Practical guide

How to create a due diligence report with AI

A useful due diligence report turns a bounded question into a traceable evidence record, distinguishes verified facts from representations and unknowns, tests material risks and dependencies, and states what the evidence permits a decision maker to conclude. AI can help organize documents, identify gaps, compare statements, and draft issue summaries; it cannot verify a source it has not.

By Two Prune

A useful due diligence report turns a bounded question into a traceable evidence record, distinguishes verified facts from representations and unknowns, tests material risks and dependencies, and states what the evidence permits a decision maker to conclude. AI can help organize documents, identify gaps, compare statements, and draft issue summaries; it cannot verify a source it has not inspected or replace qualified legal, financial, security, technical, or regulatory judgment.

Define the mandate before reviewing documents

Record the decision, authorized reviewer, subject, scope, excluded areas, materiality criteria, evidence cutoff, time period, jurisdictions, affected groups, and permitted data handling. Name the decision options: proceed, proceed with conditions, investigate, renegotiate, hold, or stop.

Due diligence is context-specific. The OECD Due Diligence Guidance for Responsible Business Conduct describes a risk-based process of embedding responsibility, identifying and assessing impacts, ceasing or mitigating them, tracking results, communicating, and enabling remediation where appropriate. The NIST AI Risk Management Framework Core provides voluntary outcomes for mapping context, measuring risks, managing them, and governing roles and evidence. These sources support disciplined review; they do not establish that a report is legally complete or that a transaction, supplier, or system is safe. OECD: Due Diligence Guidance for Responsible Business Conduct NIST: AI RMF Core

Sources: [1] [2]

Build the report with SCOPE

### S — Scope, decision, and source plan

Translate the mandate into review questions. For a technology supplier, these might cover ownership, financial dependency, security, privacy, data rights, model or service limitations, accessibility, subcontractors, incident history, business continuity, support, and termination.

Create a request list that names the evidence, owner, period, format, and acceptance test. Separate required evidence from useful context. Record unavailable, refused, stale, or restricted items rather than silently narrowing the review.

### C — Chain of evidence and contradictions

Use a source ledger:

| Evidence ID | Claim or question | Source, owner, and date | Direct, representation, estimate, or inference | Verification | Limitation | | --- | --- | --- | --- | --- | --- | | DD-01 | exact issue | stable record | classification | test performed | gap or boundary |

Trace decisive claims to original contracts, policies, technical records, financial statements, audit evidence, incident records, demonstrations, or authoritative sources as appropriate. A vendor answer is a representation until corroborated. A certification can define covered scope but does not prove that every control operated for the reviewed use.

Compare documents for contradictions in version, ownership, dates, populations, terms, system boundaries, and exceptions. Ask AI to list conflicts only from supplied material, then verify every conflict against the source.

### O — Operations, ownership, and obligations

Map the real workflow: who supplies data, configures access, reviews output, handles incidents, approves changes, supports users, manages subcontractors, and executes exit. Link contractual or policy obligations to operational evidence and accountable roles.

Test dependencies such as external services, key personnel, data feeds, model providers, hosting regions, licences, customer responsibilities, and manual controls. Distinguish a documented process from evidence that it operated.

### P — Problems, proportionality, and possible treatments

Describe each issue with condition, criteria, cause where supported, consequence, affected scope, evidence strength, current treatment, owner, and due date. Rank issues using pre-defined materiality and consequence criteria rather than dramatic wording.

Treat uncertainty proportionately. Options include accept with recorded rationale, seek evidence, repair, narrow scope, add a condition, price or contract for the risk, test, monitor, establish a fallback, or stop. A red-amber-green label without the underlying evidence and treatment is not sufficient.

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

Summarize open exceptions, responsible approvers, compensating steps, expiry, and stop triggers. State the decision supported and its conditions. Name the evidence that would change the conclusion and the date for review.

End with residual uncertainty and excluded work. Do not phrase lack of evidence as evidence of absence.

Report structure

The executive conclusion should not be stronger than the detailed findings. Terms such as verified, compliant, secure, independent, or complete require an explicit criterion and scope.

  • Decision and executive conclusion
  • Mandate, scope, exclusions, materiality, and methods
  • Subject and dependency map
  • Evidence received, missing, and quality limitations
  • Findings by review question
  • Contradictions and unresolved representations
  • Issue register and treatments
  • Conditions, exceptions, and approvals
  • Decision options and recommendation
  • Residual uncertainty and next evidence gate
  • Source ledger and technical appendices

Controlled use of AI

Approved AI tools can classify supplied documents, extract candidate terms, compare versions, build timelines, map claims to sources, and challenge whether findings support the recommendation. Retain the document set, prompts or instructions where material, output used, human checks, and rejected suggestions.

Do not upload restricted material to an unapproved service. Verify quotations, numbers, entity names, dates, and citations. AI-generated issue language should never create an allegation or certainty absent from the evidence.

Worked report: AI support vendor

A synthetic organization considers a vendor that summarizes internal support conversations. The commercial presentation says customer data is not used for training, all subprocessors are approved, and the service is available in one region.

The SCOPE mandate covers a 90-day internal pilot, not production approval. The source ledger classifies the presentation as a representation. Contract terms allow service-improvement use, the subprocessor list is undated, and the architecture diagram shows a separate analytics service in another region.

The contradiction review does not conclude misconduct. It records inconsistent statements, asks for current data-flow evidence, and routes interpretation to privacy, security, procurement, and legal owners. An audit report covers the core platform but excludes the analytics service, so its assurance scope cannot close that issue.

Operational mapping finds that users can paste sensitive cases even though the intended pilot excludes them. The interface has no preventive control. The team adds approved synthetic cases, access restrictions, visible instructions, sampling, an incident route, and a stop trigger. A manual reviewer owns every summary before use.

The issue register records three high-priority gaps: unresolved data use, incomplete subprocessor/version evidence, and no verified regional boundary for analytics. The disposition is hold external-data testing and run only a synthetic sandbox review. Conditions require an amended data-use term, dated subprocessor evidence, verified flow, deletion test, accessibility review, and named exit owner.

The report does not claim the vendor is unsafe, compliant, or unsuitable. It states that available evidence is insufficient for the requested pilot scope. The evidence artifact contains the mandate, request list, source ledger, contradiction table, dependency map, audit-scope note, issue register, conditions, approvers, residual uncertainty, and next review date.

Acceptance and hold conditions

Accept the report only when the decision is explicit, scope and exclusions are visible, decisive claims trace to current evidence, missing items and contradictions are preserved, materiality is consistent, treatments have owners, and the conclusion matches the record.

Hold it when AI-generated citations are unverified, representations are described as facts, missing evidence becomes “no issue found,” assurance scope is overstated, materiality changes for a preferred outcome, confidential data left approved controls, or conditions lack authority and expiry.

Two Prune can place due-diligence reporting inside a realistic AI-enabled assessment and evaluate how a participant scopes work, navigates documents, uses AI, tests claims, handles uncertainty, and communicates a defensible report. It does not perform professional due diligence, certify a subject, or make employment decisions.

Sources

OECD: Due Diligence Guidance for Responsible Business Conduct

  • National Institute of Standards and Technology: AI Risk Management Framework 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

Explore this topic