Prioritisation
Make defensible decisions about what to build — using frameworks, not instinct. Adapted for AI product work.
Four frameworks, four jobs. Use them in layers — not as alternatives.
RICE
Reach × Impact × Confidence ÷ Effort
Comparing multiple features against each other. Best when you have data on users and effort.
Example
Feature: Intent auto-routing in support Reach: 8,000 (tickets/month affected) Impact: 3 (medium — 20% routing improvement) Confidence: 60% (model hasn't been trained yet) Effort: 4 person-months RICE = 8000 × 3 × 0.6 ÷ 4 = 3,600
AI-Specific Twist
Confidence is the most important variable in AI. A feature that requires a new model architecture should have confidence ≤50% until a spike confirms feasibility. Never score AI features at 80%+ confidence before data and model exploration.
Common Trap
Treating all effort estimates as equal. A 3-month ML engineering story is not the same as a 3-month frontend story — data uncertainty and model iteration add invisible multipliers.
RICE is the most common ranking framework. The confidence variable is where AI PMs most often go wrong.
RICE Calculator
RICE Score
3,000
5,000 × 3 × 60% ÷ 3 = 3,000
Confidence Calibration for AI
Similar model, proven data, done this before
Existing data, new model architecture — spike completed
New data source, new approach — spike in progress
Unvalidated hypothesis — no spike yet
Interview Line
"For AI features, I cap confidence at 60% until a feasibility spike validates the approach. RICE scores look very different when the model hasn't been proven yet — and that's the honest number to report."
Kano categorizes before RICE ranks. AI features shift categories faster than traditional software — re-run this model every 6–12 months.
Basic Needs
What users expect as table stakes. Absence causes dissatisfaction; presence is invisible.
AI Examples
Model output is never confabulated without a visible disclaimer
PII from one user never surfaces to another
System degrades gracefully when model is unavailable
Response latency stays under 3s for 95% of requests
Prioritization Strategy
These ship first, always — regardless of RICE score. No framework justifies deferring a security or reliability basic need.
Decision Sequence
Five signals that should change your priority decision — even when RICE says otherwise.
Data is not ready
The feature depends on data that does not exist, is too small, or is too dirty to train on reliably.
PM Move
Deprioritize the model-dependent story. Prioritize a data pipeline or labeling story instead — this unblocks all future AI work.
Wrong Move
Proceeding with model training on "good enough" data. The model will underperform, stakeholders will lose confidence, and you'll spend 2 sprints explaining why the AI doesn't work.
Which prioritization framework do you use, and how do you choose between them?
I use different frameworks at different stages of the planning cycle — they complement each other rather than compete. Kano runs quarterly to categorize the backlog by user expectation type: basic needs, performance features, and delighters. This sets the hierarchy. Then RICE runs within each category to rank by impact-per-effort. MoSCoW handles release scoping for each sprint, forcing explicit decisions about what ships now. OKRs anchor everything — before any feature makes the backlog, it must trace to a Key Result. For AI specifically, the critical addition before RICE is a feasibility check. If data or model approach is uncertain, I add a spike. RICE scores for unvalidated AI features should have confidence capped at 50–60% until the spike completes.
Common Trap
Choosing one framework and applying it rigidly. RICE alone will deprioritize basic needs in favor of high-reach features. Kano alone gives you no ranking within categories. The frameworks are layers, not alternatives.
How do you prioritize when you do not know whether the model will work?
High uncertainty is the default state in AI product work, not the exception. My approach is to treat unknown feasibility as a signal to run a spike before committing sprint capacity. A spike is a time-boxed investigation — usually 3–5 days — with a binary output: go or no-go. If go, we add the full feature story with an updated RICE confidence score. If no-go, we document the finding and either shelve the feature or explore an alternative approach. For RICE scoring with uncertain AI features, I calibrate confidence explicitly: 80% if we have done this before with similar data; 60% if the data exists but the model is new; 40% if both model and data are unvalidated. This prevents over-committing to features that look attractive on paper but have hidden technical risk.
Common Trap
Treating "we'll figure it out in the sprint" as a valid response to uncertainty. That transfers the risk to the engineering team without giving them the space to de-risk it properly. Run the spike.
How do you handle a senior stakeholder who wants to reprioritize based on a competitor launch?
I start by taking the concern seriously — competitor launches can legitimately shift the Kano category of a feature from delighter to basic need, which changes its priority independent of RICE. My response is to run a fast Kano re-evaluation: has the competitor launch made our planned feature table stakes for users? If yes, I re-prioritize it and surface the trade-off: what gets deferred? I prepare a one-slide view showing the current sprint plan, the proposed change, and what ships later as a result. If the feature is still a delighter after the re-evaluation, I hold the position with data: "The competitor shipped X, but our users have not told us they need this yet. I am adding a 2-week validation study to confirm before we reprioritize." This buys time and shows the exec we are taking it seriously.
Common Trap
Immediately agreeing to reprioritize to avoid conflict. This trains stakeholders to use urgency as a prioritization mechanism, which breaks the roadmap process for everyone.
How do you say no to a feature request without damaging the relationship?
The goal is never to say no to the request — it is to say no to the timing or the form. My approach: "That is a strong idea for [outcome]. Right now, our sprint capacity is committed to [current priority] because it moves [KR]. I am adding your request to our discovery backlog. Here is what I need to justify reprioritizing it: [evidence or experiment result]. Can we schedule 30 minutes to define what a test of this idea would look like?" This keeps the relationship intact, gives the requestor a clear path to getting what they want, and puts the burden of evidence where it belongs — on validating the idea before committing sprint capacity. I also maintain a documented Won't list — explicit deprioritization decisions with rationale — so stakeholders do not re-request the same features every cycle.
Common Trap
Saying "we can't do that" rather than "not now, and here is what would change that." The first closes the conversation; the second opens a productive negotiation.
A feature has a high RICE score but does not trace to any current OKR. How do you handle it?
A high-RICE feature that does not trace to a KR is a signal, not a disqualifier. It means either the feature is genuinely strategic work the OKRs do not capture, or the OKRs are incomplete. My process: First, I check if the feature traces to a future-quarter KR — if so, it goes to the strategic backlog with a planned activation date. Second, I check if the OKRs are missing an outcome this feature would drive — if the business cares about this outcome, it should be a KR. Third, if neither applies, I ask the sponsoring stakeholder: "What success metric does this move?" If they cannot answer, the feature goes to discovery, not the sprint. I never completely block features that lack OKR tracing — sometimes the best ideas expose gaps in how we have framed our goals.
Common Trap
Using OKRs as a bureaucratic gate that kills good ideas. OKRs are a lens, not a lock. The goal is to ensure intentionality, not to prevent creativity.