1. LandOffer
  2. Blog
  3. All articles
  4. Use RICE Prioritization Without Treating the Score as a Decision

Product & Project Delivery

Use RICE Prioritization Without Treating the Score as a Decision

Work through three complete RICE estimates, confidence sensitivity and overlapping user cohorts. Keep the reproducible score separate from evidence quality, staffing, release approval and the actual allocation decision.

9 min readLandOffer team

Use RICE to expose the assumptions behind a priority, then make a separate decision about what to fund. Compare reach over the same period, define the impact scale, attach evidence to confidence, and include the whole team's effort. Recalculate the ordering when a disputed assumption changes. The highest score identifies a candidate to discuss. It does not validate the user problem or authorize delivery.

This LandOffer.ai guide reviews original method documentation checked October 27, 2026 and develops a fully fictional three-feature case. We executed the arithmetic and sensitivity checks locally. The analytics, research, estimates, people, and approvals inside the case are invented; no customers were studied, roadmap changed, or product experiment run.

An editorial model-making workshop shows colleagues weighing three project concepts while inspecting the assumptions behind them.

Establish one comparison before calculating anything

Intercom's original RICE article, by Sean McBride, defines reach, impact, confidence, and effort. It measures reach over a stated period and combines the factors as reach × impact × confidence ÷ effort. Its example impact ratings range from 3 to 0.25; confidence uses 100%, 80%, and 50%; effort is expressed in person-months across the team.

Keep those definitions attached to your worksheet. A row counted in users per month cannot be compared directly with one counted in transactions per quarter. Nor does a team estimate of three person-months mean three elapsed calendar months. Parallel work, part-time availability, dependencies, and specialist constraints affect duration separately.

Choose the outcome the candidates are intended to help. Otherwise, a high impact rating might mean revenue to one person and easier preparation to another. A shared label conceals the disagreement rather than resolving it.

For our fictional catalog-publishing workspace, the intended outcome is help catalog coordinators prepare complete batches for editorial review. Product lead Orin owns the comparison, researcher Beata owns evidence descriptions, engineer Clive owns implementation assumptions, and service owner Lotte decides the bounded allocation. This is an internal product task, not a portfolio case or a pitch for a preferred feature.

Give each input a unit, a reason, and a limitation

The team uses a November 1–30 planning window and counts unique coordinators expected to encounter each proposed change. The synthetic reference population is 600 active coordinators. It considers A, warnings for missing edition metadata; B, copying a prior catalog setup; and C, applying an editorial note to several selected entries.

Orin rejects an initial note that called B's reach “2,400.” That number represented setup actions, including repeated actions by the same people. The corrected estimate is 400 coordinators, making it comparable with A's 240 and C's 160. Keeping the mistaken 2,400 action count would give B a score of 600 and put it first, even with low confidence. These are stipulated exposure estimates, not measured adoption forecasts for real users.

For impact, A receives 2 because the team expects it to address a more consequential completeness problem. B receives 1 for avoiding repetitive preparation. C receives 0.5 for a smaller notes task. Those are judgments on the declared outcome; 2 does not mean a proven doubling of completion or a two-percentage-point improvement.

Confidence records the support for the estimates as a package. The team assigns A 80%, B 50%, and C 100%. Here, 100% means its highest support category, not certainty that C will work. A precise reach count cannot compensate for weak evidence that the proposed interaction produces the intended benefit.

Effort includes product, design, engineering, and verification. Clive records A at 3 person-months, B at 2, and C at 0.5. The smallest item still needs review and release work. A zero denominator would make the calculation undefined, not identify a free project.

Worked asset: the completed three-feature estimate

The following evidence cards are author-written scenario inputs. They demonstrate what an inspectable estimate record contains; they are not summaries of interviews or logs collected for this article.

Candidate and estimate Fictional support recorded Limitation retained
A: missing-metadata warnings. Reach 240; impact 2; confidence 80%; effort 3 person-months Card E1 identifies the eligible coordinator cohort. E2 describes incomplete review submissions. Clive's estimate includes rule evaluation and review work The proposed warnings may interrupt valid exceptional entries; interaction and exceptions remain untested
B: copy prior setup. Reach 400; impact 1; confidence 50%; effort 2 person-months E1 identifies repeat-setup coordinators. E3 records requests for reuse, with a draft implementation breakdown Requests do not establish whether copied settings remain correct or whether coordinators would inspect them
C: bulk editorial notes. Reach 160; impact 0.5; confidence 100%; effort 0.5 person-months E1 identifies the notes cohort. E4 describes a narrow existing task and a reviewed small-change estimate Broad rollout behavior and accidental application to unintended entries remain untested despite the high support category

The exact effort breakdown is also retained. A comprises 0.5 product, 0.5 design, 1.5 engineering, and 0.5 verification person-months. B comprises 0.25, 0.25, 1, and 0.5 respectively. C comprises 0.1, 0.1, 0.2, and 0.1. These sum to the denominators used below; none is an actual staffing commitment or time measurement.

Candidate Substitution using decimal confidence Calculated score Base ordering
A 240 × 2 × 0.80 ÷ 3 128 Second
B 400 × 1 × 0.50 ÷ 2 100 Third
C 160 × 0.5 × 1.00 ÷ 0.5 160 First

We ran these calculations with Python's Decimal arithmetic. The result reproduces the stipulated formula; it does not verify any input. C's small effort denominator explains why a low-impact idea can lead this comparison. “First” means highest on this model, not greatest total benefit, best strategy, or approved release.

The cohorts can overlap. In the authored fixture, A covers coordinator identifiers 1–240, B covers 201–600, and C covers 1–160. Their counts sum to 800, but their union contains only 600 people. Adding row reach would count some coordinators repeatedly; adding scores would likewise require assumptions about overlap and combined effects that this model has not supplied.

Review confidence without turning it into persuasion

Atlassian's product-discovery handbook places structured prioritization beside qualitative judgment and continuing learning. It describes confidence as a prompt to discuss whether further product or technical validation is needed. Use that discussion to identify the weak claim, rather than raising a percentage because the feature's sponsor wants a higher rank.

Beata's objection to B is specific: knowing that 400 coordinators repeat setup does not establish that copying a previous configuration is safe or useful. A request for “reuse” might mean a checklist, a default, or a fully copied batch. The estimate currently refers to the last option, which could preserve outdated settings.

The appropriate response is to retain that gap and investigate the proposed behavior. Raising B from 50% to 80% merely because more colleagues agree would change the score without adding evidence. Agreement about priority and support for an impact estimate answer different questions.

The confidence field also compresses several uncertainties. Beata records that B's reach support is stronger than its interaction support. If new work changes only the effort estimate, update effort rather than hiding that change inside confidence. Keeping the explanation prevents the same uncertainty from being discounted twice without anyone noticing.

Worked asset: change confidence and expose the ranking boundary

The sensitivity table changes only the listed assumptions. All other inputs remain at their base values. These are alternative calculations, not research results or newly earned confidence ratings.

Scenario A score B score C score Ordering and interpretation
Base: A 80%, B 50%, C 100% 128 100 160 C, A, B
B rises to 80% only 128 160 160 B and C tie; no automatic tie-break
B rises to 100% only 128 200 160 B, C, A; a high-support scenario, not established evidence
A falls to 50% only 80 100 160 C, B, A; A's position depends on its evidence support
B rises to 80% and C falls to 50% 128 160 80 B, A, C; explicitly a two-input scenario

For B, holding reach, impact, and effort fixed gives a score of 200 × confidence. It ties C's base 160 at confidence 0.80. It ties A's base 128 at 0.64. The latter is an algebraic threshold, not a recommendation to invent a 64% confidence category or claim a measured probability.

This helps Orin focus the next discussion. The initial B-versus-C ordering changes if the evidence supports moving B into the team's medium-confidence category. Beata therefore identifies B's copied-settings assumption as a useful question to investigate. The calculation cannot specify the right study, prove that the study will resolve it, or guarantee a rank change.

An effort change could matter just as much. If C's estimated effort doubles to 1 person-month while its other inputs remain fixed, its score becomes 80. A then leads at 128, followed by B at 100. If verification work was omitted, revise the effort estimate openly. Keeping an earlier rank can feel reassuring while leaving the missing work unfunded.

Movable project models and their evidence ribbons show how changing an assumption can reorder the comparison without approving any project.

Record the funding decision separately from the score

Intercom explicitly allows work outside score order for dependencies or essential requirements. Preserve that exception visibly. Do not inflate a mandatory item until it happens to win the ranking; explain the requirement and which authority accepted the trade-off.

Lotte's completed fictional planning record keeps the base ordering C, A, B. She authorizes a bounded C build-and-verification allocation of 0.5 person-month and a separate B investigation allocation of 0.5 person-month within a provisional 3-person-month discretionary envelope. The remaining 2 person-months are uncommitted. This is not a claim that the entire envelope is interchangeable or already staffed.

C's allocation uses its recorded breakdown. B's investigation separately reserves 0.2 product, 0.2 design/research, and 0.1 engineering person-months; it does not authorize the 2-person-month feature. Lotte and the relevant functional owners accept these fictional allocations before the example proceeds. No actual personnel or calendar commitment was made here.

C must make the selected entries visible before applying a bulk note and preserve a cancellation route. Production release requires a separate saved acceptance decision. Those are proposed acceptance requirements, not observed behavior. A remains uncommitted while Clive checks how valid metadata exceptions would be handled. B's original build score remains 100 until evidence justifies revising its inputs; funding investigation does not raise it automatically.

Lotte records the revisit trigger: bring the C interaction findings and B copied-settings evidence to the November 12 review, with negative or inconclusive findings retained. The record ends at allocation. It does not invent a successful prototype, recovered staff time, increased completion, or a changed customer outcome.

Keep the model reproducible when evidence changes

Store the dated input version, unit definitions, evidence cards, formula, effort scope, and decision note together. A later reviewer should be able to reproduce the old ordering before replacing an estimate. Changing the cell without preserving why it changed makes a disagreement harder to reconstruct.

The Confluence RICE template uses its own rating descriptions, including 1–5 impact and effort scales. That differs from this case's Intercom-style impact weights and person-month denominator. Copying values between the two without remapping definitions would create a different model. No Jira or Confluence configuration was tested here.

For your next review, resolve the reach unit first, name the least-supported estimate, and calculate a plausible change openly. Save the ranking beside the actual allocation, deferred work, owner, and revisit trigger. At a tie, record the reason for choosing a candidate or funding more investigation; do not alter an input solely to remove the tie. More workplace guides are available from LandOffer.ai; the useful result here is a decision whose assumptions another person can challenge.

Sources and Further Reading