Decision Frameworks with AI

Decision frameworks like RICE, MoSCoW, and weighted scoring matrices are designed to impose systematic evaluation on complex choices. In practice, the humans applying these frameworks introduce the same biases the frameworks are meant to eliminate. A product manager scoring 'impact' for their pet feature will unconsciously inflate the estimate. A team evaluating 'effort' will anchor on whoever speaks first. The framework provides structure, but the inputs are still subjective and inconsistent across evaluators and across time.

AI applies the same criteria to every option with the same rigour. It doesn't have a favourite feature, doesn't anchor on the first estimate it hears, and doesn't weight criteria differently based on who proposed the option or how eloquently they argued for it. This mechanical consistency is the specific value AI adds to decision-making — not better judgment than humans, but more consistent application of the judgment framework the team has already defined.

The inconsistency compounds when teams make decisions over time. The same team applying the same RICE framework in January and March will produce different scores for identical inputs because their reference points, priorities, and biases have shifted. AI provides temporal consistency — the same inputs produce the same outputs regardless of when the evaluation is run, making it easier to compare decisions across different time periods.

Start by defining the evaluation framework explicitly: what criteria matter, how they're weighted, and what scale each is scored on. Then provide the AI with the same information your team would use — option descriptions, available data, constraints, strategic context, and any relevant historical decisions. Ask the model to score each option against each criterion with explicit reasoning for each score, not just a number.

The reasoning is as valuable as the scores. When the AI explains why it scored one option's 'confidence' as low ('this estimate depends on a partnership agreement that hasn't been finalised, introducing execution risk'), it surfaces assumptions that might otherwise go unexamined in a meeting. The team can then challenge the reasoning, provide additional context, adjust the inputs, and re-run the analysis.

Present the AI's analysis as input to the decision, not as the decision itself. The team reviews the scores, debates the reasoning, overrides where they have information the AI doesn't, and makes the final call. The AI's contribution is ensuring every option was evaluated against every criterion with the same rigour — eliminating the common failure mode where option A gets careful analysis, option B gets a cursory look, and option D is forgotten entirely.

Human scenario planning typically generates two or three scenarios: optimistic, pessimistic, and 'most likely.' These tend to be anchored close to the expected outcome because they're constrained by the imaginer's experience and assumptions about what's possible. AI can generate a wider range of plausible scenarios, including compound events that humans tend to evaluate independently rather than in combination.

For each scenario, the AI can trace second and third-order consequences — a form of downstream impact analysis — that humans struggle to compute in real time: 'If the API partner raises prices, your cost per transaction increases, which reduces margin on the free tier, which means the conversion to paid needs to be higher than your current forecast to maintain unit economics.' This chain of consequences is the kind of analysis that gets lost in meetings but is tractable for a model with all the relevant inputs in its context window.

The most valuable scenarios are often the ones that combine multiple modest changes rather than focusing on single dramatic events. A scenario where competitors lower prices by 10%, hiring takes 30% longer than planned, and a key integration launches two months late may be more likely and more impactful than the dramatic scenarios teams usually plan for. AI is well-suited to exploring these combinations systematically.

AI cannot supply the inputs that require lived experience: how the team will react to a particular direction, whether a specific customer relationship can survive a pricing change, whether the engineering team has the morale and trust to tackle a risky rewrite, or whether a market opportunity is real or hype. These are human judgments that depend on social intelligence, emotional awareness, and contextual knowledge the model doesn't have.

AI also reflects the patterns in its training data when making value judgments. If asked to evaluate whether a marketing strategy is 'bold enough,' the model draws on patterns from its training corpus, which may not reflect your specific market, culture, or risk tolerance. The model is most valuable when applying a framework you've defined with criteria you've chosen, not when making judgment calls about what framework to use or what criteria matter.

Over-reliance on AI-assisted frameworks can create a false sense of rigour, making confidence calibration essential. A neatly scored RICE matrix looks objective, but if the inputs are guesses ('I think reach is about 10K users'), the output inherits that uncertainty without flagging it. AI makes the framework application consistent, but it cannot make uncertain inputs certain. The discipline of acknowledging input uncertainty — and adjusting confidence in the output accordingly — remains a human responsibility.

Rather than prompting from scratch each time a prioritisation decision arises, teams benefit from maintaining a library of decision templates — similar to a prompt library — — pre-built prompts that encode the organisation's preferred frameworks, criteria definitions, and scoring scales. A RICE template includes the team's specific definitions of Reach, Impact, Confidence, and Effort, calibrated to the organisation's context rather than using generic definitions that may not map to their work.

Templates should include few-shot examples of previously scored options with the reasoning documented. These examples calibrate the AI's scoring to the team's standards — showing that a 'high impact' score in your context means something specific, grounded in past decisions the team agreed on. Without calibration examples, the AI applies its own interpretation of impact, which may not match your organisation's scale.

Version the templates alongside the team's strategic priorities. When OKRs shift, the criteria weightings in the decision template should shift accordingly. A template that weighted 'revenue impact' heavily last quarter might need to weight 'infrastructure reliability' more heavily this quarter if the team's priorities have changed. Keeping templates aligned with current strategy ensures that AI-assisted decisions reflect the team's actual priorities, not last quarter's.