Blueprint: identifying and prioritizing AI use cases (inventory, decision matrix and pilots that get measured)

An AI use case is not a technology ("use a chatbot") but a business task with an owner, a volume and a measurable current cost ("answering the 400 monthly customer queries about order status, which today consume 60 hours of two people"). Most AI idea lists fail by confusing those two things: they are born from demos somebody saw, not tasks somebody suffers, which is why none reaches production. The classic symptom has a name in the job market: in our corpus of 79 real AI leadership postings across Switzerland and Europe, identifying and prioritizing use cases appears as an explicit duty in nearly a third, precisely because companies already tried the wish list and it did not work.

The quick test for whether this is you: ask for your company's current list of AI ideas and check how many have three numbers next to them (hours or euros the task costs today, monthly volume, and who owns it). If the answer is none, you do not have a technology or talent problem: you have a method problem, and it gets fixed in two weeks with what follows.

The whole process fits in seven steps and two calendar weeks for a mid-sized company. The reading rule: nothing enters the matrix without its three numbers, nothing enters a pilot without a baseline, and nothing scales without its decision date honored. It is the same circuit the Chief AI Officer blueprint prescribes for the first 90 days; this is the fine mechanics of the use-case part.

The inventory is a week of structured conversations, not an ideation workshop. The goal is not for people to propose AI uses (they will propose demos they saw): it is to extract the tasks that fit the profile AI solves well today, which are almost always the ones nobody brags about doing.

Three criteria, each scored 1 to 5, and the total score is value plus feasibility minus risk. Three criteria and not one more: every extra column added to a prioritization matrix is a new place to hide a personal preference. The full rubric, so that two different people score alike:

Three typical tasks of a mid-sized services company, scored with the rubric above. They are worked examples to calibrate your eye (your company's numbers will differ), and they show the pattern that surprises almost every committee: the glamorous case loses to the boring case with volume.

Every approved pilot gets written on one page with the table's fields, and that page is the contract. The field almost everyone omits and that saves the most pilots is the last one: the date on which the pilot stops existing as a pilot, whether scaled or killed. Without a date, every pilot tends toward eternity, because killing it looks like failure; with a date, killing it is following the plan.

One: no task enters the matrix without its three numbers; what cannot be measured cannot be prioritized, only wished for. Two: two pilots maximum at once; the third steals the supervision the first two need to be conclusive. Three: no pilot without a prior baseline; a baseline estimated from memory after launch always comes out inflated. Four: every decision date is honored, and scale or kill are the only two valid outcomes; "keep piloting" is the forbidden answer. Five: every quarter, the whole matrix gets rescored, because feasibility changes every few months with each model generation, and January's unfeasible case can be June's obvious one.

The fifth rule is what turns this from a one-off exercise into an operating system: the living matrix is the company's AI roadmap, and whoever maintains it (internal or fractional) holds the map that job postings call strategy.

This blueprint is not consultant theory: it is the filter the three production systems of this lab passed through. The editorial operation won its pilot because it was the highest-volume task with owned data and review-controllable risk; the market research system scored high on feasibility (public sources, digital text) and its risk was mitigated with the data contract; and more than one flashy candidate was killed on time, its decision date honored, because the numbers did not work. This house's matrix gets rescored every quarter, and from it will come the next cases published here, with their numbers and their failures, as always.

If you want the matrix applied to your company under our honesty rules (real baseline, dated decision, governance from day one), the usual door is below, on the card of the person who signs.

Two shifts will change your matrix within a year. The first is the feasibility column: agents able to execute complete tasks are raising the score of cases that were unfeasible a year ago (the ones requiring navigating systems, crossing sources or sustaining long processes), so quarterly rescoring stops being hygiene and becomes advantage: whoever rescores first, captures first. The second is the risk column: the EU AI Act's obligations apply in phases and will turn filters that are best practice today into requirements with sanctions, making customer-facing cases without governance more expensive and, in relative terms, the boring internal ones cheaper. Our read: the competitive edge of the next two years will not be having AI ideas, which are already free, but the speed and honesty of the idea-pilot-decision circuit. That circuit is exactly what this blueprint leaves built.