To create a decision memo with AI, define authority, options, criteria, and evidence first. AI organizes the draft; the authorized person decides.

Key Takeaways

Name one decision, one decision owner, and one deadline.

Define scope, constraints, and mandatory criteria before comparing options.

Include the status quo when it is a real alternative.

Give every material claim an evidence ID, date, owner, and applicability note.

Calculate scores and financial effects outside the language model.

Show uncertainty, sensitivity, conflicts, and dissent rather than smoothing them away.

Record the human decision separately from the AI-assisted recommendation.

What does it mean to create a decision memo with AI?

A decision memo is a compact record that frames a specific choice, compares viable alternatives against established criteria, presents evidence and uncertainty, and gives an authorized person enough context to decide. It is not merely a summary. A summary can describe information without explaining which choice is required, who has authority, or what trade-offs matter.

NASA’s decision-analysis guidance says a decision report should cover the decision, criteria, alternatives, evaluation methods, process and results, recommendation, and final decision.［1］ That structure makes a useful memo contract even outside engineering, because it separates analysis from authority.

Step 1: Frame the decision and authority

Write a one-sentence decision question with a verb and a bounded object: select, approve, defer, stop, fund, adopt, or test something. Avoid broad prompts such as “What should our strategy be?”

Field	What to record
Memo ID/version	Stable record and revision
Decision question	Exact choice requiring action
Decision owner	Person or body authorized to decide
Decision deadline	Latest useful date
Intended outcome	Result the decision should support
Scope/non-goals	Included and excluded questions
Constraints	Mandatory policy, safety, capacity, cost, or schedule limits
Contributors	Analysts, subject owners, and affected reviewers
Review trigger	Event or evidence that requires reconsideration

The memo author may recommend an option but cannot grant authority. If authority is unclear, stop at a draft and route the question to the governance owner. Do not ask AI to infer who can approve from titles, meeting notes, or organizational habits.

Step 2: Build a decision evidence ledger

Create the evidence ledger before prose. Each entry should support a defined claim and state where it applies.

Evidence field	Purpose
Evidence ID	Stable citation used in the memo
Claim	What the evidence can establish
Source/owner	Original system, document, study, or expert owner
As-of date	Time boundary
Scope	Product, population, market, process, or scenario
Method	How the evidence was produced
Quality/limitations	Missingness, bias, uncertainty, or conflict
Permission	Whether the material may be processed by AI

Use only the minimum approved extract. Replace names and confidential values with synthetic placeholders when the workflow is being designed. Never paste legal privilege, unreleased financials, customer data, credentials, or sensitive personnel material into an unapproved model.

AI may classify evidence, identify unsupported claims, and draft questions. It must not generate citations that were not supplied. NIST identifies confabulation as confident false or erroneous output and notes that generated logic or citations can also mislead users.［4］ Every factual sentence in the memo therefore needs a route back to an independently opened source.

Step 3: Define the options before scoring

List mutually understandable alternatives. Include “maintain the current state” or “defer pending evidence” when those are real choices. Do not hide an option simply because the drafting team dislikes it.

For each option, record:

a stable option ID and neutral description;
prerequisites and dependencies;
expected effects and affected groups;
irreversible steps and exit costs;
evidence supporting and challenging the option;
unresolved questions and the owner of each;
whether it meets every mandatory constraint.

Avoid option labels such as “safe plan” and “expensive plan.” Those names preload the conclusion. Give AI a naming rule that is descriptive and symmetric.

Step 4: Establish criteria and scales

Define criteria before evaluating alternatives. NASA recommends objective, measurable criteria that differentiate options and distinguishes mandatory from enhancing criteria.［1］ A mandatory failure should be shown directly, not diluted by unrelated high scores.

Criterion field	Definition
Criterion ID/name	Stable identity
Decision relevance	Why it matters
Measure/source	Evidence or calculation used
Scale	Operational definition for every rating
Direction	Higher, lower, pass/fail, or bounded range
Mandatory?	Whether failure removes an option
Weight	Decision-owner-approved importance, if used
Uncertainty	Range, confidence, or unresolved condition

Weights are preferences, not facts. AI can check that weights sum correctly or explain their effect, but only the decision owner can approve them. If a team cannot agree on weights, show several plausible sets and test whether the recommendation changes.

Step 5: Choose an analysis proportionate to the decision

Not every memo needs a complex matrix. A reversible, low-impact choice may need a short comparison and named risks. A costly, safety-critical, legally constrained, or hard-to-reverse decision needs deeper evidence, specialized review, and explicit uncertainty.

Possible methods include a requirements screen, a trade-off table, a controlled cost-benefit model, a pilot, user research, technical testing, or scenario analysis. The NASA Systems Engineering Handbook says the rigor of analysis should be commensurate with the consequences of the decision.［2］ Record why the method fits and what it cannot show.

When monetary or quantified effects matter, use a governed calculation and the cost-benefit analysis workflow. When future conditions could reverse the result, use AI-assisted scenario analysis. The decision memo should reference those artifacts rather than re-create their arithmetic in prose.

Step 6: Calculate outside AI and lock the result

Build formulas in a controlled spreadsheet, script, or analysis system. Test units, signs, periods, missing values, duplicate records, boundaries, and reconciliation totals. Save the model version and output before asking AI to draft commentary.

Result field	What the memo needs
Analysis ID/version	Reproducible calculation reference
Option/criterion	Exact matrix location
Input source	Evidence IDs and cutoff
Value/rating	Locked result
Scale/unit	Operational meaning
Uncertainty	Range or unresolved state
Reviewer	Person who reproduced the result

Do not copy a generated total into the official matrix. If AI detects a possible inconsistency, return to the controlled model, verify it, and create a new version.

Step 7: Draft, challenge, and record the memo

Give the model the decision contract, option records, criterion definitions, evidence ledger, and locked results. Request a structured draft with these labels:

Decision required — the exact choice and deadline.
Context — only verified background needed to understand it.
Options — neutral descriptions, including status quo where applicable.
Evaluation — results against established criteria.
Uncertainty — unknowns, conflicts, assumptions, and sensitivity.
Recommendation — the analyst’s proposed course and rationale.
Dissent — credible objections or minority views.
Decision record — blank until the authorized person decides.

Require evidence IDs after material claims. Prohibit new numbers, sources, weights, approvals, and facts. Ask the model to return unsupported when a statement cannot be traced.

Step 8: Separate facts, inferences, and recommendations

Review every paragraph using a simple label.

Label	Meaning	Required support
Fact	Directly established by evidence	Evidence ID and scope
Calculation	Deterministic output	Model version and inputs
Inference	Interpretation connecting evidence	Reasoning and uncertainty
Assumption	Condition not yet established	Owner and validation plan
Recommendation	Proposed action	Criteria, trade-offs, and authority boundary

Do not let a fluent transition turn an inference into a fact. “The pilot had fewer errors” may be factual; “the option will reduce errors everywhere” is a broader prediction requiring more evidence.

Step 9: Test uncertainty and sensitivity

Ask which uncertain inputs or preferences could change the ordering. The Green Book explains that risk and uncertainty should be considered in appraisal and that sensitivity analysis can show how uncertain assumptions affect results.［3］ Use that idea proportionately, without importing public-sector rules into a different organization as mandatory policy.

Run at least these checks:

replace provisional inputs with their plausible low and high values;
test alternative decision-owner-approved weights;
identify evidence that, if wrong, could reverse the recommendation;
show options that remain close rather than forcing a false winner;
describe what additional research would reduce material uncertainty;
record the cost and delay of obtaining that evidence.

If the recommendation changes easily, say so. The honest conclusion may be a bounded pilot, a delayed decision, or a choice among closely ranked alternatives.

Step 10: Run an adversarial and stakeholder review

Assign reviewers to challenge different aspects: evidence quality, operational feasibility, security, finance, legal or policy constraints, affected-user impact, and implementation capacity. AI can generate challenge questions, but the responses must come from evidence owners.

Preserve substantive dissent in the memo. Summarize the strongest alternative case, the evidence behind it, and why the recommendation still differs. Do not ask the model to “resolve disagreement” by averaging incompatible judgments.

If competitors are discussed, first use the evidence-backed competitive analysis. Do not import unsourced feature claims or market assumptions into the memo.

Step 11: Record the human decision and follow-up

Leave the decision section visibly blank or marked pending decision until the authorized person acts. Then record:

selected option or explicit deferral;
decision-maker and timestamp;
rationale, including any departure from the recommendation;
conditions, approvals, or funding limits;
action owners and due dates;
evidence or metrics required after implementation;
review trigger and scheduled review date.

The decision record should be append-only or versioned. Do not rewrite the recommendation to make it appear that the analysis predicted the final choice.

Final checklist

One decision, owner, and deadline are explicit.
Scope, constraints, options, and status quo are visible.
Criteria and scales were defined before scoring.
Every material fact maps to an evidence ID and applicability note.
Calculations are locked and reproducible outside AI.
Assumptions, conflicts, uncertainty, and sensitivity are shown.
Dissent receives a fair, evidence-based summary.
The recommendation is separate from the authorized decision.
Follow-up actions, owners, and review triggers are recorded.

FAQ

How is a decision memo different from a report?

A report may describe a subject without requiring a choice. A decision memo is organized around one decision, the authority to make it, viable options, explicit criteria, evidence, uncertainty, a recommendation, and a separate final record.

What is the minimum useful decision memo?

At minimum, include the decision question, owner, deadline, options, mandatory constraints, key evidence, material unknowns, recommendation, and a blank decision record. Add analysis depth in proportion to consequence and reversibility.

Should the status quo always be an option?

Include it when maintaining the current state or deferring action is genuinely possible. If policy, safety, or a fixed deadline makes it impossible, explain that constraint rather than pretending it is viable.

When should I use a weighted decision matrix?

Use one when several differentiating criteria need a transparent comparison and the decision owner can define scales and weights. Do not use weights to disguise a mandatory failure or force agreement where evidence is missing.

What should I do when sources conflict?

Keep both sources, document their scope, date, method, and owner, and state which claims remain unresolved. Ask what additional evidence could distinguish them. Do not let AI select the more convenient source.

Can AI write the recommendation?

AI can draft recommendation language from supplied criteria, evidence, and locked results. The analyst must verify the reasoning, and the authorized decision-maker remains free to select another option or request more analysis.

How should dissent be handled?

Record the strongest evidence-based objection, not a weakened version. Identify the reviewer, affected criterion, supporting evidence, and condition under which the alternative would become preferable.

Can AI fill in the final decision section?

Only after a verified human decision has been supplied as input, and even then the text remains a draft for confirmation. AI must never infer approval from meeting tone, silence, calendar events, or an earlier recommendation.

When should a decision memo be reopened?

Use named triggers such as a material evidence revision, breached assumption, legal change, failed pilot condition, cost threshold, or implementation result. Preserve the original decision and create a new version for reconsideration.

Sources

NASA — Decision Analysis — https://www.nasa.gov/reference/6-8-decision-analysis/
NASA — Systems Engineering Handbook — https://www.nasa.gov/wp-content/uploads/2018/09/nasa_systems_engineering_handbook_0.pdf
HM Treasury — The Green Book 2026 — https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government/the-green-book-2026
NIST — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Sources checked 4 September 2026.

Related articles:

How to Create a Decision Memo with AI