Practical guide

How to create a decision memo with AI

A useful decision memo lets an authorized reader choose, defer, or reject an action without reconstructing the analysis. It names the decision, compares real alternatives, traces material facts, explains uncertainty, recommends one course, and makes execution consequences visible. AI can help organize evidence, test the argument, and improve clarity; it cannot supply missing authority or turn an.

By Two Prune

A useful decision memo lets an authorized reader choose, defer, or reject an action without reconstructing the analysis. It names the decision, compares real alternatives, traces material facts, explains uncertainty, recommends one course, and makes execution consequences visible. AI can help organize evidence, test the argument, and improve clarity; it cannot supply missing authority or turn an unsupported preference into a decision.

Freeze the decision brief before drafting

Record the decision owner, decision date, question, affected population, constraints, non-negotiable requirements, evidence cutoff, and available options. Separate the person writing the memo from the person authorized to decide. If the question changes during analysis, revise the brief and explain why.

The UK Treasury Green Book treats appraisal as the assessment of costs, benefits, and risks across options rather than a justification for a preferred answer. U.S. OMB Circular A-4 similarly emphasizes defining need, considering alternatives, using a consistent analytical baseline, and describing uncertainty. These are public-decision sources, not universal memo templates; FRAME adapts their decision discipline to an ordinary workplace artifact. HM Treasury: The Green Book Office of Management and Budget: Circular A-4

Sources: [1] [2]

Build the memo with FRAME

### F — Fix the question, owner, and boundary

Open with one sentence that can be answered by a decision. State who decides, by when, what is inside scope, what is excluded, and what happens if no action is taken. Define terms that could change the choice, such as “pilot,” “production,” “approved source,” or “material risk.”

Do not begin with background. A reader should know the choice before seeing the history.

### R — Record sources, facts, and assumptions

Create a claim register before writing prose:

| Claim ID | Decision-relevant statement | Source and date | Fact, estimate, or assumption | Limitation | Memo use | | --- | --- | --- | --- | --- | --- | | C-01 | exact claim | stable source | classification | uncertainty or gap | section / option |

Prefer current first-party records for organizational facts and authoritative primary sources for external requirements. Mark estimates and assumptions in the sentence where they matter. Do not let AI-generated citations, reconstructed quotations, or unsourced market figures enter the memo as evidence.

### A — Analyze alternatives on common criteria

Include a credible status quo and at least one materially different alternative. Evaluate every option against the same criteria: decision fit, expected benefit, cost or effort, implementation time, reversibility, dependencies, affected groups, operational risk, and evidence strength.

Use ranges or scenarios when point estimates would overstate precision. Explain weighting rather than hiding judgment inside a total score. If a criterion is a gate, show it as a gate; do not compensate for a failed legal, security, accessibility, or safety requirement with a high benefit score.

### M — Make the recommendation and execution path

State the recommended option, why it wins under the stated criteria, and which evidence was decisive. Then specify the first owner, action, deadline, resources, dependencies, approval conditions, and stop or reconsider triggers.

Add the strongest reasonable objection. Answer it with evidence or accept it as residual uncertainty. A memo that removes every objection is usually hiding the tradeoff.

### E — Expose uncertainty, dissent, and the next evidence gate

Close with what is known, unknown, disputed, and time-sensitive. Record material dissent without converting it into anonymous “stakeholder concern.” Name the evidence that would change the recommendation and when the decision should be reviewed.

Use dispositions such as approve, approve with conditions, run a bounded test, request evidence, defer, or reject. Avoid “proceed cautiously” unless the memo defines an action and authority.

Use AI as a controlled drafting partner

AI can help extract candidate claims from supplied records, normalize option criteria, identify missing counterarguments, test whether the recommendation follows from the table, shorten repetitive prose, and simulate questions from finance, security, operations, or affected users.

For every material AI contribution, retain the input scope, output used, human verification, and disposition. Ask for traceability, not invention. A useful challenge prompt is: “List each sentence whose truth is necessary for this recommendation, identify its cited evidence, and flag any sentence that is an estimate, inference, or unsupported assumption.”

Do not paste restricted information into an unapproved system. The accountable author verifies sources, calculations, option treatment, and final wording.

Worked memo: whether to adopt an AI meeting assistant

A synthetic operations team asks whether to enable an AI meeting assistant for customer calls. The initial draft recommends an organization-wide rollout because the vendor claims large time savings.

FRAME changes the question to: “Should the operations director authorize a six-week, 20-user pilot for internal project meetings, subject to security and records controls?” Customer calls and automated follow-up messages are excluded because their recording, consent, and approval requirements are unresolved.

The claim register finds that the time-saving figure comes from vendor marketing with no matching local workflow evidence. It remains context, not a forecast. Internal records show that manual notes take 12–25 minutes per meeting, but the sample covers only eight employees over one week. The memo labels that range descriptive and incomplete.

The alternatives are no change, improve the existing note template, run the bounded pilot, or enable the tool broadly. Common criteria expose two gates: approved data handling and a reliable human review step before any note is shared. Broad enablement fails both gates. The template change is low risk but does not test transcription quality or review burden.

The recommendation is the bounded pilot. Conditions include internal meetings only, visible participant notice, no sensitive cases, a named tool administrator, human approval before distribution, deletion rules, and a pre-registered evaluation. The evidence artifact is a memo packet containing the brief, source register, option table, security disposition, pilot measures, dissent note, approval conditions, and review date.

At the end, the memo records the strongest objection: the pilot population may be unusually comfortable with AI and underestimate support needs. The next evidence gate therefore samples setup failures, correction time, missed or fabricated action items, opt-outs, accessibility issues, and review effort. The memo does not claim the tool will save time, improve decisions, or be safe for customer calls.

Acceptance and hold tests

Accept the memo only when the decision is explicit, alternatives are real, material claims trace to current sources, calculations can be reproduced, tradeoffs use common criteria, authority is named, and the recommendation has conditions plus a next evidence gate.

Hold it when the author cannot distinguish fact from assumption, AI-generated sources are unverified, a preferred option receives different criteria, a mandatory gate is hidden inside a weighted score, execution has no owner, or the requested decision exceeds the reader’s authority.

Two Prune can place decision-memo work inside a realistic AI-enabled assessment and evaluate how a participant frames a choice, navigates sources, uses AI, verifies claims, weighs tradeoffs, and communicates a defensible recommendation. It does not approve the decision or make employment decisions.

Sources

HM Treasury: The Green Book

  • Office of Management and Budget: Circular A-4

Sources: [1] [2]

Sources

  1. 1.HM Treasury: The Green Book · UK Government
  2. 2.Office of Management and Budget: Circular A-4 · www.whitehouse.gov

Explore this topic