--- title: "How to prioritise AI use cases: value, feasibility and risk" url: https://zerofive.ai/en/blog/strategy/prioritising-ai-use-cases canonical: https://zerofive.ai/en/blog/strategy/prioritising-ai-use-cases language: en published: 2026-04-02 updated: 2026-09-18 author: "ZeroFive.AI" tags: AI use cases, AI project prioritisation, AI portfolio, AI backlog, artificial intelligence ROI abstract: "A method for prioritising AI use cases: the three evaluation axes, the most common mistakes and how to build a backlog a board can approve." --- # How to prioritise AI use cases: value, feasibility and risk There is a recurring moment in adoption journeys, and it arrives right after the ideation workshops: the company finds itself with a list of thirty, fifty, sometimes a hundred AI use case ideas, all plausible, all sponsored by someone, and a budget that funds five. What happens next decides more of the whole programme's success than any technology choice, and yet it is the step treated with the least method: in most cases the winner is whoever shouts loudest in committee, or the idea most similar to the demo seen at the last event. The data suggests what this selection by acclamation costs. MIT Project NANDA (The GenAI Divide, July 2025) observed that over half of GenAI budgets concentrate on sales and marketing, where results are the most visible to narrate, while the best returns show up in the automation of back office and operational functions, where nobody issues press releases. Capital, left to instinct, goes where things glitter. ## The three axes, and how to measure them without cheating Serious prioritisation evaluates every use case on three axes, and the discipline lies in the how, even before the criteria. Value must be expressed in a business metric someone signs: hours recovered on a process at their full cost, margin points, reduction of a quantified risk, incremental revenue with its assumptions made explicit. The signature is the important part, because a value estimated by the AI team is worth half of one estimated by the process owner who will later answer for it. Feasibility measures the distance between the use case and the conditions for delivering it, and the component that almost always dominates is data, with the five readiness questions we covered elsewhere: access, measured quality, ownership, time, legitimacy. Alongside data weigh the integration required into existing systems and the skills available, in-house or purchasable. Risk, the third axis, crosses the AI Act classification (a use case in Annex III territory carries a set of obligations that changes its timeline and cost), the exposure on personal data and the impact of a system error on the customer or the person. High risk does not disqualify, it informs: it moves the use case later in the sequence, to when governance will be ready to carry it. ## The mistakes that warp the ranking Four distortions recur in almost every portfolio we review. The first is scoring without declared weights, where every function fills the matrix inflating its favourite axis: weights must be decided beforehand, together, and then applied to everyone. The second is the single horizon, which races three-month quick wins against two-year bets on the same scale: a healthy portfolio separates them by construction, typically with a majority share of fast-return initiatives that fund and legitimise the few structural bets. The third distortion is ignoring dependencies: three use cases sharing the same data foundation are worth, together, more than their sum, and sequencing them well collapses the marginal cost of the second and the third. The fourth is the ranking carved in stone, while scores age quickly, because data improves, model costs fall, regulation advances: the backlog is a living artefact, to be revisited at a fixed cadence, with the same seriousness as the first draft. ## From score to resolution The ranking, alone, remains an exercise until it is translated into a format the top can decide upon. The translation that works has three components: the first use cases with their signed value and the prerequisites to close, the go/no-go criteria each will have to pass to move phase, and the list of what was deliberately excluded with the reasons why, which is the most underrated part and the one protecting the programme from endless reopenings. A backlog without explicit exclusions is a wish list with numbers next to it. It is the work our AI Assessment covers, the second link of the chain after the maturity measurement: use case map, scoring on the three axes with shared weights, quick wins identified and a backlog the board can approve knowing what it is choosing and what it is choosing to forgo. If your list of thirty is already on the table and the budget for five too, the method to get from one to the other takes a few weeks to build: calendly.com/fabiolalli/zerofive, or hello@zerofive.ai. And if you have already made your ranking, the stress test is a single one: who signed the numbers in the value column?