Practical guide

AI due diligence report verification checklist

Use this checklist to test whether a due diligence report is decision-ready, not merely polished. It checks the mandate, source traceability, calculations, contradictions, issue treatment, approvals, and residual uncertainty. It can be used on a report drafted with or without AI. A completed checklist is review evidence, not proof that the subject is safe, compliant, complete, or suitable.

By Two Prune

Use this checklist to test whether a due diligence report is decision-ready, not merely polished. It checks the mandate, source traceability, calculations, contradictions, issue treatment, approvals, and residual uncertainty. It can be used on a report drafted with or without AI. A completed checklist is review evidence, not proof that the subject is safe, compliant, complete, or suitable.

Source and scope boundary

The OECD Due Diligence Guidance for Responsible Business Conduct describes a risk-based process that includes identifying and assessing impacts, acting on findings, tracking results, communicating, and enabling remediation where appropriate. The NIST AI Risk Management Framework Core provides voluntary governance outcomes for mapping context, measuring risk, managing it, and recording accountability. They support the structure below but do not replace legal, financial, security, technical, accessibility, procurement, or jurisdiction-specific review. OECD: Due Diligence Guidance for Responsible Business Conduct NIST: AI RMF Core

Sources: [1] [2]

Verify with TRACE

Record pass, repair, not applicable with reason, or hold for every item. Add an evidence ID and reviewer initials. Do not use a percentage score to conceal a failed mandatory item.

### T — Terms of review and decision threshold

Hold if: the decision is undefined, the reviewed subject differs from the proposed use, or an exclusion removes a material dependency without disclosure.

### R — Records, references, and representations

Hold if: an AI citation was not opened, a representation is reported as verified fact, or missing evidence is described as “no issue found.”

### A — Assertions, analysis, and arithmetic

Hold if: a material number cannot be reproduced, incompatible periods are compared, or certainty exceeds the underlying record.

### C — Contradictions, controls, and consequence

Hold if: a contradiction affecting the decision remains hidden, control design is treated as operating effectiveness, or a material dependency has no owner.

### E — Exceptions, escalation, and evidence gate

Hold if: an exception has no expiry, the approver cannot accept the consequence, or the report recommends action beyond the reviewed scope.

  • [ ] The report names the decision it supports and the authorized decision owner.
  • [ ] Scope, excluded areas, entities, products, jurisdictions, populations, time period, and evidence cutoff are explicit.
  • [ ] Materiality and severity criteria were set before findings were ranked.
  • [ ] Required professional reviews and applicable obligations are identified rather than inferred.
  • [ ] The report distinguishes proceed, proceed with conditions, investigate, renegotiate, hold, and stop.
  • [ ] Confidentiality, permitted tools, data handling, retention, and reviewer access are documented.
  • [ ] Each decisive claim links to a stable source, owner, date, version, and covered scope.
  • [ ] Direct evidence, management or vendor representation, estimate, inference, and AI-generated suggestion are visibly different.
  • [ ] Missing, refused, inaccessible, stale, superseded, or restricted records remain in the evidence ledger.
  • [ ] Quotations, dates, names, contract terms, certifications, and citations were checked against originals.
  • [ ] Assurance reports and certifications state their covered entity, service, period, criteria, exceptions, and exclusions.
  • [ ] External authoritative sources are current enough for the stated decision date.
  • [ ] Every conclusion states the evidence and reasoning that support it.
  • [ ] Calculations can be reproduced from named inputs, units, formulas, filters, and rounding rules.
  • [ ] Totals reconcile with component populations; samples reconcile with their sampling frame.
  • [ ] Currency, period, tax, inflation, discounting, and scenario assumptions are explicit where relevant.
  • [ ] Alternative explanations and counterevidence were considered.
  • [ ] Severity follows the pre-defined criteria rather than the preferred commercial outcome.
  • [ ] The executive summary is no stronger than the detailed findings.
  • [ ] Versions, dates, entity names, ownership, data flows, system boundaries, populations, and obligations were compared across sources.
  • [ ] Contradictions are recorded with both sources and a disposition; they are not silently harmonized by AI.
  • [ ] Documented controls are separated from evidence that those controls operated for the reviewed scope and period.
  • [ ] Dependencies, subprocessors, external services, licences, manual reviews, customer duties, and exit arrangements are mapped.
  • [ ] Each issue states condition, criterion, supported cause, consequence, affected scope, current treatment, evidence strength, owner, and due date.
  • [ ] Affected users and foreseeable accessibility, privacy, security, operational, and fairness consequences are routed to qualified owners.
  • [ ] Conditions and exceptions identify authority, exact scope, compensating action, expiry, and stop trigger.
  • [ ] Open issues have accountable owners, closure evidence, and dates.
  • [ ] The recommendation states residual uncertainty and the evidence that would change it.
  • [ ] Dissent and unresolved reviewer comments are preserved.
  • [ ] Approval refers to the exact report version and reviewed subject.
  • [ ] A next review is triggered by time, change, incident, new evidence, or expansion of use.
  • [ ] The final package retains the mandate, source ledger, calculations, contradictions, issue register, approvals, and checklist.

Downloadable checklist record

Copy this compact record into a spreadsheet or document for each item:

| Check ID | TRACE area | Requirement | Disposition | Evidence ID | Repair or exception | Owner | Due date | Reviewer | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | T-01 | Terms | Decision and authority named | pass / repair / N/A / hold | exact record | action and boundary | role | date | initials |

Version the checklist separately from the report. Record checklist version, report hash or version, reviewer, review time, and any changed disposition after repair.

Worked verification: synthetic support-summary vendor report

A synthetic due diligence report recommends a 90-day pilot of a service that summarizes internal support conversations. The report contains 34 references, a risk matrix, a cost estimate, and an executive conclusion that the service is “secure and compliant.”

TRACE does not award a high score for surface completeness. Under Terms, the mandate covers a synthetic-data sandbox, while the recommendation permits real support conversations. That scope mismatch is a hold.

Under Records, the vendor presentation says customer data is not used for training. The contract permits service-improvement use, and the report cites only the presentation. The checklist classifies the statement as a representation, adds the contract, and records the contradiction. A security audit covers the core service but excludes an analytics subprocessor, so the checklist prevents the report from generalizing its assurance.

Under Analysis, the annual cost total uses 10,000 conversations per month, while the operating appendix uses 14,000. Recalculation changes the estimated variable charge by 40 percent. The repair records the population, formula, currency, and sensitivity rather than choosing the lower figure.

Under Controls, the report lists a human review policy but provides no sampled operating evidence. The checklist changes “all summaries are reviewed” to “the procedure requires review; operation was not tested.” It also adds the analytics service, deletion process, user access, incident route, and customer configuration to the dependency map.

Under Exceptions, the team replaces the broad recommendation with hold real-data use; permit only a time-limited synthetic evaluation after named controls are configured. Closure evidence includes amended data-use terms, a current subprocessor record, verified data flow, deletion test, accessibility review, reconciled cost model, and an authorized privacy and security decision.

The completed artifact contains 42 checklist rows: 27 pass, 10 repair, 2 reasoned not-applicable items, and 3 holds. Those counts are workflow indicators only. The three holds control the disposition; they are not averaged away.

Practitioner QA before acceptance

Have a reviewer who did not author the report select the five most decision-critical claims and trace each to its original record. Reperform every material calculation. Sample at least one claim from the executive summary, one control statement, one assurance statement, one cost or benefit figure, and one condition.

Reject the checklist when reviewers tick boxes without evidence IDs, change N/A to pass, average mandatory holds into a score, or verify only grammar and links. Accept it only when another qualified reviewer can reproduce the reasoning and see precisely what remains unknown.

AI may help extract candidate claims, compare versions, or prepare the checklist table. A human reviewer remains responsible for opening sources, checking calculations, resolving contradictions, applying criteria, and routing professional conclusions.

Two Prune provides realistic AI-enabled work assessments and evidence about observed task performance. It can assess whether a participant demonstrates source checking, calculation verification, uncertainty handling, and clear delivery in a bounded simulation. It does not perform due diligence, certify a report, promise an outcome, or make employment decisions.

Sources

OECD: Due Diligence Guidance for Responsible Business Conduct

  • NIST: AI RMF Core

Sources: [1] [2]

Sources

  1. 1.OECD: Due Diligence Guidance for Responsible Business Conduct · www.oecd.org
  2. 2.NIST: AI RMF Core · National Institute of Standards and Technology