RICE is a good default. Reach times Impact times Confidence, divided by Effort. It forces you to quantify and it kills pet features. But for AI products and habit products, vanilla RICE consistently mis-ranks the work. I learned this scoring the Listen2RE backlog, where the framework kept telling me to build the wrong thing.
The two blind spots
Blind spot one: habit impact. RICE's "Impact" score treats a feature that creates a daily return loop the same as a feature that delivers a one-time hit of value. For a product whose entire thesis is a daily habit, that is a category error. A streak mechanic with tiny reach can outweigh a flashy feature with broad reach, because it compounds.
Blind spot two: AI-readiness. Every AI feature carries hidden effort that RICE's "Effort" rarely captures: the evaluation harness, the prompt iteration, the human-review fallback, the cost-per-call. Two features with identical engineering effort can have wildly different true cost once you price in the AI tax.
The fix: a habit multiplier and an honest effort number
I keep RICE's bones and add one multiplier:
Score = (Reach × Impact × Confidence ÷ Effort) × Habit-impact
Habit-impact runs from 0.5 to 2.0. Anything that strengthens the core return loop scores above 1.0. Anything orthogonal to the loop scores below it. And I fold the full AI tax into Effort — if a feature needs an eval set and a human-review path, that is real person-weeks, so it goes in the denominator.
A real example
On Listen2RE, three candidates:
| Feature | Plain RICE rank | With habit multiplier |
|---|---|---|
| Content-library expansion | high (broad reach) | killed (0.6x, weak loop) |
| Streak anchored to morning session | mid | shipped first (1.7x, tiny effort) |
| Per-learner topic sequencing | mid | promoted (1.8x, high loop) |
Library expansion was the most-requested feature and the highest plain-RICE score. It was also the wrong build: high effort, weak habit impact, competing with the daily loop for attention. The multiplier surfaced that. The streak anchor — small reach, trivial effort, enormous loop impact — jumped to the top, exactly where it belonged.
For B2B2C, swap the formula entirely
When I prioritized the ERP platform, reach was the wrong numerator because reach is not adoption. A feature can reach every campus and be adopted by none. There I used (Value to daily user × Adoption likelihood) ÷ Rollout cost, which deliberately sequenced the unglamorous workflow features the clerk needed ahead of the dashboards the director asked for — because renewal follows the clerk, not the buyer.
The takeaway
Frameworks are lenses, not laws. RICE is a fine lens until your product's value comes from compounding behavior or your effort hides an AI tax. Then you adjust the lens. The discipline that matters is not the formula — it is being honest about what actually drives the outcome, and weighting for that.