A spreadsheet can make uncertain guesses look precise. AI can make that spreadsheet faster. Neither tells you which problem deserves the team's attention unless the underlying evidence and tradeoffs are clear.
Use AI to improve the quality of a prioritization discussion: organize inputs, expose assumptions, compare alternatives, and show what would change the recommendation.
Start with a goal and a constraint
“Which feature should we build?” is incomplete. A more useful question is: “Which opportunity is most likely to improve first-week activation for new team administrators, within the engineering capacity available this quarter?”
The goal identifies the outcome. The segment prevents unrelated customers from being combined. The capacity constraint forces realistic tradeoffs. Include mandatory obligations separately; a required security fix should not disappear because a growth idea earns more points.
Compare opportunities before solutions
Suppose the team is considering an AI assistant, a simpler invitation flow, and better import guidance. Describe the underlying customer obstacles first. The assistant may be one way to address several different problems, each supported by different evidence.
Create a table for the opportunity, affected segment, evidence, expected behavior change, alternatives, effort range, dependency, and biggest uncertainty. Include “do nothing yet” and a smaller manual or content-based intervention.
Try this prompt:
Compare these opportunities against the stated goal. Use only supplied evidence. Separate measured values from estimates. Do not invent reach, impact, confidence, or effort. Identify duplicated problems, missing alternatives, dependencies, and assumptions that could reverse the ranking. Recommend the next evidence-gathering step for each weakly supported option.
If you score, make assumptions inspectable
A method such as RICE can organize a conversation: reach × impact × confidence ÷ effort. It remains a model built from your definitions and estimates. Use a consistent time period and effort unit. Define what each impact level means. Keep confidence tied to evidence quality rather than presentation quality.
Ask AI to test ranges rather than invent exact inputs. If an option wins only when optimistic impact and low effort are combined, the ranking is fragile. If two options exchange places under plausible assumptions, gathering evidence may be more valuable than arguing over decimals.
Do not treat strategic, legal, or safety obligations as just another compensating factor in a weighted average.
Make the tradeoff visible
For each serious candidate, write what choosing it delays. A roadmap conversation becomes more honest when the team can see the opportunity cost. Include implementation and operating effort: support, review, training, infrastructure, and maintenance.
For an AI feature, estimated development speed should not hide ongoing model costs or human correction effort. A simpler feature may deliver more net value even when its demo is less impressive.
Finish with a decision and a review trigger
Record the selected opportunity, evidence, rejected alternative, owner, next test, and conditions for revisiting the choice. A useful review trigger might be a change in a key dependency or evidence that the suspected obstacle is uncommon.
A roadmap is an expression of current judgment. Better evidence should be allowed to change it. AI is valuable when it helps the team see those conditions clearly before committing.
Try it: Use the Prioritization Decision Brief. For AI-specific choices, read the existing guide on whether a feature should use AI.