Practical guide

AI-assisted report QA checklist: free template and guide

An AI-assisted report is ready for use only when its purpose, evidence, analysis, claims, limitations, and approval state can be inspected. Good grammar and a confident executive summary do not establish that the source record is complete or that the conclusion follows from it. This copyable checklist uses REPORT: remit, evidence, procedures, outputs, review, and traceability.

By Two Prune

An AI-assisted report is ready for use only when its purpose, evidence, analysis, claims, limitations, and approval state can be inspected. Good grammar and a confident executive summary do not establish that the source record is complete or that the conclusion follows from it. This copyable checklist uses REPORT: remit, evidence, procedures, outputs, review, and traceability.

Source basis and limits

The UK Government's 2025 AQuA Book says analytical assurance should be proportionate to business risk, continue throughout the analytical lifecycle, distinguish the analyst from the independent assurer, and retain documentation of sources, assumptions, decisions, limitations, and changes. Its delivery guidance asks teams to communicate the assurance state, check whether outputs meet the commission, and preserve supporting documentation. The NIST Generative AI Profile describes information integrity in terms of source linkage, verification, uncertainty, provenance, and the level of vetting. These sources support the controls below. They do not approve a report or replace the receiving organization's specialist standards. AQuA Book NIST AI 600-1

Sources: [1] [2]

Copyable REPORT checklist

### Document control

| Field | Entry | | --- | --- | | Report title, ID, version, and status | | | Commissioner and intended decision | | | Audience and permitted uses | | | Evidence cut-off and reporting period | | | Author, independent reviewer, approver | | | Applicable policy or professional standard | | | Next review or expiry date | |

### R - Remit and decision fit

Decision-fit line: This report supports [decision] for [audience] using evidence through [date]. It does not establish [material non-claim].

### E - Evidence inventory

| Evidence ID | Source owner | Date/version | Population and period | Quality concern | Used for claim | Access route | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | |

### P - Procedures and calculations

| Test | Input | Method or formula | Expected condition | Result | Reviewer evidence | | --- | --- | --- | --- | --- | --- | | | | | | | |

### O - Outputs and claims

Use a claim ledger for statements that could change the decision.

| Claim ID | Report claim | Evidence IDs | Inference required | Confidence and limit | Disposition | | --- | --- | --- | --- | --- | --- | | | | | | | Keep / narrow / correct / remove / escalate |

### R - Review and release

The author should complete self-checks before an independent reviewer repeats the high-risk tests. Keep the roles distinct for consequential analysis.

| Review issue | Severity | Evidence | Required action | Owner | Disposition and date | | --- | --- | --- | --- | --- | --- | | | | | | | |

### T - Traceability and maintenance

  • [ ] State the question the report was commissioned to answer.
  • [ ] Name the decision, audience, scope, exclusions, and time period.
  • [ ] Distinguish observed results from forecasts and recommendations.
  • [ ] Confirm that the report addresses the commission rather than an easier substitute question.
  • [ ] State uses that the evidence cannot support.
  • [ ] Open every material source rather than relying on an AI summary.
  • [ ] Confirm identity, date, version, units, denominator, and coverage.
  • [ ] Mark missing, inaccessible, conflicting, duplicated, or stale evidence.
  • [ ] Check whether citations support the precise proposition beside them.
  • [ ] Record restricted sources without exposing their contents to an unapproved tool.
  • [ ] Reproduce material totals, rates, comparisons, and rankings from source values.
  • [ ] Preserve units, rounding, exclusions, weighting, and missingness rules.
  • [ ] Test at least one plausible alternative assumption or denominator.
  • [ ] Reconcile totals across the narrative, tables, appendices, and source files.
  • [ ] Check that code, spreadsheet formulas, filters, and transformations match the described method.
  • [ ] Label AI-generated classifications or summaries and validate a risk-based sample.
  • [ ] Separate fact, estimate, inference, opinion, recommendation, and unknown.
  • [ ] Make counter-evidence and high-consequence exceptions visible in the main report.
  • [ ] Check that charts use an honest scale, labelled series, readable text, and accessible alternatives.
  • [ ] Confirm that the executive summary preserves the report's material uncertainty.
  • [ ] Remove causal, comparative, or outcome language not supported by the method.
  • [ ] Scale review depth to effect, complexity, longevity, novelty, and reversibility.
  • [ ] Involve a subject specialist for claims outside the author's competence.
  • [ ] Record unresolved disagreement and any release condition.
  • [ ] Confirm privacy, security, legal, accessibility, finance, or communications review where relevant.
  • [ ] Obtain evidence of approval; do not infer approval from silence.
  • [ ] Retain source snapshots, calculation artifacts, prompts or AI outputs needed to reproduce material claims, and review dispositions.
  • [ ] Maintain a version history linking every material change to its reason and evidence.
  • [ ] Define correction, withdrawal, and stakeholder-notification routes.
  • [ ] Set a review or expiry date for recurring reports.
  • [ ] Check that published copies, dashboards, and downstream extracts use the approved version.

Bounded AI workflow

AI can inventory approved source IDs, suggest a report structure, identify internal contradictions, generate test cases, and improve clarity. Ask it to attach source identifiers to extracted claims and to list uncertainty rather than fill gaps. Then reopen original sources, reproduce calculations, inspect transformations, and have accountable people decide what can be released.

Do not let a model invent evidence, citations, customer outcomes, approvals, legal conclusions, or reviewer dispositions. A generated checklist result is not independent assurance.

Worked example: fictional monthly service-performance report

A service team drafts a report saying response performance improved from 82% to 91% after an AI assistant was introduced. The executive summary recommends expanding the tool. The underlying workbook contains monthly ticket counts and response times; the AI draft cites the workbook generally.

REPORT first clarifies the commission: explain current service performance and the exceptions that require an operating decision. The evidence inventory finds that the earlier month covered all channels while the later month omitted an email backlog during a migration. The procedures table reproduces the 91% figure for the reduced population, then recalculates a comparable 84% rate when the backlog is included. It also finds that priority-one cases improved while priority-two cases worsened.

The claim ledger narrows “performance improved” to two supported observations and removes the causal statement because no comparison isolates the assistant's effect. The main summary now exposes the channel exclusion and the priority-two exception. A service analyst checks calculations; an independent reviewer traces the headline numbers; an operations owner decides to repair the email feed and continue a bounded pilot rather than approve expansion.

The release pack retains the source snapshot, formula check, claim dispositions, and approved report. A correction route is ready if the missing backlog changes after reconciliation.

Adaptation guidance

For a short narrative report, keep the six REPORT sections but reduce the number of rows. For a recurring dashboard pack, emphasize data lineage, change control, repeat-run tests, and expiry. For a regulated or high-consequence report, add the applicable professional control mapping and independent specialist review. Always preserve the distinction between assurance evidence and the approval decision.

Version history

| Version | Date | Change | Evidence affected | Author | Reviewer | Approver | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | |

Two Prune boundary

Two Prune can assess whether a participant frames a reporting task, navigates supplied evidence, uses AI thoughtfully, checks claims and calculations, exposes uncertainty, and delivers a useful report in a realistic simulation. It reports observed assessment evidence. It does not approve reports, provide analytical assurance, validate source systems, accept specialist risk, or make employment decisions.

Sources

GOV.UK: The AQuA Book

  • NIST AI 600-1: Generative AI Profile

Sources: [1] [2]

Sources

  1. 1.AQuA Book · UK Government
  2. 2.NIST AI 600-1 · doi.org