Executive Decision Memo Prompt System with Verification
A controlled prompt sequence for drafting, challenging and verifying an executive decision memo without allowing AI fluency to replace evidence.
Advanced Executive Program in Management & Business Administration develops connected capability across strategy, finance, customers, operations, technology and people. This practical prompt system can be used before enrollment, during executive study or inside an authorized workplace exercise.
A single prompt can produce polished but unreliable prose. The system separates evidence extraction, option design, challenge, calculation checks and final verification so that every material claim remains traceable.
Human-owned memo standard
How can an executive use AI to improve a decision memo without inventing facts or surrendering accountability? Use the tool on this page to produce a six-prompt system, verification worksheet and safe-use rules. Start with a bounded decision, retain the evidence behind every material claim, and distinguish what is known from what is assumed. The tool is designed to improve management preparation and review; it does not replace the authority, specialist judgement or procedures required by an employer.
Where AI assistance creates hidden evidence risk
The prompt system creates a compact common language for those connections. It does not force every organization into the same answer. Instead, it makes local definitions, evidence, constraints and accountability visible. That is useful when a management team agrees on the goal but disagrees about the route, or when confident recommendations rely on incompatible assumptions.
Prompt roles and verification gates
| Element | Management purpose | Minimum evidence |
|---|---|---|
| Evidence boundary | Tell the model which sources are authorized and how unsupported claims must be marked. | The model may organize supplied evidence but cannot silently complete missing facts. |
| Decision frame | Define the choice, owner, deadline, scope and constraints. | A memo without a decision frame becomes a topic summary. |
| Option builder | Generate genuinely feasible alternatives, including maintaining the current course. | Options must state dependencies, reversibility and resource consequences. |
| Critic prompt | Ask a separate pass to find contradictions, weak evidence and omitted stakeholders. | Challenge should be independent from drafting instructions. |
| Verification prompt | Require a claim-to-source table and recalculation of visible numbers. | The output is a review aid, not proof that the sources themselves are correct. |
| Human sign-off | Assign named people to validate evidence, authority, risk and final wording. | Accountability cannot be delegated to the prompt system. |
1. Evidence boundary
Tell the model which sources are authorized and how unsupported claims must be marked. The model may organize supplied evidence but cannot silently complete missing facts.
Review question: What observable evidence would confirm that evidence boundary is working in the selected scope, and who has authority to respond when it is not?
2. Decision frame
Define the choice, owner, deadline, scope and constraints. A memo without a decision frame becomes a topic summary.
Review question: What observable evidence would confirm that decision frame is working in the selected scope, and who has authority to respond when it is not?
3. Option builder
Generate genuinely feasible alternatives, including maintaining the current course. Options must state dependencies, reversibility and resource consequences.
Review question: What observable evidence would confirm that option builder is working in the selected scope, and who has authority to respond when it is not?
4. Critic prompt
Ask a separate pass to find contradictions, weak evidence and omitted stakeholders. Challenge should be independent from drafting instructions.
Review question: What observable evidence would confirm that critic prompt is working in the selected scope, and who has authority to respond when it is not?
5. Verification prompt
Require a claim-to-source table and recalculation of visible numbers. The output is a review aid, not proof that the sources themselves are correct.
Review question: What observable evidence would confirm that verification prompt is working in the selected scope, and who has authority to respond when it is not?
6. Human sign-off
Assign named people to validate evidence, authority, risk and final wording. Accountability cannot be delegated to the prompt system.
Review question: What observable evidence would confirm that human sign-off is working in the selected scope, and who has authority to respond when it is not?
Controlled drafting sequence
Step 1: Prepare inputs
Remove restricted data and label every source, date and owner. The immediate output is an authorized evidence pack.
Step 2: Run evidence extraction
Ask for relevant facts, uncertainties and conflicts only. The immediate output is a structured evidence register.
Step 3: Frame the decision
Ask the model to restate the choice and list missing information. The immediate output is a validated decision question.
Step 4: Develop options
Request alternatives with assumptions, consequences and reversibility. The immediate output is an option table.
Step 5: Draft the memo
Generate a concise recommendation using only accepted evidence. The immediate output is a reviewable first draft.
Step 6: Run the critic
Use a new context where possible and instruct it to seek disconfirming evidence. The immediate output is a challenge report.
Step 7: Verify claims
Map each material statement and calculation to source evidence. The immediate output is a claim ledger.
Step 8: Approve or stop
A human owner resolves findings and signs the final decision. The immediate output is a defensible memo or an explicit information gap.
Memo example: channel investment
An executive team is considering a customer-service automation pilot. The evidence pack contains contact volumes, service levels, five customer interviews, vendor estimates and security constraints. The prompt system must not invent adoption rates, savings or legal conclusions.
| Evidence or choice | Current entry | Interpretation | Management response |
|---|---|---|---|
| Evidence prompt | Extract facts and unresolved conflicts | Volume and wait time are sourced; savings remain estimates | Accept with labels |
| Options prompt | Build three reversible choices | No change, narrow pilot, full deployment | Keep first two for comparison |
| Critic prompt | Seek failure paths and excluded groups | Accessibility and escalation workflow missing | Return to design |
| Verification prompt | Recalculate stated savings | Draft double-counted reduced handle time | Correct calculation |
| Human sign-off | Confirm authority and risk acceptance | Pilot owner approves narrow scope | Proceed with review gate |
Unsafe prompt and review patterns
- Placing confidential material in an unapproved tool.
- Asking one prompt to draft and approve its own work.
- Accepting citations that were not in the supplied sources.
- Allowing estimates to become facts.
- Using AI to make legal or regulated decisions.
- Publishing the memo before human verification.
Governance questions
Should the same model run the critic?
A separate context or model can reduce anchoring, but independence of instructions and evidence review matters more than the brand of model.
Can the system verify truth?
It can test internal consistency and traceability. Humans must still assess whether sources are authentic, current and sufficient.
What data should be excluded?
Exclude personal, confidential, privileged, security-sensitive or contract-restricted data unless the approved environment and purpose explicitly permit it.
When should the process stop?
Stop when a material claim lacks evidence, authority is unclear, risk requires specialist judgement or the decision cannot be made within stated constraints.
AI governance evidence and scope
The U.S. National Institute of Standards and Technology maintains the AI Risk Management Framework and its Generative AI Profile, NIST AI 600-1. NIST’s AI RMF Core describes governance, contextual mapping, measurement and management, including documented human-AI roles and testing, evaluation, verification and validation. This article applies those principles to one narrow workflow: drafting and challenging an executive decision memo. It does not claim that following five prompts makes an organization compliant with the AI RMF or any law.
The source-bounded design answers a predictable failure: fluent text can detach a claim from its evidence. Separating evidence extraction, framing, alternatives, criticism and memo drafting reduces the number of goals in each prompt and creates human review points. It does not eliminate hallucination or automation bias. The human reviewer must open critical sources, recompute important numbers and assess whether the evidence is authorized and fit for the decision.
Users must confirm approved tools, contracts, data locations, retention, privacy, security, professional obligations and any legal requirements. If a model output contains potentially exposed sensitive information, follow the authorized incident route rather than copying it into another system. High-consequence decisions remain with properly authorized humans and qualified specialists. The prompts are reusable text, not an autonomous agent and not a technical security control.
Verification trial before workplace use
Run the prompt sequence first on a fictional decision with a deliberately seeded source pack. Include one conflicting number, one stale source, one unsupported claim, one ambiguous population and one calculation that can be recomputed. The expected result is not persuasive prose; it is correct identification and handling of the traps.
Ask two human reviewers to verify the evidence ledger independently. Measure fabricated citations, missed conflicts, incorrect calculations, unsupported memo claims and review time. Repeat with an ordinary manual memo so the organization can see whether AI assistance changes both speed and error, not speed alone. A small trial does not establish general performance across models or decisions.
Test the stopping rules by removing a critical source and by placing a prohibited instruction inside the supplied content. The system should return NOT ESTABLISHED or stop; it should not follow a document’s embedded instruction or replace the missing fact from memory. Confirm that prompts, outputs and logs remain in the approved environment.
Only after the failure taxonomy is understood should the organization pilot a low-consequence real case. Keep the final decision and any external communication outside the model workflow until authorized humans complete verification and required specialist review.
The trial should include model and prompt version, temperature or other material settings when available, source-pack checksum, reviewer identity and date. A later model update may change behaviour even when the visible prompt is identical. Do not generalize one successful trial across languages, domains or data classifications. Re-test when the model, retrieval method, document parser, source type or decision consequence changes. Preserve only the evidence permitted by retention and privacy rules.
Human sign-off record
The final memo should include a compact sign-off table that separates contribution from authority.
| Review | Human question | Sign-off evidence |
|---|---|---|
| Decision owner | Is this the decision I am authorized to make? | Name, date and final disposition |
| Source owners | Are decisive facts represented with correct definition and limitation? | Verified claim IDs |
| Finance | Do calculations reconcile and use compatible assumptions? | Recalculation reference |
| Operations | Are sequence, capacity and dependencies feasible? | Accepted implementation boundary |
| Required specialist | Does the option satisfy the review within that specialist’s authority? | Qualified conclusion or conditions |
| Recorder | Does the stored memo match the approved version? | Version and controlled location |
The AI output cannot occupy any sign-off field. If a reviewer changes a claim, rerun the dependent calculation or argument rather than editing only the sentence. If reviewers disagree, preserve the disagreement and route the decision to the authorized owner. “Reviewed” is too vague when the person checked only style.
Before release, the recorder samples the memo backward: take each recommendation and trace it to options, reasoning, verified claims and sources. Take each action and trace it to capacity, authority and a review rule. This reverse test catches a common defect—well-supported analysis followed by an implementation promise that was never evaluated.
Prompt-change control
Treat a material prompt revision like a method change. Record the reason, expected effect and examples used to test it. A wording change that appears cosmetic can alter source use, confidence or output structure. Keep a stable fictional test pack containing conflicting evidence, missing data and a verifiable calculation, then compare the revised sequence with the prior version.
Do not judge only whether the new memo reads better. Compare claim traceability, fabricated references, missed limitations, calculation errors, unsupported recommendations and reviewer correction time. If one quality dimension improves while another weakens, retain the earlier version or narrow the use case until the trade-off is understood.
Memo release checklist
- The decision question and owner match the authorized record.
- Every decisive factual claim has been opened and verified by a human.
- Important calculations have been recomputed outside the model output.
- Facts, estimates, assumptions and hypothetical scenarios are distinguishable.
- Feasible alternatives include the current course and a staged option where appropriate.
- Specialist conclusions are quoted or summarized within their approved scope.
- Implementation owners accept capacity, dependencies and dates.
- The memo includes outcome, harm, review and stop conditions.
- The stored version matches the approved text and preserves the source ledger.
If any critical item fails, do not compensate with stronger wording. Return the memo to the relevant verification gate.
Controlled prompt system for an executive decision memo
The system below uses generative AI only as a drafting and challenge aid. A named human owns the decision, evidence selection, specialist review and final wording. Do not place confidential, personal, regulated, privileged or contract-restricted information into an unapproved model. Do not ask a model to invent missing facts, browse without source control or make the decision.
Record zero: authorization and data boundary
Complete this before any prompt:
Decision ID and owner:
Approved AI tool / environment:
Permitted data classification:
Prohibited content:
Source pack version and cutoff:
Required reviewers:
Decision deadline:
Final record location:
If the environment or data permission is uncertain, stop. Redaction is not automatically sufficient: combinations of apparently harmless details can identify a customer, employee or transaction. Use a fictionalized case when the learning purpose does not require workplace data.
Prompt 1 — source-bounded analyst
ROLE: You are a source-bounded analyst, not the decision maker.
TASK: Convert the supplied source pack into an evidence ledger for [decision].
RULES:
1. Use only the supplied material. Do not rely on memory or add facts.
2. Give every material statement a source ID and exact location.
3. Separate observation, calculation, estimate, assumption and opinion.
4. Mark conflicts, missing definitions, stale evidence and incompatible populations.
5. If support is absent, write NOT ESTABLISHED.
OUTPUT: table with claim, type, source ID/location, date/population, limitation,
and decision relevance; then a gap list. Do not recommend an option.
Verification gate V1: a human opens every source cited for a decision-critical claim. Sample citations are not enough for the facts that could reverse the decision. Recompute important arithmetic outside the model. Reject invented source IDs, quotations or precision not present in the pack.
Prompt 2 — decision framer
Using only the verified evidence ledger, draft:
- one-sentence decision question;
- owner, authority, scope, horizon and exclusions;
- constraints versus preferences;
- stakeholders materially affected;
- latest responsible decision date;
- uncertainties that could change the choice.
Offer two alternative frames and explain how each changes the option set.
Do not select a preferred option.
Verification gate V2: the decision owner chooses or rewrites the frame. Confirm that a mandatory obligation has not been converted into a weighted preference and that the frame does not smuggle in the sponsor’s solution.
Prompt 3 — alternative designer
Generate a strategy table with genuinely different feasible alternatives,
including current course, delay and a bounded learning option when possible.
For each, list mechanism, resources, dependencies, reversibility, first
irreversible step and evidence needed. Do not score options. State when a
suggested alternative conflicts with a verified constraint.
Verification gate V3: functional owners validate feasibility and capacity. Delete alternatives that depend on nonexistent authority or capability; do not retain them to make the preferred option look better. Add a neglected option identified by an implementer where appropriate.
Prompt 4 — red-team critic
Act as an independent critic. For each alternative, identify:
- the strongest disconfirming evidence;
- hidden assumptions and denominator problems;
- downside and affected stakeholder;
- dependency or capacity failure;
- evidence that would reverse the ranking;
- a pre-mortem explaining a plausible failure.
Do not introduce unsourced factual claims. Label hypothetical tests clearly.
Verification gate V4: assign each challenge to a human owner: resolved, accepted as uncertainty, requires investigation, or outside scope with reason. Do not ask the drafting model to judge its own unsupported claim as verified.
Prompt 5 — memo drafter
Draft a two-page executive decision memo from the verified frame, evidence
ledger, alternative table and resolved challenge log. Structure:
1. Decision requested and latest responsible date
2. Recommendation with confidence
3. Evidence and calculations
4. Alternatives and trade-offs
5. Risks, unknowns and required specialist review
6. Implementation, owner, resources and dependencies
7. Measures, review date, adapt/stop triggers
Use source IDs after every material claim. Preserve disagreement. Write
NOT ESTABLISHED for an unresolved factual gap. Do not create quotations.
Verification gate V5: the named owner compares every sentence with the controlled inputs, removes rhetorical certainty not supported by evidence, secures required professional review and signs the final human-authored record. Store prompts and outputs only if permitted by retention policy.
Worked example: channel investment
A services firm must choose whether to invest €420,000 in a partner channel. The source pack contains two years of direct-channel cohort data, a partner proposal, a capacity estimate and customer interviews. The first AI draft states that partner customers have “higher lifetime value,” but the evidence ledger shows no partner cohort: the claim is NOT ESTABLISHED. It also compares gross partner leads with qualified direct opportunities, an incompatible denominator.
The framer produces two questions. Frame A asks whether to launch the full partner program. Frame B asks which reversible step can test partner-sourced economics before a €420,000 commitment. The owner selects Frame B because the deadline is self-imposed and the major uncertainty is conversion quality.
The alternative prompt produces full launch, no launch, a 90-day two-partner pilot and purchase of market research. Operations notes that onboarding both partners uses the same scarce implementation specialist, so the supposed two-partner diversification is not operationally independent. The red team identifies a selection risk: partners may nominate unusually strong early prospects. The memo therefore recommends a capped pilot with pre-agreed eligibility, a €65,000 ceiling and no exclusivity.
The decision record states the calculation rather than merely the conclusion. Pilot success requires at least 30 qualified opportunities, conversion no more than five percentage points below the comparable direct cohort, contribution margin no more than three points below, and onboarding effort below 18 hours per activated customer. The stop trigger is any material compliance breach or a forecast that exceeds the authorized capacity. The decision owner, not the AI system, accepts those trade-offs.
Claim-verification table
| Memo claim | Source location | Check | Status | Human owner |
|---|---|---|---|---|
| Addressable cohort size | S1, table 3 | Population and date match | Verified / revise | Commercial |
| Contribution calculation | S2, cells D12:D18 | Independently recompute | Finance | |
| Implementation effort | S3, section 4 | Confirm role and calendar capacity | Operations | |
| Customer concern | S4, interview IDs | Avoid generalizing qualitative sample | Customer lead | |
| Legal feasibility | Specialist note | Required approval exists | Legal / compliance |
Stopping rules
Stop the AI-assisted workflow if a required source cannot be opened; critical citations are fabricated; the model repeatedly ignores the source boundary; prohibited data entered the system; material claims cannot be verified before the decision deadline; or the use case crosses a policy boundary that requires a different approved process. Escalate potential data exposure under the organization’s incident procedure rather than deleting local evidence and pretending it did not occur.
NIST’s AI Risk Management Framework emphasizes documented roles, context, testing, evaluation, verification and validation. This prompt system operationalizes those ideas for a narrow writing task; it is not a claim of formal compliance. The Advanced Executive Program in Management & Business Administration develops the managerial judgment needed to connect AI use with strategy, risk and execution.