ICE Scoring

There are always more things to build than slots to build them in. Five items fit in the next quarter and thirty are sitting in the list. Yet the meeting where the order gets decided often has no standard at all, and when that happens the loudest voice sets it. The request sales pushed hardest moves to the front, and the competitor feature the CEO saw yesterday moves ahead of that. Neither move had anything to do with how much either one was worth.

The real damage isn't the order, it's that nothing is written down. Ask why A went first and nobody can explain it, and three months later the same candidates come back and the argument starts over. The people whose requests slipped never heard why, so they argue louder next time.

Scoring frameworks exist to break that argument into axes. Instead of "A is more urgent," someone says "A is an 8 on impact and a 4 on confidence," and it becomes visible which axis the disagreement actually lives on. Among those frameworks, ICE sits at the light end. With only three axes, ten candidates can be scored in half an hour — which avoids the failure where the scoring itself becomes such a chore that the team quietly stops doing it.

Rate each candidate from 1 to 10 on three axes and multiply the three numbers. In practice it runs in this order.

1. Collect the candidates into one list. Scattered items can't be compared, so a place that holds everything — a product backlog — has to exist first.
2. Match the sizes. Put "reword a button" and "replace the payment system" in the same table and any framework produces nonsense. Split day-sized and month-sized work into separate tables and score each one on its own.
3. Score impact: how much changes if this works. Something that could double signup conversion is a 9; fixing wording that mildly annoyed a few people is closer to a 2.
4. Score confidence: how much evidence there is that the impact will actually show up. Already verified by an A/B test is a 9; a hunch from a meeting room is a 3. This axis is what stops a plausible-sounding idea from sailing to the top on its own.
5. Score ease: how easily it can be done, with easier scoring higher. A day of work is a 9; two months plus an external integration is a 2. The people who build it supply this number.
6. Multiply. Impact 8, confidence 6, ease 7 comes to 336. The figure means nothing by itself; what means something is that the candidate next to it, scored the same way, came out at 90. Multiplying rather than averaging keeps a near-zero axis from being rescued by high scores elsewhere.
7. Don't ship the resulting order as-is — filter it once. Anything with a date in a contract, anything with legal exposure, and anything blocked behind other work moves forward regardless of score. For each candidate you defer, write one line about why it isn't happening this time.

It works best when the same people score repeatedly. The scale lives in each person's head, so there is no guarantee that Mia's 7 and Theo's 7 are the same size. But when the same three people keep scoring quarter after quarter, their scales converge, and while the absolute numbers stay untrustworthy the relative order becomes useful.

It also fits a moderate number of candidates. Ten to twenty is a half-hour session. Push past fifty and the later scores get assigned carelessly, which is worse than not scoring at all. That lightness is the entire reason the framework exists, so adding axes or weights takes away the reason you picked it.

There are three real limits. First, reach is missing. Something enormously good for 1 percent of users and something mildly good for 80 percent come out at similar scores. If the number of people affected varies widely across candidates, a different framework serves better. Second, because the three values are multiplied, a single point swings the result hard. Whether confidence is a 4 or a 5 flips the ranking, and that call is itself a judgment. Third, scores rise when the person scoring has a stake in the outcome. Ask sales to rate the impact of a request sales brought in and it rarely lands below an 8. That's why nobody scores alone, and why product, engineering, and sales sit down together.

"Let's do one ICE pass and cut the bottom ten from this quarter."

Not a proposal to rank things precisely, but to clear out the obvious no's quickly.

"I have impact at 8 too. I want to understand why you put confidence at 3."

Naming the axis where the disagreement lives, which narrows the discussion.

"This is first on ICE, but it drops once you account for reach."

Pointing out that the axis this framework leaves out would change the answer here.

At Pagedeck, Mia puts twenty candidates in a table before quarterly planning and gets Theo and Sofia in a room to score them within the hour. Theo supplies ease, and Sofia supplies impact for anything that came in through sales. If one person scored all three axes, that person's instincts would simply become the order.

Even with scores in hand, the table doesn't become the plan. It gets used to clear out the bottom half, and the six or seven left at the top go back on the table for a conversation about where the company is trying to go this quarter. Anything within about 30 points is treated as a tie, and ties get settled by direction rather than arithmetic.

The most common mistake is treating the score as the decision. A number comes out, so working straight down the list looks reasonable. But all three values were assigned by feel, which makes the gap between 336 and 320 meaningless. What the framework is genuinely good at is removing the candidates that clearly don't belong. Teams that follow the ranking literally start optimizing for the ranking — someone writes an 8 next to their own confidence — and the numbers stop working as evidence.

Mistaking simple arithmetic for precision is the other one. One multiplication, a tidy three-digit result, and it looks exact. It's the opposite: three axes means a lot of discarded information in exchange for finishing fast. If that trade isn't acceptable, RICE adds an axis, MoSCoW draws a boundary instead of a ranking, and a value-effort matrix separates the big blocks with no calculation at all.