Practical guide
AI decision memo verification checklist
A decision memo is ready only when an independent reviewer can identify the exact choice, trace material claims, reproduce calculations, compare genuine alternatives, understand objections, and see who owns approval and follow-through. Polished prose is not enough if the recommendation depends on an invented citation, unequal option treatment, hidden assumption, or authority the reader does not.
A decision memo is ready only when an independent reviewer can identify the exact choice, trace material claims, reproduce calculations, compare genuine alternatives, understand objections, and see who owns approval and follow-through. Polished prose is not enough if the recommendation depends on an invented citation, unequal option treatment, hidden assumption, or authority the reader does not hold.
Freeze the review packet
Before testing the memo, record:
If the question, evidence, or recommendation changes during review, create a new version. Do not attach a current recommendation to evidence from an earlier decision frame.
HM Treasury’s Green Book requires options, costs, benefits, risks, and uncertainty to be considered rather than using appraisal to justify a predetermined answer. U.S. OMB Circular A-4 similarly emphasizes defining the need, analyzing alternatives against a baseline, and presenting uncertainty. DECIDE converts that decision discipline into an executable memo review. Passing it does not establish that a recommendation is suitable for every organization or jurisdiction. HM Treasury: The Green Book Office of Management and Budget: Circular A-4
- [ ] memo title, version, author, reviewer, decision owner, and decision date
- [ ] decision question, scope, exclusions, constraints, and consequence of no action
- [ ] evidence cutoff and source register
- [ ] options considered and comparison criteria
- [ ] calculations, assumptions, and sensitivity files
- [ ] recommendation, conditions, dissent, and next evidence gate
- [ ] approved location for restricted inputs and review records
Sources: [1] [2]
Run the DECIDE checks
### D — Decision, authority, and boundary
### E — Evidence and claim traceability
Select three decisive claims and trace them backward from recommendation to sentence, calculation, source, and source version. A broken link is a repair, even when the statement sounds plausible.
### C — Comparison, calculations, and contradictions
### I — Implementation, conditions, and reversibility
### D — Dissent, distributional effects, and dependencies
### E — Editorial release and residual uncertainty
- [ ] The opening sentence states one decision that can be approved, limited, deferred, or rejected.
- [ ] The named reader has authority for that decision, or the required approver is explicit.
- [ ] Scope, exclusions, deadline, affected groups, constraints, and no-action consequence are visible.
- [ ] Terms that could change the choice—such as pilot, production, approved, material, or safe—are defined.
- [ ] The memo does not present background as if it were the decision.
- [ ] Every material factual claim links to a stable source, date, owner, and relevant excerpt or calculation.
- [ ] Facts, estimates, assumptions, inferences, vendor statements, and opinions are labelled distinctly.
- [ ] Sources are current enough for the decision and authoritative for the claim made.
- [ ] Quotations and citations introduced with AI assistance are checked against the original source.
- [ ] Missing or conflicting evidence remains visible rather than being resolved by confident prose.
- [ ] The memo includes a credible status quo and materially different alternatives.
- [ ] Every option is tested against the same criteria, baseline, time horizon, population, and units.
- [ ] Mandatory legal, security, privacy, accessibility, safety, or operational gates are not hidden inside a weighted score.
- [ ] Headline calculations expose numerator, denominator, assumptions, range, and treatment of missing data.
- [ ] At least one material calculation is independently reproduced.
- [ ] Sensitivity or scenario analysis is included where a plausible change could reverse the recommendation.
- [ ] Contradictory sources and reviewer disagreements have a recorded disposition.
- [ ] The recommendation names the first action, owner, deadline, dependencies, resources, and approval conditions.
- [ ] The implementation path matches the option that was analyzed.
- [ ] Stop, rollback, or reconsider triggers are measurable.
- [ ] Downstream communications, access, data, procurement, and operating responsibilities are assigned.
- [ ] A pilot is not described as reversible when data transfer, commitments, or external effects cannot be undone.
- [ ] The strongest reasonable objection appears in its best form.
- [ ] Material dissent is attributed to a role or evidence source rather than hidden as “stakeholder concern.”
- [ ] Benefits, burdens, exclusions, and risks to affected groups are considered where relevant.
- [ ] Upstream and downstream dependencies have owners and evidence.
- [ ] An alternative explanation is recorded for each decisive observed pattern.
- [ ] Title, summary, option table, recommendation, and appendix agree.
- [ ] Tables and charts use clear units, populations, dates, and definitions.
- [ ] The memo distinguishes what is known, unknown, disputed, and time-sensitive.
- [ ] The next evidence gate names the question, method, owner, threshold, and review date.
- [ ] The final file, attachments, approval record, and source register share a version identifier.
- [ ] Residual uncertainty is visible to the decision owner.
Copyable verification record
| Check | Memo element | Evidence or rule | Test | Expected | Observed | Disposition | Owner | Residual uncertainty | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | DM-01 | claim, option, calculation, condition, or approval | exact source/version or criterion | trace, recalculate, compare, challenge | measurable result | retained evidence | pass / repair / hold / remove | named role | remaining question |
Attach the decision brief, claim register, source excerpts, option matrix, calculation workbook, sensitivity results, challenge notes, dissent record, approvals, and release copy. A screenshot of a source or a chat transcript does not replace a stable source and verified claim.
Worked verification: AI meeting-assistant memo
A synthetic memo recommends enabling an AI meeting assistant for all customer and internal calls. Its summary cites a vendor claim that the tool “saves 30% of meeting time” and says adoption is low risk because users can edit notes.
The Decision check fails because the operations director cannot approve customer-call recording or organization-wide data handling. The memo is reframed as a six-week internal-project-meeting pilot requiring security, records, accessibility, and privacy dispositions.
Evidence tracing finds that the 30% figure is a marketing statement with no comparable local workflow evidence. It is removed from the forecast. Internal observation shows that eight volunteers spent 12–25 minutes preparing notes for one week; the memo labels the sample descriptive and incomplete.
The Comparison check finds that broad rollout is scored on “potential efficiency” while the status quo is scored on observed cost. The reviewer applies common criteria to no change, a better note template, a bounded pilot, and broad enablement. Broad enablement fails two mandatory gates: approved data handling and reliable human review before distribution.
The calculation reviewer reproduces the time range and finds that the original average excluded two meetings with no notes. The missing cases are restored, and the memo uses a range instead of a point forecast. A sensitivity shows that correction time could erase any apparent saving.
The Implementation check requires internal meetings only, visible notice, no sensitive cases, named administration, human approval before sharing, deletion rules, and a rollback route. The dissent record preserves an accessibility concern and assigns a test rather than declaring it resolved.
The verified disposition is approve the bounded test with conditions, not approve the tool. The evidence packet includes the repaired claim register, common option matrix, recalculation, sensitivity, control dispositions, dissent, pilot conditions, stop triggers, and six-week review date. It does not claim time savings, decision improvement, safety, or suitability for customer calls.
Independent release challenge
Give a reviewer one decisive recommendation sentence. Ask them to identify the authority, trace every necessary claim, reproduce a calculation, explain why the preferred option wins on common criteria, state the strongest objection, and name the event that would reverse the decision. Expand review if any answer depends on undocumented guidance.
Hold release when the decision is ambiguous, authority is missing, a material source cannot be verified, alternatives receive unequal treatment, a required gate is averaged away, calculations do not reproduce, implementation lacks an owner, or residual uncertainty disappears from the summary.
Two Prune can place decision-memo verification inside a realistic AI-enabled assessment and evaluate how a participant navigates evidence, checks AI-assisted work, tests calculations, weighs tradeoffs, and communicates a defensible recommendation. It does not approve the memo or make employment decisions.
Sources
HM Treasury: The Green Book
- Office of Management and Budget: Circular A-4
Sources: [1] [2]
Sources
- 1.HM Treasury: The Green Book · UK Government
- 2.Office of Management and Budget: Circular A-4 · www.whitehouse.gov