subject is AI

The Global Internal Audit Standards provide a useful progression. Standard 14.2 requires internal auditors to analyze relevant, reliable, and sufficient information and determine whether a difference exists between evaluation criteria and the current condition. Standard 14.3 requires potential findings to be evaluated for significance, including collaboration with management to identify root causes when possible and determine potential effects. Standard 14.4 moves the analysis toward recommendations or action plans that resolve the difference, mitigate risk to an acceptable level, address root cause, or improve the activity.

Standard 15.1 completes the chain through final engagement communication. For assurance engagements, findings, significance and prioritization form part of the required communication, alongside the applicable engagement conclusions, responsible individuals, and planned completion dates for actions.

That does not mean every audit report must display five labeled boxes for criteria, condition, cause, effect, and recommendation. The sequence is a practical reasoning model, not a mandated report layout. The Standards themselves distinguish requirements from considerations for implementation and examples of evidence of conformance. What matters is that the logic exists, is supported, and survives challenge even when the final wording is concise.

A finding does not prove that the auditor understands AI jargon. It proves that the auditor can explain a deviation well enough for someone to decide what to do about it.

What makes AI findings harder to write

AI creates four recurring complications.

First, criteria often come from multiple layers: law and regulation, internal policies, contractual commitments, adopted standards, control architecture, and voluntary frameworks. Second, some AI behavior is probabilistic or changes with configuration, data, vendor updates, or context of use. Third, the cause is rarely just “the model”; it may sit in data, workflow design, human oversight, change management, vendor governance, or unclear accountability. Fourth, a technical risk matters only when it is connected to an organizational objective, decision, or exposure.

Element	Question it must answer	Common AI writing trap	Better practice
Criteria	What should happen, and why does it apply?	Treating voluntary good practice as a binding requirement.	Identify the authority: law, policy, contract, adopted standard, or guidance used as a benchmark.
Condition	What is actually happening?	Generalizing from one surprising output or isolated test.	Document population, sample, version, configuration, period, prompts or test cases, and reproducible evidence where feasible.
Cause	Why does the gap exist?	Blaming “the model” without investigating the process.	Assess ownership, data, design, changes, human oversight, monitoring, and third parties before concluding.
Risk or effect	What could happen, and why does it matter?	Listing generic labels such as bias, privacy, or hallucination.	Link the gap to a decision, stakeholder, objective, and residual risk.
Action	What should change to reduce the risk?	Prescribing a specific model or tool without necessity.	Strengthen the control objective, assign accountability, define evidence, and leave management responsible for the solution unless a specific requirement dictates otherwise.

This discipline also prevents auditors from confusing uncertainty with deficiency. A generative system may produce variable outputs without that variability, by itself, constituting a finding. A finding emerges when the observed behavior demonstrates a meaningful gap against an applicable criterion and creates an exposure worth communicating.

Do not turn voluntary frameworks into compliance findings

AI audits often combine frameworks, but those sources do not carry the same authority.

The Global Internal Audit Standards govern how internal audit performs and communicates its work; they are not, by themselves, a catalog of AI controls that the audited organization must implement. The IIA’s Artificial Intelligence Auditing Framework provides professional guidance on AI governance, management, risks, and controls, but it remains supporting guidance rather than a legal requirement.

ISO/IEC 42001:2023 specifies requirements for an AI management system. It can provide strong evaluation criteria where the organization has adopted the standard, the engagement is assessing conformity with it, or another basis makes those requirements applicable. An auditor should not label an issue an “ISO nonconformity” when ISO/IEC 42001 was never part of the engagement criteria.

The NIST AI Risk Management Framework is voluntary and outcome-oriented. Its Generative AI Profile, NIST AI 600-1, adds risk-management considerations for risks that are unique to or exacerbated by generative AI. NIST also notes that AI RMF 1.0 is being revised. That makes it valuable for benchmarking and criterion design, not an automatic source of “noncompliance.”

Likewise, COSO’s 2026 guidance on effective internal control over generative AI translates familiar internal-control concepts into GenAI-specific practices. It can sharpen control expectations without becoming legislation.

The practical rule is simple: before drafting the finding, document why the criterion applies. That sentence can prevent half of the debate later.

Example: turning an observation into a defensible finding

Consider a hypothetical organization using a generative assistant to support customer refund decisions. Its internal AI policy requires documented human review for refunds above a defined threshold and a quarterly performance review of the system.

A weak observation might read:

“The AI assistant does not have sufficient human oversight and could make incorrect decisions.”

That is a concern, not yet a defensible finding.

A stronger version could be built as follows:

Criteria. The internal AI policy requires evidence of human review for above-threshold cases and a quarterly performance assessment of the assistant.

Condition. In a sample of 50 above-threshold refunds, 12 had no evidence of human review. The team also could not provide the quarterly performance assessment for the period tested. The audit team validated the result against transaction records and the approval repository.

Cause. Control responsibility is split between Operations and the AI product team. The responsibility matrix does not assign an owner for verifying human review or retaining monitoring evidence.

Risk or effect. Incorrect or inconsistent decisions could be approved without timely detection, creating customer, financial, conduct, and reputational exposure. Significance should be assessed under internal audit’s methodology, considering likelihood, impact, existing controls, and risk tolerance.

Action. Management should assign a control owner, define minimum evidence for human review, establish performance-review metrics and thresholds, and perform a risk-based retrospective review of affected cases.

The improvement is not more AI vocabulary. It is the move from a generic concern to a traceable relationship among expectation, evidence, exposure, and decision.

Five quality checks before you issue the finding

Before closing an AI audit finding, test it against five questions:

Applicability: Can you explain in one sentence why the criterion is binding, adopted, or being used as guidance for this engagement?
Evidence: Is the condition supported by relevant, reliable, and sufficient information, including enough technical context to understand or reproduce the test?
Causality: Is the root cause supported, or is it honestly presented as a hypothesis that still needs validation?
Risk: Does the effect connect to objectives, decisions, and residual risk rather than relying on generic AI labels?
Action: Does the response address cause or reduce risk to an acceptable level without taking management’s responsibility for choosing and implementing the solution?

If one of those questions cannot be answered, the finding probably needs more work.

Better writing moves the disagreement to the right place

A strong finding does not eliminate disagreement. It moves the disagreement to the right level. Instead of debating whether “AI is risky” or whether a report sounds too severe, management and internal audit can discuss the evidence, residual risk, organizational tolerance, and the most proportionate response.

That is what mature AI assurance should aim for: fewer impressive statements and more traceable logic. Models, vendors, and testing techniques will change. A well-built finding remains useful because it is anchored in decisions, controls, evidence, and accountability.