Practical guide
AI control assessment verification checklist
Use this checklist after a control assessment has been drafted and before its conclusion is used for remediation, risk acceptance, or executive reporting. It tests whether the record is traceable, reproducible, and limited to what the evidence supports. A polished narrative is not enough: the reviewer should be able to reconstruct the population, criteria, procedure, exceptions, and authority.
Use this checklist after a control assessment has been drafted and before its conclusion is used for remediation, risk acceptance, or executive reporting. It tests whether the record is traceable, reproducible, and limited to what the evidence supports. A polished narrative is not enough: the reviewer should be able to reconstruct the population, criteria, procedure, exceptions, and authority behind the conclusion. This is an educational verification aid. It does not create an audit opinion or replace the standards, qualifications, and independence requirements that apply to a real engagement.
Freeze the verification packet
Record these identifiers before reviewing the assessment:
If a source changes during verification, create a new packet version. Do not silently replace the population or conclusion in place.
The U.S. Government Accountability Office’s 2025 Green Book distinguishes control design, implementation, and operating effectiveness and emphasizes quality information. Those concepts support separate checklist tests for the control description, evidence, and operation, even when the Green Book is not the organization’s governing standard. GAO Green Book
- control statement and version;
- objective, risk, owner, operator, frequency, and system;
- assessment period and evidence cutoff;
- source-population extract, row count, control total, and checksum where available;
- test procedure, sample basis, and assessment criteria;
- AI service and material configuration when an AI tool contributed;
- preparer, reviewer, exception authority, and final decision maker;
- assessment file, exception register, action plan, and retest record versions.
Sources: [1]
Run the VALID checklist
### V — Verify the population and source trail
Verification failure: the assessment reports a sample result but cannot show how the sample came from a complete population.
### A — Assess criteria, calculations, and attributes
Verification failure: the conclusion changes the threshold after exceptions appear or relies on an unstated “reasonable” standard.
### L — Link exceptions to consequence and scope
Verification failure: “resolved” combines a corrected transaction, an accepted risk, and a promised future action.
### I — Inspect independence, authority, and AI contribution
The Institute of Internal Auditors’ Global Internal Audit Standards address engagement planning, evidence, findings, conclusions, action plans, communication, and monitoring. They provide a current professional reference for this review discipline; the applicable engagement rules must still be determined by qualified people. IIA Global Internal Audit Standards
### D — Document disposition, remediation, and retest
Verification failure: a corrected configuration is treated as proof that the control operated throughout the earlier assessment period.
- [ ] The population covers the stated period, system, business unit, threshold, and transaction type.
- [ ] The extract reconciles to an independent total or the assessment explains why it cannot.
- [ ] Exclusions, filters, joins, deduplication, and late-arriving records are documented.
- [ ] Every tested item has a stable source anchor.
- [ ] Screenshots and exported files preserve the source version and surrounding context.
- [ ] Missing, unreadable, or contradictory evidence is recorded as a limitation rather than inferred.
- [ ] AI-extracted fields were checked against the rendered source for material items.
- [ ] The expected control behavior is stated before the observed results.
- [ ] Each test attribute maps to a policy, procedure, configuration, or approved criterion.
- [ ] Thresholds, dates, currencies, percentages, and formulas were independently recalculated.
- [ ] Identity, authority, segregation, timing, sequence, and version attributes are tested where relevant.
- [ ] The procedure distinguishes design, implementation, and observed operation.
- [ ] A similar-looking record is not treated as evidence for the tested item.
- [ ] Any AI-proposed criterion is confirmed by an accountable human against the actual rule.
- [ ] Results use explicit states such as pass, exception, unable to test, and not applicable.
- [ ] An evidence gap is distinguishable from observed control non-operation.
- [ ] Each exception names the failed criterion, affected item, source evidence, and consequence.
- [ ] The reviewer considered whether the exception indicates a wider population issue.
- [ ] Similar exceptions were clustered only after item-level results were preserved.
- [ ] Root-cause statements are labelled as confirmed, supported, or still a hypothesis.
- [ ] Compensating controls have their own evidence and test, not just an owner assertion.
- [ ] The control operator, assessment preparer, reviewer, and risk authority are identified.
- [ ] Any concentration of those roles has an explicit independence decision.
- [ ] The person accepting an exception has authority for its scope, value, and duration.
- [ ] AI source facts, transformations, suggestions, and human judgments remain distinguishable.
- [ ] The reviewer can reproduce material AI-assisted comparisons without relying on the model’s prose.
- [ ] Confidential information was processed only in the approved environment and retained according to policy.
- [ ] A reviewer challenge is documented for high-consequence or contradictory results.
- [ ] The conclusion is limited to the tested entity, period, population, and procedure.
- [ ] Design, implementation, and operation conclusions are stated separately when applicable.
- [ ] Exceptions have an owner, due date, action, interim protection, and approver.
- [ ] Risk acceptance states the authority, rationale, scope, expiry, and monitoring condition.
- [ ] Remediation is not closed merely because an action plan exists.
- [ ] A retest records the new evidence, procedure, reviewer, date, result, and remaining uncertainty.
- [ ] The final report links to the frozen packet, exception register, and disposition record.
Sources: [2]
Copyable verification record
Use one row for every material test result or exception.
| ID | Population / sample anchor | Criterion and calculation | Observed evidence | Result | Consequence and scope | AI contribution | Human challenge | Disposition / authority | Retest and uncertainty | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | CA-01 | extract row / source event | expected attribute and independent calculation | exact record and timestamp | pass / exception / unable / N/A | affected process and possible expansion | extract / compare / draft | check performed | action, owner, approver | date, result, what remains |
Attach the source manifest, reconciliation, selection logic, calculation sheet, prompt or reproducible instruction summary, exception register, approvals, and retest evidence. A dashboard status without those anchors is a summary, not the assessment record.
Worked verification: privileged-access review
A quarterly assessment states that every privileged account was reviewed by the system owner. The source extract lists 146 accounts, and a review workbook contains 145 rows. The draft concludes that the control operated because every workbook row has an approval date.
VALID changes the result. The population check finds one service account omitted during a spreadsheet join. The source trail also shows that nine approvals occurred after the deadline. The assessment criterion requires review by quarter-end, so those nine are exceptions even though approval later occurred. One reviewer approved their own account; the authority check routes that item to a different approver.
The verified record therefore reports 146 in scope, 136 timely approvals with appropriate authority, nine late reviews, and one omitted account. The missing row is both a population-process defect and an item needing review. The owner repairs the join, adds a population-count reconciliation, and completes the outstanding review. A retest confirms the repaired logic on the next extract, while the report retains the earlier exceptions instead of rewriting history.
Sampling and escalation rule
If a checklist failure affects the population, selection logic, criterion, or reviewer authority, stop and expand the review before relying on the conclusion. If it affects one isolated item and the population remains sound, record the exception and apply the pre-defined expansion rule. Never reduce the sample or change the rule merely to obtain a cleaner result.
Two Prune can assess how a participant frames a control question, navigates evidence, uses AI, checks calculations, handles exceptions, and communicates a verification record in a realistic work simulation. It does not execute the customer’s control assessment, decide risk acceptance, or replace the qualified reviewers responsible for the real process.
Sources
- 1.GAO Green Book · U.S. Government Accountability Office
- 2.IIA Global Internal Audit Standards · www.theiia.org