Practical guide

AI-assisted decision memo template and guide

A decision memo should let an authorized reader understand the choice, evidence, alternatives, trade-offs, uncertainty, recommendation, and conditions for review. AI can help organize controlled material, compare options, and challenge omissions. It must not invent evidence, choose values, conceal disagreement, or become the decision-maker. Copy this template into the approved working system and.

By Two Prune

A decision memo should let an authorized reader understand the choice, evidence, alternatives, trade-offs, uncertainty, recommendation, and conditions for review. AI can help organize controlled material, compare options, and challenge omissions. It must not invent evidence, choose values, conceal disagreement, or become the decision-maker. Copy this template into the approved working system and adapt its depth to the consequence and reversibility of the decision.

Source basis and limits

The 2026 UK Green Book says appraisal should compare alternatives using evidence and present costs, benefits, distributional effects, risks, and uncertainty transparently. The Government Functional Standard for Analysis calls for timely, objective advice suitable for the decision audience and balanced communication of limitations and uncertainty. The AQuA Book requires proportionate assurance, traceable data and assumptions, independent review, and clear intended-use limits. NIST's Generative AI Profile recommends verifying sources and citations in generative-AI outputs. These sources support the memo controls below; they do not approve a recommendation, an AI tool, or a decision in a particular organization. The Green Book 2026 GovS 010: Analysis The AQuA Book NIST AI 600-1

Sources: [1] [2] [3] [4]

Copyable memo

### Decision header

Memo ID and version: Decision: Decision owner: Decision deadline: Author: Independent reviewer or assurer: Specialist reviewers: Evidence cut-off: Information classification and distribution:

### 1. Recommendation in one paragraph

State the recommended option, the principal reason, the most important trade-off, the confidence or evidence limitation, and the action requested. Do not write this section until the options and evidence have been reviewed.

Recommended disposition: proceed / proceed with conditions / investigate / repair / defer / stop. Action and owner: Condition or trigger: Review date:

### 2. Decision and boundaries

Decision question Write one choice that an authorized owner can make now.

Purpose and intended use Explain why the decision is needed and how the memo may be used.

In scope List the population, workflow, period, geography, product, or policy boundary.

Out of scope Name adjacent questions the memo does not answer.

Authority and constraints Record legal, policy, privacy, security, accessibility, financial, operational, and ethical constraints and their owners.

Consequence and reversibility Describe who benefits, who carries cost or risk, and what can be reversed.

### 3. Evidence register

| Source ID | Evidence | Owner | Date or period | Population | Quality or relevance | Limitation | Claim supported | | --- | --- | --- | --- | --- | --- | --- | --- | | E-01 | | | | | | | |

Separate source fact, analytical result, stakeholder view, professional judgment, assumption, and hypothesis. State how missing, conflicting, inaccessible, or stale evidence is handled. Do not infer “no effect” from absent evidence.

### 4. Objectives and decision criteria

List the outcomes or constraints that matter before scoring options. For each criterion, state the rationale, evidence route, unit or descriptive scale, owner, and whether it is a threshold or trade-off.

| Criterion | Why it matters | Evidence and method | Threshold or trade-off | Distributional effect | Uncertainty | | --- | --- | --- | --- | --- | --- | | | | | | | |

Avoid one weighted total when it would hide critical failures, non-compensable constraints, or disputed values. Show sensitivity to plausible changes in assumptions or priorities.

### 5. Options

Include business as usual, a non-technology or process option, the leading option, and a staged or reversible alternative where relevant.

| Option | Description | Evidence of benefit | Cost and burden | Risks and controls | Dependencies | Reversibility | Key uncertainty | | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | |

Explain why any longlisted option was removed. Do not construct a weak comparator to make the preferred option appear stronger.

### 6. Analysis and trade-offs

Summarize the method, assumptions, calculations, scenarios, and sensitivity checks. Identify:

Link detailed models, calculations, interviews, or legal analysis rather than compressing them into unsupported prose.

### 7. Uncertainty and evidence confidence

| Uncertainty | Source | Direction or range of effect | Decision sensitivity | Mitigation or evidence request | Revisit trigger | | --- | --- | --- | --- | --- | --- | | | | | | | |

Use ranges, scenarios, or descriptive uncertainty appropriate to the decision. Avoid unwarranted precision. State whether more evidence is likely to change the decision or only delay it.

### 8. Implementation and controls

If the recommendation is accepted, specify:

### 9. Assurance and decision record

Checks completed: Calculations reproduced by: Sources and citations reopened by: Specialist reviews and scope: Material disagreement: Known limitations: Assurance disposition: sufficient / sufficient with condition / repair / hold.

Decision taken: Decision owner and date: Conditions: Rejected option and reason: Next review:

  • which option performs best on which criterion;
  • material unmonetized or qualitative effects;
  • effects on different groups;
  • dependencies outside the decision owner's control;
  • rival explanations or dissenting views;
  • critical exceptions that cannot be averaged away; and
  • assumptions whose failure would change the recommendation.
  • accountable owner and collaborators;
  • first reversible action;
  • resource and dependency commitments;
  • privacy, security, accessibility, legal, financial, and operating controls;
  • monitoring and critical-failure indicators;
  • pause, stop, rollback, and escalation conditions; and
  • next decision and evidence date.

Bounded AI protocol

Use AI only for defined tasks with approved inputs.

| Pass | AI may | AI must not | Human check | Evidence retained | | --- | --- | --- | --- | --- | | Inventory | List sources and metadata | Add an unapproved source | Reconcile source register | Versioned inventory | | Compare | Surface agreements and conflicts | Resolve a conflict silently | Reopen authoritative records | Conflict disposition | | Structure | Draft headings and tables | Choose decision criteria or weights | Decision owner approves criteria | Approved outline | | Draft | Propose concise wording with source IDs | Invent facts, calculations, citations, or approval | Verify every material claim | Accepted and rejected edits | | Challenge | Suggest rival explanations and failure cases | Issue final assurance | Analyst and assurer adjudicate | Challenge log |

Record the tool, date, approved data boundary, responsible person, and any retained prompt or output required by policy. A qualified human remains accountable for evidence, recommendation, assurance, and decision.

Worked memo: a fictional knowledge-base migration

A service team must decide whether to replace its internal knowledge base in one release. The preferred vendor's demo appears faster, and an AI draft recommends immediate migration. The completed memo changes the recommendation.

Business as usual, process repair, phased migration, and full migration are compared. Evidence includes search logs, support tickets, accessibility testing, security review, content inventory, vendor documentation, migration samples, staff interviews, and cost ranges. The team cannot reliably monetize time saved because the observation sample is small and task mix differs, so the effect remains a range and qualitative consideration.

The evidence register reveals that 18% of high-use articles have no current owner, the vendor's import test lost embedded accessibility descriptions, and one required integration is documented only on a higher service plan. The AI comparison had treated missing article owners as a migration cleanup benefit and stated that the integration was included. Human source review corrects both claims.

Sensitivity analysis shows that a full migration becomes unattractive when cleanup takes more than six weeks. A phased option preserves the current system, migrates one low-risk collection, tests accessibility and search quality, and creates an exit. The memo recommends the phased option with conditions: named content owners, an accessible import test, verified plan entitlement, and a six-week review. It does not claim that the vendor will improve productivity or service outcomes.

Memo QA checklist

[ ] One authorized owner can make the stated decision by the stated date.

  • [ ] Purpose, scope, exclusions, constraints, consequences, and reversibility are clear.
  • [ ] Sources, results, views, judgment, assumptions, and hypotheses are distinguishable.
  • [ ] Business as usual and credible alternatives are represented fairly.
  • [ ] Criteria and thresholds were set before a preferred result was known.
  • [ ] Costs, benefits, risks, distributional effects, dependencies, and uncertainty are visible.
  • [ ] Critical failures and non-compensable constraints are not hidden by a total score.
  • [ ] Calculations, citations, and decision-driving claims were independently checked.
  • [ ] The recommendation states conditions, controls, stop rules, and a revisit trigger.
  • [ ] The final record preserves the decision, authority, rejected options, and limitations.

Version history

| Version | Date | Change | Evidence affected | Author | Approver | | --- | --- | --- | --- | --- | --- | | 0.1 | | Initial brief | | | |

Two Prune boundary

Two Prune can assess whether a participant frames a decision, navigates supplied evidence, uses AI thoughtfully, compares options, verifies claims and calculations, handles uncertainty, and communicates a recommendation in a realistic simulation. It reports observed assessment performance. It does not approve a decision, validate a business case, certify compliance, predict workplace performance, or make employment decisions.

Sources

HM Treasury: The Green Book 2026

  • UK Government Functional Standard GovS 010: Analysis
  • GOV.UK: The AQuA Book
  • NIST AI 600-1: Generative Artificial Intelligence Profile

Sources: [1] [2] [3] [4]

Sources

  1. 1.The Green Book 2026 · UK Government
  2. 2.GovS 010: Analysis · UK Government
  3. 3.The AQuA Book · UK Government
  4. 4.NIST AI 600-1 · National Institute of Standards and Technology