Key takeaway
The model should widen and structure your thinking. Decision ownership, values, evidence, and accountability remain human.01
Describe the decision precisely
State the decision, owner, deadline, scope, success criteria, constraints, and what happens if no action is taken. Check whether you are deciding between real options or accepting a framing that excludes better alternatives.
Avoid asking only “What should I do?” The response will reflect hidden assumptions you have not examined.
02
Supply evidence and stakeholder views
Provide the available data, source dates, current plan, prior decisions, and stakeholder concerns. Label opinions and facts. Include constraints that cannot be changed and preferences that can.
Ask the model to identify evidence gaps and assumptions rather than smoothing them into a confident recommendation.
03
Generate genuinely different options
Request options built on different strategic choices, not superficial variations. Include maintaining the current state, delaying for evidence, or running a limited experiment where relevant.
Define common comparison criteria before evaluating options. Otherwise, the response may describe one option generously and another narrowly.
04
Challenge the recommendation
Use a pre-mortem, second-order effects, stakeholder impact, reversibility, and sensitivity analysis. Ask what evidence would change the recommendation and which assumption carries the greatest risk.
Do not ask the model to imitate a hostile critic for entertainment. Ask for specific failure mechanisms and observable signals.
Challenge instruction
Assume the recommended option failed after six months. Identify the most plausible causes, the earliest signal for each, and one mitigation we can implement before committing.
05
Prefer reversible learning
When uncertainty is high and delay is affordable, design the smallest test that distinguishes between competing beliefs. Define the metric, guardrail, duration, decision rule, and what the test cannot prove.
A pilot should reduce uncertainty, not merely postpone a difficult decision. Assign an owner and a date for interpreting the result.
06
Document the human decision
Write a memo containing the evidence, options, recommendation, dissent, assumptions, and review triggers. Identify which values or trade-offs were human judgments rather than model conclusions.
For legal, financial, employment, health, safety, or other high-impact decisions, follow applicable policy and involve qualified professionals. Do not use an AI recommendation as authority.
Review
Practical checklist
- The decision, owner, and deadline are clear.
- Success criteria and constraints are explicit.
- Facts and stakeholder opinions are labeled.
- Options use the same evaluation criteria.
- Second-order effects and failure modes are reviewed.
- A reversible experiment is considered.
- The final human rationale is documented.
FAQ
Common questions
Can AI remove bias from a decision?
No. It can help expose assumptions and generate alternative perspectives, but models and supplied context can contain bias. Human review and fair processes remain necessary.
Should I ask several models for recommendations?
Multiple models may reveal different framings, but agreement does not prove correctness. Compare their assumptions and evidence rather than counting votes.
What decisions should not be delegated to AI?
Decisions involving rights, safety, professional obligations, or material consequences require accountable human processes and, when applicable, qualified professional judgment.
Ready to apply the method?
Build your own prompt