Board decisions from audit evidence
AI Audit for Boards
Use AI audit evidence to approve, restrict, defer or request further work while preserving material limitations and accountability.
A board should use an AI audit to understand what evidence supports a material decision, what remains uncertain and who is accountable for the next action. Directors do not need to become model engineers to challenge an unsupported conclusion. They do need a report that distinguishes tested controls from promises, management risk acceptance from verified remediation and a narrow technical finding from a broad claim of organisational compliance.
This guide is for directors, audit committees and executives preparing board papers. It concerns the use of audit evidence in oversight, not investment advice or a statement that every AI decision requires full-board approval. The allocation of decisions depends on the organisation's governance, delegated authorities and applicable requirements. The examples below are hypothetical decision tools to adapt with the relevant professional advisers.
Ask what decision the audit was designed to support
Begin with the decision, not the headline rating. Was the audit commissioned to support a limited pilot, an expansion of automated authority, supplier selection or continued operation after an incident? A report that answers one question should not be used to approve another without examining the gap. “Audited AI” is too broad a basis for a material decision.
The IIA's Three Lines Model distinguishes governing-body oversight, management responsibilities and independent internal audit. That is useful context for keeping roles clear. An auditor's findings inform oversight; they do not automatically transfer management's responsibility for the system or the board's role in the decisions reserved to it.
Ask management to state the proposed use, affected people and processes, permitted actions and consequence if the system is wrong. Then ask which claims the audit examined. If the requested approval goes beyond those claims, the board can require more evidence, narrow the proposed use or record a decision to proceed with clearly identified uncertainty where it is authorised to do so.
Example: AI-assisted franchise-expansion recommendations
Consider a hypothetical group using AI to help prepare franchise-expansion papers. The application summarises source information about candidate locations and proposes factors for analysts to examine. It does not approve capital expenditure, sign leases or commit the company to a location. Existing human approval processes remain responsible for those decisions.
Management asks permission to expand the assistant's use in preparing capital proposals. The audit examines whether source references are traceable, whether assumptions are visible to analysts and whether the application stays within its advisory boundary. It does not validate the commercial attractiveness of a particular investment or guarantee that the underlying market information is complete.
The audit finds that some selected recommendation statements cannot be reproduced from the retained source versions. It also confirms the tested application cannot submit a capital approval. The board therefore has two different pieces of evidence: a working authority boundary in the tested conditions and an unresolved evidence-quality issue in the decision support. Combining them into a single green or red status would obscure the actual choice.
The first board question: what exactly was examined?
Request the system and workflow boundary, versions, review period, evidence access, languages and important exclusions. Ask whether the work examined design, controlled tests or actual operation. A design assessment can be valuable before launch, but cannot establish how a control operated during a period that has not occurred. A supplier-controlled system may also have inaccessible components that limit the conclusion.
Directors do not need every technical identifier in the meeting paper, but the material boundary must be understandable. In the franchise example, a review of draft preparation does not establish the accuracy of every financial assumption in a capital proposal. A test of the current application does not automatically cover a planned integration that will submit recommendations into an approval system.
Use tested versus not-tested coverage to prevent scope from disappearing in summary. The report should name exclusions that could change the decision, not merely list them in a long appendix. The board can then ask whether the missing area is acceptable for the proposed use or needs additional work.
The second question: what does the evidence establish?
Ask the presenter to connect the material conclusion to the procedures and observations behind it. A policy says what should happen; a configuration shows part of implementation; a controlled test shows behaviour in specified conditions; operating records can support a conclusion about selected events and a period. These forms of evidence should not be treated as interchangeable.
For the recommendation assistant, a source link displayed in the interface does not establish that the source supports the statement. A retained source version and a reproducible comparison may provide stronger evidence. Likewise, an approval button does not establish that the application can commit expenditure; the relevant question is the authority and confirmed effect at the destination.
Ask how uncertain and contradictory cases were handled. Were unresolved outcomes treated as successful? Were difficult cases excluded? Were results from targeted tests presented as a population rate? A board does not need to recalculate every metric, but should expect the reviewer to explain the denominator, selection and limits of any number central to the decision.
The third question: which finding changes our decision?
Severity should reflect consequence, exposure and uncertainty in the actual use. A technically impressive weakness may have limited impact in a tightly restricted pilot, while a mundane access or record failure can be material at enterprise scale. Ask management and the reviewer to explain the decision implication rather than relying only on a colour or severity label.
In the hypothetical case, unreproducible source support may prevent reliance on the assistant's analysis without additional human verification. The tested restriction on capital submission reduces a different risk. The board might permit a bounded drafting use while requiring source verification and retaining existing approval authority. It should not infer that a working authority boundary makes the advisory content reliable.
| Audit evidence | Decision implication | Further condition |
|---|---|---|
| Selected statements lack reproducible source support | Do not rely on those statements as verified decision evidence | Require source-version traceability and appropriate retesting |
| Tested application cannot submit capital approvals | Advisory use has a tested authority restriction | Reassess before adding any action-capable integration |
| Underlying market data completeness was not examined | No audit conclusion on the completeness of investment inputs | Use the existing qualified business review for those inputs |
| Supplier internals remain partly inaccessible | Some claims cannot be established from available evidence | Decide whether the proposed use is acceptable within that limit |
The table is an illustrative governance aid. It does not recommend a franchise investment or prescribe the board's legal duties. Its purpose is to show that different evidence can support different restrictions and conditions within the same decision.
The fourth question: who owns the risk and the correction?
Identify the executive responsible for the use and the people able to change the relevant conditions. Technology may repair traceability, the supplier may provide missing records and the business team may verify assumptions. An audit finding assigned vaguely to “AI” or “IT” can remain unresolved because no one owns the full workflow.
Separate management acceptance of residual risk from technical closure. Management may decide to continue a restricted use while a finding remains open. That decision should identify the authority, rationale, controls, time or trigger for review and the evidence still required. It should not change the finding's status to verified simply because someone accepted the risk.
Ask whether remediation depends on another programme, supplier release or unavailable dataset. A date without a credible dependency plan can create false confidence. The remediation planning guide helps translate findings into sequenced actions, while the board should focus on material obstacles and decisions within its remit.
The fifth question: who made the judgement, and can we rely on it?
Ask about the reviewer's competence, independence and access. Did the firm build the system or design the controls it now evaluates? Does it expect significant remediation revenue? Who reviewed material conclusions? These questions do not imply that every prior relationship invalidates the work, but they determine how the role and limitations should be understood.
Open for AI offers development and delivery services as well as audits, so independence must be assessed for the particular engagement. A useful technical assessment can have a disclosed role limitation; it should not be represented as an independent regulated opinion without basis. Independence and conflicts explains the questions a sponsor should resolve before the board relies on the report.
Also clarify the report's professional form. A scoped technical audit, formal professional assurance and management-system certification are different deliverables. If an external recipient requires a particular form, confirm suitability with the appropriate practitioner. Audit versus assurance helps prevent a board paper from giving a technical report a status it was never commissioned to provide.
Build the board paper around a decision
A useful paper begins with the proposed decision and why it is needed now. It then states the permitted use, material evidence, findings, limitations, management response and options. Keep the technical report available for challenge, but do not make directors infer the decision from a long list of model metrics.
Present genuine options with consequences. “Approve or lose the benefits of AI” is not a balanced decision frame. Options may include a narrower pilot, delayed integration, additional evidence, supplier restrictions or a phased release with explicit conditions. Explain which uncertainty each option retains and who will monitor it.
For the franchise assistant, an option could permit analysts to use drafts while requiring independent source checks before any capital paper is submitted through the existing process. Another could defer wider use until traceability is corrected. The audit informs these choices; management must explain operational feasibility and the responsible authority must make the decision.
A decision record that preserves conditions
| Field | Example entry |
|---|---|
| Decision requested | Permit a bounded advisory drafting use for franchise-expansion papers |
| Evidence considered | Named audit report and version, coverage statement, material finding and management response |
| Conditions | Human source verification; no automated capital submission; specified users and workflow |
| Unresolved limitation | No conclusion on complete market data or partly inaccessible supplier internals |
| Accountable executive | Named owner of the use and remediation programme |
| Return trigger | Material authority change, new evidence, incident or failure to meet the agreed remediation condition |
| Closure evidence | Retest of source traceability and review of any changed integration boundary |
This is not a legally complete minute template. It illustrates the information needed to preserve the meaning of a conditional decision. If conditions are recorded only in the spoken discussion, a later team may interpret “approved” as unrestricted permission. The decision record should distinguish the allowed use from changes that require further authority.
Challenge aggregate portfolio reports
A statement that most AI systems have been audited can hide a material exception in the remaining systems or an unresolved finding in a completed review. Ask what the denominator includes, whether informal uses are still being discovered and what “audited” means. Completion of work is not the same as absence of significant findings.
Shared dependencies can create concentration. A central identity or model gateway issue may affect several applications, but its impact depends on their actual configuration and use. Conversely, a shared control report does not automatically establish every local workflow. The enterprise audit guide explains how portfolio reporting can reuse evidence while retaining application-specific questions.
Do not average away a material issue. A high score for several low-consequence controls cannot compensate for an unresolved authority boundary in a system that can make consequential changes. Ask for the decisions and restrictions attached to the exceptions, not only a trend line showing increasing completion.
Revisit reliance when the system or use changes
An audit conclusion has a baseline. Changes to the model, data, users, workflow or authority can affect it. Management should identify which changes require review and who decides whether existing evidence remains relevant. A report from the previous release should not become a permanent approval badge for every later feature.
In the hypothetical example, connecting the assistant to a capital-approval interface would change the authority question even if the generated text remained similar. Expanding to new languages or source types could change the performance question. The board should expect a change-impact explanation and the necessary additional evidence before relying on the earlier conclusion for a materially different use.
For material fixes, ask for verified retesting rather than a statement that a ticket is closed. The retest should address the original condition and relevant neighbouring behaviour, with changed versions and remaining limitations visible. A board can then distinguish implementation progress from evidence that the required control now works.
Stress-test the conditions attached to a decision
A conditional decision is useful only when the condition can be recognised and acted on. Ask management what would happen if the source data changed materially, the service introduced a new model or a user sought to give the assistant more authority. The answer should identify which evidence may no longer apply, who assesses the change and what happens to use while the question is resolved.
For the franchise-expansion example, the board might permit a bounded advisory pilot using specified information and an existing human approval process. If the proposal later allows the assistant to rank applicants using additional personal data or initiate commitments, that is a changed decision boundary. It should not be described as a routine continuation simply because the interface looks the same. The relevant decision authority needs a clear route to reconsider the proposal.
Ask who can pause the affected activity without waiting for the next scheduled board meeting. The actual delegation depends on the organisation's governance and applicable requirements; an audit report does not create it automatically. Management should explain the temporary action, escalation route and evidence needed for restart. A condition that nobody has authority to enforce is weak even when it sounds precise in the minutes.
Keep evidence confidence separate from expected business value
A promising benefit estimate does not make weak evidence stronger. Equally, a technically well-supported control does not prove that the proposed business investment is worthwhile. Directors should see the business case and audit conclusions as related inputs to a decision, with their different assumptions visible. An audit of source fidelity in an advisory system does not validate a forecast of successful franchise expansion.
Where management presents several options, show how the unresolved findings affect each option. A limited pilot may require different controls and evidence from broad deployment. Deferring one feature may allow a narrower use to proceed, but that should be a reasoned management proposal within the board's authority. The auditor should explain the evidence implications without being asked to choose the commercial strategy.
Hear contrary evidence without blurring accountability
Allow management and the reviewer to state a material disagreement in precise terms. Ask whether they dispute a fact, a method, the likely consequence or the proposed response. A disagreement about the business response can remain even when both accept the same finding. The minutes should preserve the distinction and identify the decision taken, rather than recording artificial consensus.
Before closing the agenda item, ask what information would cause the board or its delegated authority to revisit the decision. Record the owner and trigger alongside the action date. This makes the report a usable part of governance while preserving that directors, management and reviewers have different responsibilities and that statutory duties require advice appropriate to the organisation.
Questions that improve the next meeting
Ask the presenter: What decision does this evidence support? What was not examined that could change that decision? Which finding requires action now? What is management accepting, and under whose authority? What new evidence will return to us? These questions make the report useful without requiring directors to conduct the technical audit themselves.
For an audit scoping discussion for a board or audit committee decision, bring the proposed use, the approval being sought and the material concerns. Open for AI's AI auditing services can be scoped around the evidence needed for that decision. The report should make it possible to approve, restrict, defer or request further work with a clear understanding of what is known and what remains uncertain.
FAQ
Your questions, answered.
What should a board ask when presented with an AI audit report?
Ask which decision the work supports, what was examined, what the evidence establishes, which material limits remain and who owns the next action. Clarify reviewer competence and independence. The board should not treat a headline rating as approval for a use outside the tested scope or assume that a technical report provides certification or a formal professional opinion. Ask the presenter to connect one material conclusion to evidence and the proposed decision.
Does an AI audit transfer responsibility from management to the auditor?
No. The auditor provides findings and conclusions within the engagement scope. Management remains responsible for the system and its use, while governance bodies retain the decisions assigned to them. The precise allocation depends on the organisation and applicable requirements. A decision record should identify the accountable executive, restrictions and evidence needed for follow-up. Ask who has authority to enforce any condition attached to the decision between scheduled meetings.
Can a board approve AI use with an open audit finding?
A responsible authority may consider a bounded use with explicit conditions and residual uncertainty, where that decision is within its remit. It should understand the consequence, available controls and evidence limits, and record the rationale and review trigger. Risk acceptance is not verified remediation. Some findings may make the proposed use inappropriate until further work is completed. Keep the open condition visible and specify what would end, extend or revoke the bounded permission.
What does a clean AI audit rating actually mean?
Its meaning depends on the criteria, procedures, population, version, period and report form. It should not be read as a guarantee of future performance or complete compliance. Ask for material exclusions and how uncertain results were treated. A favourable conclusion on selected controls may coexist with unexamined claims that matter to the proposed decision. A favourable extract should not circulate without the conditions that make the conclusion valid.
How should directors assess an auditor who also built the AI system?
Request an engagement-specific assessment of self-review and other conflicts, including firm-level incentives and planned remediation work. Ask what safeguards apply and who owns the independent judgement where one is required. Prior knowledge can be useful without establishing independence. The report should describe the actual role and limitations rather than relying on the provider’s external status. If the intended independence requirement cannot be met, change the arrangement or the deliverable.
What should a board minute record about an AI decision?
Record the decision, evidence considered, permitted use, material limitations, conditions, accountable owner and triggers for return. Include the relevant report or version reference. The example in this guide is not a legally complete minute template; adapt it through the organisation’s governance process. The purpose is to prevent a conditional approval from later being treated as unrestricted permission. Distinguish approval of the proposed use from acceptance of the report as an information item.
Does auditing most enterprise AI systems establish portfolio assurance?
Not automatically. Completion counts can hide unresolved findings, unexamined high-consequence uses and an incomplete inventory. Ask what the denominator includes and which decisions remain restricted. Shared-control evidence also requires application-specific relevance. A high average score should not offset a material unresolved issue in a system whose authority or data exposure is consequential. Also ask whether one shared defect affects multiple systems despite appearing as a single finding.
When should an AI audit issue return to the board?
Use the organisation’s delegated authorities and agreed triggers. Material changes to use or authority, significant incidents, new evidence, failed remediation conditions or a decision beyond management’s remit may require return. Do not rely on an invented universal schedule. The original decision should state who monitors the condition and what evidence is needed before reliance can be extended. Provide a change-focused paper showing what has altered since the original decision and why it matters.