RICE and ICE, worked
Score three real candidates with RICE and again with ICE, see why the two rank them differently, and put an honest number in the confidence column.
On this page
A prioritisation framework is a way of making the same argument about every item on a backlog, so that the argument can be checked. RICE and ICE are the two most teams reach for, and on the same three candidates they return different winners. The difference is not noise: one of them has a reach column and the other does not, and the backlog below has a feature that 12,000 people would use.
The three candidates
The product is the support tool this course follows: 12,000 agents across all customers, one team of four engineers, and one quarter ahead of it, which is 12 person-months. Three items are competing for that quarter.
- A, bulk-close for duplicate tickets. About a third of agents work queues where duplicates pile up after an outage; the estimate is 4,000 agents a quarter. The ticket data on duplicate rates exists, so the reach is grounded; what the feature does for those agents is a judgement. Three person-months.
- B, CSV export. Every tenant has a reporting team that asks for it; reach is all 12,000 agents, though most will touch it once a month. One person-month, and the team has built exports before.
- C, AI ticket summaries. A model writes a three-line summary at the top of every ticket. Reach is the 8,000 agents on tenants that have opted into model features. No eval exists yet, no prototype has been in front of an agent, and the team's estimate is six person-months including the eval work.
RICE, column by column
RICE comes from Intercom's product team: reach × impact × confidence ÷ effort. Each column has a scale, and the scale is the point — a column with no scale is an opinion with a decimal point.
- Reach is people (or events) per quarter. Not "everyone eventually": the number who meet the feature in the period the score is for.
- Impact is how much the feature moves the outcome for one of those people, on a fixed scale: 3 massive, 2 high, 1 medium, 0.5 low, 0.25 minimal.
- Confidence is how much of the above you can back with evidence, on three settings: 100% when reach and impact both rest on data, 80% when one of them is a judgement, 50% when both are. There is nothing below 50%. An item you would score at 20% is not a low-confidence item; it is a moonshot, and it leaves the sheet to be prototyped rather than ranked.
- Effort is person-months, the only column where a bigger number is worse.
Scored:
A = 4,000 × 2 × 0.8 ÷ 3 = 2,133
B = 12,000 × 0.5 × 1.0 ÷ 1 = 6,000
C = 8,000 × 3 × 0.5 ÷ 6 = 2,000
B wins, by a factor of three, and B is the boring one. That is not a flaw in the framework. B reaches three times as many people as A, costs a third as much, and is certain. The low impact (0.5) is real — an export a reporting team runs monthly does not change an agent's day — but it is multiplied by 12,000 and divided by 1, and nothing on the sheet survives that. (The exact division is 2,133.3; the trailing digits mean nothing, which choosing the framework comes back to.)
What holds C down is the two columns that are easiest to change. Confidence 0.5 and effort 6 between them divide C's 24,000 of reach × impact by 12. Lift confidence to 0.8 and C scores 3,200, ahead of A; cut effort to 4 and it scores 3,000. Neither is a reason to score C higher today. They are reasons to spend two days on a prototype before the quarter is committed, which is what the one-day prototype is for.
ICE on the same three
ICE is impact × confidence × ease, each scored 1–10, and it was designed for growth experiments where every test hits roughly the same population. Scored on the same three candidates:
A = 6 × 8 × 7 = 336
B = 2 × 10 × 9 = 180
C = 9 × 5 × 3 = 135
A wins. The same team, the same backlog, the same quarter, and a different first item. The cause is visible in the two sheets: ICE has no reach column, so B's 12,000 agents have nowhere to go. B is left with an impact of 2 out of 10, because for any one agent an export does little, and 2 × 10 × 9 cannot catch 6 × 8 × 7. RICE multiplied the small impact by the whole population; ICE never saw the population.
Two other things change on the way from RICE to ICE. Ease is the inverse of effort, so C's six months become a 3 and B's one month a 9, which is the same information in a less useful form — the 1–10 scale has no unit, so a 9 and a 7 cannot be added to see whether the quarter fits. And every column is now 1–10 with no anchors, so two PMs scoring the same item can disagree by a point or two per column without either being wrong.
This is why ICE belongs to experiments and RICE to backlogs. When every candidate is a variant shown to the same traffic, reach is a constant and dropping it costs nothing; a 1–10 impact is enough to order twenty tests. When candidates reach 4,000 people or 12,000, dropping reach drops the largest number on the sheet.
Confidence is the honest column
Reach can be counted and effort can be estimated by the people who will do the work. Impact is a judgement against a fixed scale. Confidence is the column that says how much of the rest is evidence, and it is the one a sheet lies in first.
The three settings are what make it usable. 100% means reach came from data (the 12,000 is a count of seats) and impact from something observed (reporting teams exporting by hand today). 80% means one of the two is a judgement: A's reach is counted, its impact of 2 is a guess about agents nobody has watched use bulk-close. 50% means both are judgements: C's 8,000 counts opted-in tenants, not agents who want summaries, and its impact of 3 is what a slide said.
Below 50% there is no number, by design. A PM who writes 30% is saying the item is a bet, and a bet is not ranked against a CSV export — it is tested. The two-day prototype of C in front of five customers either produces evidence (which moves C to 80% and the score to 3,200) or it does not (and C leaves the sheet, rather than being re-scored at 25% and sitting in the ranking as a 1,000). The exercise Three features, one quarter works that decision through; The RICE sheet that lied shows what a sheet looks like when every confidence cell says 100%.
What a score is for
A RICE score is an argument in a fixed form: this many people, this much each, this much of it evidenced, at this cost. The number ranks; the columns are what you discuss. When sales says C should be first, the answer is not "it scored 2,000" but "its confidence is 50% because no agent has seen a summary, and that is the cheapest column to change — here is the two-day plan." The score ends the argument about ordering so that the time goes on the column that is wrong.
The framework also does not know about the quarter. B (1) plus A (3) uses 4 of 12 person-months; B plus C uses 7. The sheet orders; the capacity check is yours, as is the question of what the other 5 or 8 months do. And it does not know about dates, which is what cost of delay adds. The habit of stating the inputs before multiplying is the same one the back-of-envelope estimate teaches: assumption first, arithmetic second, and a number somebody else can re-derive.
In an interview
The question comes as "here are three features, how would you prioritise", and the interviewer is testing whether you can produce a ranking that someone else could check, then argue with your own ranking. Name the framework, name the scale for each column, score out loud with the arithmetic visible ("B is 12,000 × 0.5 × 1.0 ÷ 1, so 6,000"), and then say which column you trust least and what you would do about it. A candidate who scores and stops has shown they can multiply; a candidate who says "C's score is held down by confidence, and a prototype moves that column for two days of work" has shown they know what the score is for.
The mistake that loses points is scoring confidence by how much you want the feature. Confidence 100% on an AI feature with no eval and no prototype is the most common lie on a RICE sheet, and an interviewer who sees it will ask what evidence the 100% rests on. The scale has three settings and a floor; use them, and say why each item sits where it does.
Check yourself
1. A fourth candidate, D, reaches 2,000 agents at impact 3 with confidence 80% and effort 2 person-months. Where does it rank, and does it change the first item?
D = 2,000 × 3 × 0.8 ÷ 2 = 2,400. It slots between B (6,000) and A (2,133), so the first item is still B. The capacity check: B (1) plus D (2) plus A (3) is 6 person-months, half the quarter, so all three fit and C's six months is the real question again.
2. The team revises C's effort from 6 to 4 person-months after a design cut. Does C overtake A?
C = 8,000 × 3 × 0.5 ÷ 4 = 3,000, against A's 2,133 — yes. But the cut moved a column that was already an estimate by the people doing the work; confidence at 0.5 still says nobody has used a summary. Rank C above A and still spend the two days on the prototype before committing four months.
3. You are scoring twelve growth experiments that all go to the same 10% of traffic. RICE or ICE?
ICE. Reach is the same constant on every row, so the RICE reach column would multiply everything by the same number and change no ranking. A 1–10 impact and a 1–10 ease are enough to order twelve tests, and the precision RICE would add is precision about a constant.