AI Product Feature Triage: How to Decide Which AI Bets Get Resources
TL;DR
Every AI PM has the same problem: 50 promising ideas, 5 engineers, and a quarterly clock. Triage is what turns the inevitable shortage into a strategic advantage. This guide gives you the five-criteria scoring rubric, the AI-specific multipliers (capability fit, vendor risk), and the kill rules that prevent the backlog from becoming graveyard.
The Five-Criteria Triage Rubric
1. User pain (1-5)
How acute is the problem this solves? Evidence from research, support tickets, or behavior data — not from intuition.
2. Strategic fit (1-5)
How well does this advance the quarterly theme and broader strategy? Misaligned wins are still net negative.
3. Effort (1-5, lower better)
Engineering, design, eval, GTM. Honest estimate including the long tail of polish.
4. Capability fit (1-5)
AI-specific. Can current models reliably do this? "Probably" is a 3; "demonstrated in our eval" is a 5.
5. Vendor and infra risk (1-5, lower better)
AI-specific. Multi-vendor support? Cost-at-scale viable? Lock-in concerns? Hidden risks here.
How to Score and Combine
A simple formula: (User Pain × Strategic Fit × Capability Fit) ÷ (Effort × Vendor Risk). The top 3-5 features rise to the top quickly. The rest go to the "not now" list with explicit reason codes.
Score with the team, not solo
Triage scoring is collaborative. The PM proposes; eng challenges effort; design challenges fit. Quality of decisions follows quality of debate.
Re-score quarterly
Capability fit changes as models improve. Effort changes as infra matures. A 'not now' can become a 'ship now' with one model release.
Annotate evidence
Each score has a one-line evidence note. Prevents scoring drift and makes the rationale auditable later.
Avoid 'averaged scoring'
When two reviewers disagree (4 vs 1), don't average to 2.5. Surface the disagreement; resolve it deliberately.
The AI-Specific Multipliers
Generic prioritization frameworks miss what makes AI features uniquely complex. Two multipliers carry the AI-specific weight: capability fit and vendor risk. Both can flip a high-scoring feature into a deferral.
Capability fit makes or breaks
If the model can't reliably do the task at quality, the rest doesn't matter. Test capability before committing engineering.
Vendor risk is real
Single-vendor dependencies, cost trajectories, and lock-in shape long-term viability. Score honestly.
Capability is dynamic
Re-score quarterly. Yesterday's 'not feasible' is tomorrow's 'trivial' after a model release.
Cost-at-scale models
Run the numbers at projected volume. Some features look great at small scale and become unsustainable at 10x traffic.
Make Prioritization Defensible
The AI PM Masterclass walks through real triage with templates and feedback — taught by a Salesforce Sr. Director PM.
Kill Rules That Free Up Capacity
Idea-stage kill rule
If user pain <3 OR capability fit <2, kill before any engineering. The earliest, cheapest place to say no.
Prototype-stage kill rule
If 4-week prototype doesn't hit acceptance threshold, kill. Don't escalate to full build.
Beta-stage kill rule
If 6-week beta shows low repeat usage, kill. Don't graduate to GA on hope alone.
Live-feature retirement
Quarterly review: features below usage threshold either get focused investment or get sunset. Drift kills focus.
Triage Anti-Patterns
"Everything is a 5"
When everything is high priority, nothing is. Force diversification of scores; calibrate the team to the rubric.
Skipping capability fit
Features that the model can't actually do reliably soak up engineering. Test capability before committing.
Loudest stakeholder wins
If exec mood determines triage, you have stakeholder management, not prioritization. Defend the rubric.
'Quick wins' that aren't
AI quick wins often become long projects. Be skeptical of low effort scores until proven by prototype.