How to Score and Prioritise AI Use Cases

Every AI programme eventually produces a backlog of forty use cases, a scoring spreadsheet with eleven weighted criteria, and no shipped work. The spreadsheet is not the problem, but it is a symptom. Scoring models fail in a predictable way: they optimise for defensibility rather than decision-making. Eleven criteria with weights to two decimal places produce a ranking nobody argues with and nobody believes. The scores cluster, the top six are within noise of each other, and the actual selection gets made in a meeting on grounds the model never captured. A useful scoring model has few criteria, coarse scales, and an explicit tie-break. Ours has four: value at stake, feasibility given your readiness scores, time to first evidence, and reversibility. Scored 1-3, not 1-10 — a ten-point scale invents precision you do not have. The fourth criterion is the one most models omit and the one that matters most early: how cheaply can you undo this if it goes wrong.