RICE Framework with AI: Score Features on Evidence, Not Guesses
AI doesn't change the RICE formula. It changes where the inputs come from. Here is the formula, a worked example and a four-step workflow for scoring Reach, Impact, Confidence and Effort from evidence.
Share
The RICE framework scores a piece of work as (Reach × Impact × Confidence) ÷ Effort. Intercom built it and Sean McBride published it in 2018. AI does not change the formula; it changes where the inputs come from. Used well, AI pulls Reach from usage data, Impact from clustered customer feedback and Confidence from the evidence you actually hold, so the score stops being a dressed-up guess.
Key takeaways
RICE uses fixed scales: Impact runs 0.25 to 3, Confidence is 100%, 80% or 50%, and Effort is in person-months (Intercom, 2018).
The weakest inputs are Impact and Effort. Only one third of ideas tested at Microsoft improved the metric they targeted (Kohavi et al., 2012), and in a classic planning-fallacy study students took 55.5 days on tasks they predicted would take 33.9 (Buehler et al., 1994, via Wikipedia).
PMs know the inputs are thin: 35% of those who struggle most with prioritisation say the main problem is getting the right data (Craft.io, 2023).
AI is already in the workflow: 94% of 379 enterprise product professionals surveyed by Productboard use AI daily or often (Productboard, 2025).
AI should gather and cite evidence for each input. People should still own the trade-offs and the final order.
What is the RICE framework?
RICE is a scoring model for comparing product work. Intercom's team built it "from first principles" because existing methods made it hard to compare different ideas consistently, and McBride's write-up remains the reference (Intercom, 2018). Each letter is one question:
Cluster feedback, count accounts per theme, quote the source
Confidence
How sure are we about the other estimates?
100% high, 80% medium, 50% low
Whatever evidence backs Reach and Impact
Audit which inputs have evidence and which are assumptions
Effort
How much team time will it take?
Person-months (0.5 for under a month)
Past tickets of similar size
Find comparable past work and its real duration
The formula multiplies the three value factors and divides by effort, giving "total impact per time worked". McBride is clear that it "shouldn't be used as a hard and fast rule": a low-scoring item can go first if it is a dependency or table stakes for key customers. For a deeper guide to each input, see the RICE prioritization framework guide.
How do you calculate a RICE score?
Multiply Reach × Impact × Confidence, then divide by Effort. Intercom's article gives inputs for three example projects but stops before the final scores. Here they are, worked through:
Project
Reach (per quarter)
Impact
Confidence
Effort (person-months)
RICE score
Project 1
450
3
100%
2
675
Project 2
2,000
1
80%
4
400
Project 3
800
2
50%
1
800
Project 1's Reach comes from 500 customers a month reaching a point in the signup funnel, 30% of whom choose the new option, over 3 months: 500 × 0.3 × 3 = 450.
Notice what wins. Project 3 tops the list with only 50% confidence, because it is cheap. That is the score doing its job: it flags an inexpensive bet worth trying. It is also a warning. A 50% confidence item should usually earn a small test before it earns a full build slot. Run your own numbers in the free RICE calculator.
Why are most RICE scores guesses?
The formula is simple. The inputs are where teams cut corners.
The confidence crisis
Confidence is meant to discount weak estimates. In practice, people pick 80% because 50% sounds unprepared. Itamar Gilad's Confidence Meter makes the point plainly: self-conviction and anecdotes should count for almost nothing, and confidence should rise only with evidence such as tests and data (Itamar Gilad).
Impact is usually overestimated
At Microsoft, only one third of ideas tested in controlled experiments improved the metric they were designed to improve (Kohavi et al., 2012). If your team scores most ideas at Impact 2 or 3, your scale is drifting upward.
Effort is usually underestimated
The planning fallacy is well documented. In Buehler, Griffin and Ross's 1994 study, students predicted their thesis would take 33.9 days; it took 55.5 days on average, and only about 30% finished within their own estimate (Wikipedia summary of the study). Engineers estimating under roadmap pressure face the same bias.
Put the three together and a RICE score can be wrong by a multiple, not a margin. The fix is better inputs, not more decimal places.
How can AI improve RICE scoring? A four-step workflow
AI is good at reading large volumes of messy evidence: tickets, call transcripts, analytics exports, past issues. That is exactly the work most PMs skip when scoring. Use it for these four steps.
Step 1: Ground Impact in clustered customer feedback
Export the last 90 days of support tickets, call notes and feedback. Ask an AI model to cluster them by underlying problem, count distinct accounts per cluster and pull two or three verbatim quotes each. Then map clusters to Impact:
Blocks a core workflow or appears in churn reasons → candidate for 2 or 3.
Repeated friction with a workaround → 1.
Occasional request, no consequence named → 0.5 or 0.25.
Check every cluster against the quotes. AI groups by wording, and two customers can use different words for the same problem. The support ticket analysis template gives a manual baseline to compare against.
Step 2: Validate Reach with product analytics
Replace "lots of users want this" with a count. Have AI draft the analytics query (for example, distinct accounts that opened the reporting page in the last quarter), then run it yourself. Count people who hit the affected workflow, not your total user base. A product analytics check before scoring catches the common error of counting every account as reachable.
Step 3: Cross-check Effort against past delivery
Give AI the scope and a list of past tickets or epics with their actual durations. Ask for the three most similar pieces of work and how long they really took. If the team estimates 1 person-month and similar past work averaged 2, use the evidence or write down why this time differs. This is reference-class thinking: compare with what happened, not with the best case.
Step 4: Run a confidence audit
Ask AI to list, for each input, the evidence behind it and label it observed, inferred or assumed. Then set Confidence with a rule, not a mood. The ladder below is a suggested starting point, combining Intercom's three levels with Gilad's evidence-first idea:
Evidence behind Reach and Impact
Confidence
Measured usage plus a test or prototype result
100%
Usage data plus repeated, sourced customer evidence
80%
A few anecdotes, a stakeholder request or a competitor move
50%
Only the team's belief
Below 50%: run a test before scoring
Copyable confidence-audit prompt
You are auditing a RICE score. For the item below, list each input
(Reach, Impact, Confidence, Effort). For each one:
1. Quote or link the evidence provided that supports it.
2. Label it OBSERVED (measured or directly seen), INFERRED
(reasoned from evidence) or ASSUMED (no evidence given).
3. Flag any number that has no evidence.
Do not invent evidence. If none is provided, say "none provided".
Then recommend a Confidence level (100%, 80%, 50%) using this rule:
[paste your confidence ladder]
Item:
Proposed inputs:
Evidence (tickets, call notes, analytics output, past tickets):
What can't AI do in a RICE score?
AI makes the inputs cheaper to gather. It does not make them true. Three limits matter.
It can be confidently wrong. In the 2025 Stack Overflow survey, 46% of developers distrusted the accuracy of AI output and 33% trusted it; 66% named answers that are "almost right, but not quite" as their top frustration (Stack Overflow, 2025). Treat every AI-suggested number as a draft with a source attached, or discard it.
It counts mentions, not importance. Ten tickets from free users and one from your largest customer are not comparable, and a model will not know your segment priorities unless you tell it.
It does not know your strategy. RICE ranks items against each other. Whether any of them serve this quarter's outcome is a separate question; why prioritization frameworks fail covers that gap.
When should you ignore the RICE score?
RICE has a built-in tilt: dividing by Effort rewards small, safe work. Follow scores blindly and you get the incremental trap, a roadmap of minor improvements while larger bets never qualify because their Reach is low today and their Effort is high.
Override the score, and write down why, when:
It is a dependency. Something higher up the list cannot ship without it. Intercom's article names this case.
It is table stakes. Key customers will not buy or stay without it, whatever its score.
It is a deliberate strategic bet. Some teams reserve a fixed share of capacity, for example 20%, for bets that scoring would never pick. That number is a common convention, not a research finding; choose your own and keep it visible.
The evidence is too thin to score. Below 50% confidence, the right next step is a test, not a ranking.
Log each override so the reasoning survives the meeting. A decision log works well for this.
Copyable AI-assisted RICE worksheet
Item:
Outcome it serves this quarter:
REACH
Definition (who, which workflow, which period):
Number:
Source (query, dashboard, export):
IMPACT (3 / 2 / 1 / 0.5 / 0.25)
Score:
Feedback cluster(s) and account count:
Two sourced quotes:
CONFIDENCE (100 / 80 / 50%)
Score:
Evidence level per the ladder:
What would raise it:
EFFORT (person-months)
Team estimate:
Comparable past work and actual duration:
Final estimate and reason:
RICE score:
Override? (dependency / table stakes / strategic bet / too thin to score):
Owner: Review date:
Common mistakes with AI-assisted RICE
Pasting a feature list into a chatbot and asking for scores. Without evidence, the model fills gaps with plausible numbers. You get false precision faster.
Letting AI set Confidence about its own output. The confidence audit checks the evidence you supplied, not the model's certainty.
Measuring Reach as total users. Reach is the number of people who hit the affected workflow in the period.
Scoring once and never checking. After launch, compare predicted Impact with the result. That feedback is the only way your scale gets calibrated.
Hiding overrides. If the CEO's pick jumps the queue, record it as an override with a reason rather than nudging the inputs until the score agrees.
Where Specky fits
Specky's prioritisation view scores a backlog with RICE, ICE or weighted criteria. Its AI scoring suggests all four RICE inputs from the context in your Product Graph, such as customer calls, feedback, tickets and product data, and adds a one-sentence rationale for each item. Every input stays editable and the ranking stays your call. For a broader method that covers more than RICE, read how to prioritize product features.
Frequently asked questions
What is the RICE formula?
The RICE formula is (Reach × Impact × Confidence) ÷ Effort. Reach is people affected per period, Impact is scored 0.25 to 3, Confidence is 50%, 80% or 100%, and Effort is in person-months. For example, Reach 800, Impact 2, Confidence 50% and Effort 1 gives a score of 800.
What is a good RICE score?
There is no universal good score, because Reach depends on your user base and period. RICE scores are only meaningful relative to other items scored by the same team, with the same definitions, in the same round. Treat items within about 20% of each other as a tie and decide on strategy or evidence instead.
How do you set Confidence in RICE?
Tie Confidence to evidence, not feeling. Use 100% when measured usage and a test support the estimates, 80% when you have data plus repeated customer evidence, and 50% for anecdotes or stakeholder requests. Intercom's scale treats anything below 50% as a moonshot, which usually means you should test before you score.
Can ChatGPT or another AI tool calculate RICE scores?
AI can do the arithmetic and, more usefully, gather evidence for each input: clustering feedback for Impact, drafting analytics queries for Reach and finding comparable past work for Effort. It should not invent the inputs. Give it your evidence, ask it to cite what supports each number, and keep the final ranking with the team.
What is the difference between RICE and ICE?
RICE adds Reach and measures Effort in person-months, so it asks how many people a change touches. ICE rates Impact, Confidence and Ease on rough scales, usually 1–10, and is faster to apply. RICE suits mature products with usage data; ICE suits quick growth experiments. The AI use case prioritization scorecard covers a third option for AI ideas.
How do you measure Reach in RICE?
Count the people or accounts that will hit the affected workflow in a fixed period, usually a quarter. Pull it from product analytics where possible. Intercom's example multiplies 500 customers a month reaching a signup step by the 30% who choose the option by 3 months, giving a Reach of 450 per quarter.
GDPR-aligned. EU workspace data is stored in Frankfurt, Germany. GDPR and EU data Built in Vienna, Austria. A product of influence solutions fm GmbH (FN 675374y). Imprint 14-day trial. Card required, and nothing is charged until day 14. 30-day refund after subscribing. See pricing