Practical guide
AI executive briefing verification checklist
An executive briefing is ready only when a reviewer can trace its decision, evidence, options, uncertainty, and follow-through without reconstructing the author’s work. This checklist turns that review into a repeatable acceptance test. It is designed for a briefing drafted with or without AI and focuses on whether the issued document is decision-ready, not whether it sounds polished.
An executive briefing is ready only when a reviewer can trace its decision, evidence, options, uncertainty, and follow-through without reconstructing the author’s work. This checklist turns that review into a repeatable acceptance test. It is designed for a briefing drafted with or without AI and focuses on whether the issued document is decision-ready, not whether it sounds polished.
Source and use boundary
The UK Policy Profession Standards describe strong advice as clear, accurate, well-reasoned, evidence-based, impartial, and targeted to its audience. They also describe briefing as selecting and presenting information so another person can understand a subject quickly. GOV.UK publishing guidance adds that useful writing should be specific, clear, concise, and written for the reader. These are authoritative public-practice sources for review criteria, but they do not determine whether a particular recommendation is correct. UK Government: Policy Profession Standards annex GOV.UK: Use the right tone
Sources: [1] [2]
Run the DECIDE check
Score every item pass, repair, not applicable with reason, or hold. A hold is a release blocker. A repair needs an owner and completion time. Do not replace missing evidence with an average score.
### D — Decision and reader
Hold when: no decision can be identified, the named reader lacks authority, or the document disguises a consequential choice as “for awareness.”
### E — Evidence and calculations
Hold when: a recommendation depends on an untraceable claim, a material number cannot be reproduced, or conflicting evidence has been silently combined.
### C — Choices and consequences
Hold when: only the preferred option is developed, the consequences are asymmetric, or a specialist judgment is presented as settled without review.
### I — Integrity and uncertainty
Hold when: uncertainty is hidden, evidence strength is overstated, or a late edit changes the decision without renewed clearance.
### D — Delivery, controls, and dissent
Hold when: the recommendation cannot be implemented, no one owns the next action, or the document has no response to an adverse signal.
### E — Evidence pack and issue record
Hold when: the issued version cannot be reconstructed or the briefing differs from the evidence pack that reviewers cleared.
- [ ] The first paragraph names the decision, decision owner, and deadline.
- [ ] The document says whether it seeks a choice, approval, escalation, update, or preparation for discussion.
- [ ] The recommendation is expressed as an action, not a topic.
- [ ] The briefing distinguishes the accountable decision maker from advisers, reviewers, and delivery owners.
- [ ] The language, detail, and reading time fit the named audience.
- [ ] The executive page still makes sense if separated from every annex.
- [ ] Every material claim points to a stable source, owner, date, and version.
- [ ] Facts, estimates, representations, assumptions, and inferences are visibly different.
- [ ] Decisive calculations have been reperformed from the cited inputs.
- [ ] Units, populations, time periods, denominators, and definitions are consistent.
- [ ] Conflicting sources are shown and resolved, or left explicitly unresolved.
- [ ] Quotations were checked against the original source and surrounding context.
- [ ] AI-generated summaries were compared with the underlying records.
- [ ] At least one credible current-state or no-change option is included.
- [ ] Alternatives differ in a way that matters to the decision.
- [ ] Each option shows benefit, cost, delivery burden, dependencies, affected parties, and material risk.
- [ ] Forecasts state their assumptions and sensitivity.
- [ ] The briefing says what evidence would change the recommendation.
- [ ] Legal, security, privacy, accessibility, financial, or regulatory judgments are routed to qualified owners.
- [ ] The strongest counter-evidence and credible objection are represented fairly.
- [ ] Missing data, weak coverage, stale records, and measurement limits are named.
- [ ] The conclusion does not extend beyond the population and period examined.
- [ ] Confidence language matches evidence strength.
- [ ] Material edits have not removed qualifications or changed meaning.
- [ ] The briefing states the human judgment involved in source selection and recommendation.
- [ ] The chosen action has an owner, due date, resources, dependencies, and acceptance condition.
- [ ] There is a fallback or stop trigger for material downside.
- [ ] Open questions and dissent are recorded rather than erased.
- [ ] Communications and affected-user actions are assigned.
- [ ] The review point specifies what will be measured and who will decide next.
- [ ] The source list, claim ledger, calculation sheet, option analysis, and annexes use the same version.
- [ ] Clearances and specialist reviews are recorded.
- [ ] The final file has an issue date, version, author, approver, and distribution classification.
- [ ] Material AI assistance is recorded according to the organization’s policy.
- [ ] The decision, conditions, and later outcome can be linked back to the issued briefing.
Compact review record
Use this table as the downloadable review artifact:
| Check ID | Result | Evidence inspected | Defect or limitation | Owner | Due | Retest | | --- | --- | --- | --- | --- | --- | --- | | D-01 | pass / repair / N/A / hold | page and source IDs | exact observation | named person | time | reviewer and time |
The reviewer then records one disposition:
- Accept: every applicable check passes and no hold remains.
- Accept with conditions: only bounded repairs remain, with owners and a deadline before use.
- Return for repair: the decision is identifiable, but one or more material checks fail.
- Hold: evidence integrity, authority, or implementation is too uncertain for issue.
Worked review: synthetic service-recovery briefing
A service owner receives a two-page briefing recommending restricted operation after five access requests were misrouted by an AI-supported workflow. The reviewer examines the executive page, a 214-request population extract, configuration history, incident records, an option table, and the clearance log.
The decision and reader checks pass: the owner must choose by 15:00 among continued operation, restriction, and suspension. The recommendation names manual triage for access and payroll as a condition.
The evidence check initially fails. The briefing says the error rate “doubled,” but the earlier period used a different request taxonomy. Recalculation cannot produce a comparable rate. The reviewer marks E-04 repair, removes the trend statement, and retains the observable count of five misroutes.
The choices check finds another defect. Suspension has a staffing estimate, but the estimate has no source. The author links the queue model, records the eight-hour sampling window, and changes the point figure to a range. The limitation remains visible.
Integrity review confirms that three events followed an outdated taxonomy, one lacked identity context, and one involved a reviewer override. The final text does not attribute every event to the model. Delivery review confirms a configuration freeze, daily sample, named incident owner, and stop trigger for another access misroute.
The evidence pack matches version 1.3 of the briefing. A specialist has cleared the access-control statement, and the service owner’s decision will be recorded after the meeting. The reviewer issues accept with conditions: update the distribution label and attach the recalculation record before circulation. Those repairs are checked at 14:20.
This worked example demonstrates why the checklist is not a style score. A concise, persuasive document can still fail because its comparison population changed, an estimate lacks lineage, or causes were collapsed.
Using AI during verification
An approved AI tool may help map claims to supplied source IDs, compare versions, locate inconsistent numbers, or test whether a decision is visible. The reviewer must open the cited records, reperform decisive calculations, inspect context, and decide the disposition. Do not let an AI tool turn missing evidence into a pass or resolve ownership and risk acceptance on behalf of accountable people.
Two Prune assesses observed performance in realistic AI-enabled work. It can provide evidence about how a participant frames a briefing task, navigates sources, uses AI, verifies claims, applies judgment, and communicates. It does not approve executive briefings, establish workplace impact, or make employment decisions.
Sources
UK Government: Policy Profession Standards annex
- GOV.UK: Use the right tone
Sources: [1] [2]
Sources
- 1.UK Government: Policy Profession Standards annex · UK Government
- 2.GOV.UK: Use the right tone · guidance.publishing.service.gov.uk