Use AI with prioritization frameworks

Frameworks emphasize different concerns. RICE combines reach, impact, confidence, and effort. ICE uses impact, confidence, and ease for faster comparison. Value-versus-effort supports discussion without pretending to numerical precision. Cost of delay highlights urgency. Kano explores expected and differentiating user value. Risk matrices examine likelihood and consequence.

Do not ask AI, “Which framework should we use?” without context. State the decision:

• Are you sequencing work within one product outcome?
• Comparing investments across unlike products?
• Choosing experiments under uncertainty?
• Managing deadlines or regulatory obligations?
• Balancing growth, reliability, accessibility, and technical health?

No single framework captures strategy, dependencies, fairness, contractual commitments, or existential risk. Use a framework as one view and document factors handled outside it.

Before scoring, create one record per input:

Estimate: “12,000 eligible monthly accounts.”
Definition: Accounts that attempt a supported import, not all active accounts.
Source: Event query and date range.
Owner: Analytics lead.
Confidence: Medium because instrumentation misses mobile retries.
Sensitivity: If actual reach is below 7,000, the sequence may change.

AI can extract candidate values from documents and identify conflicting estimates. It should not silently choose one. Require citations or source IDs and represent unknown values as ranges.

Normalize definitions across proposals. “Reach” might mean people exposed, people who benefit, or transactions affected. “Effort” might include engineering but omit design, data labeling, evaluation, legal review, support, monitoring, and incident response. A comparable score requires comparable definitions.

Point estimates conceal uncertainty. Instead of assigning reach 12,000, use a plausible range such as 7,000–15,000. Instead of “three weeks,” estimate design, engineering, evaluation, rollout, and ongoing operations separately.

Ask AI to compute low, base, and high scenarios using values supplied by the team. Then inspect whether the ranking changes:

• If a proposal wins in every plausible scenario, the decision is robust.
• If small changes reverse the ranking, gather better evidence or treat items as effectively tied.
• If one uncertain input dominates, focus discovery there.
• If a severe risk is excluded from the formula, add a decision gate rather than a tiny score penalty.

Sensitivity analysis is often more informative than the final rank. The assistant can vary one field at a time and explain which assumptions matter most, provided the formulas and input ranges are explicit.

A model asked to “score these ideas” will often fill empty cells with plausible numbers. That behavior is convenient and dangerous. Use a strict contract:

• calculate only from supplied inputs;
• show the formula and substitutions;
• mark missing data as unknown;
• do not convert qualitative evidence into a number without an approved mapping;
• preserve ranges and confidence;
• list excluded factors;
• separate calculation from recommendation.

Check arithmetic independently in a spreadsheet or deterministic script. Generative AI is useful for structuring, explaining, and challenging assumptions, but ordinary calculation tools should own repeatable math.

Some work should not compete only on a value score. Security remediation, accessibility obligations, legal requirements, contractual commitments, and critical reliability work may be mandatory. Mark them as constraints or reserved capacity rather than distorting estimates to make them rank highly.

AI features also introduce harms that a generic framework can miss: privacy exposure, unequal quality across user groups, unsupported advice, automation bias, manipulation, and costly review queues. Evaluate consequence, reversibility, detectability, scale, and affected groups.

A high expected-value feature may still need to wait until evaluation data, permissions, or a safe fallback exists. “High priority” is not the same as “launch ready.”

Use AI to prepare a comparison brief:

1. Restate the decision and constraints.
2. Show normalized inputs with sources.
3. Calculate scenarios.
4. Identify disputed assumptions.
5. Surface dependencies and opportunity costs.
6. Present risk and strategic factors outside the score.
7. List information that could change the ranking.

In the review meeting, ask owners to defend inputs rather than argue over the generated recommendation. Record the decision, rationale, dissent, assumptions, and review trigger. A framework makes disagreement legible; it should not hide who made the choice.

Prioritization is temporal. Reach changes, dependencies slip, incidents occur, and strategy moves. Record the date and source window. Define triggers for rescoring: a major research finding, cost estimate change, new obligation, or experiment result.

After delivery, compare predicted and observed reach, impact, effort, and risk. This calibration matters more than whether the team followed a framework neatly. If impact is consistently overstated or operational effort omitted, update estimation practice.

Do not let AI learn from past prioritization decisions as if they were objective labels. Historical choices reflect power, constraints, and sometimes mistakes. Use retrospectives to examine those patterns explicitly.