Why we measure every AI build against one number

Every AI build we take on is scored the same way before any work begins: the monthly value of the workflow it replaces, in hours saved or revenue recovered, divided by the cost of building the replacement. The candidate with the best ratio is built first. A candidate that cannot produce a number is set aside, however interesting the technology behind it.

The scoring exists because the choice of first project decides whether an AI programme survives. A first build that pays back inside a quarter buys permission for the second and third. A first build chosen for ambition, and still unfinished six months later, usually ends the programme.

How do you choose the first AI project?

List every workflow where people copy, chase, re-key or assemble the same things each month, then attach two figures to each: what the manual version costs, and what a governed replacement would cost to build. Rank by the ratio. The exercise is short, and the ranking is rarely what instinct predicted, because the workflows people complain about loudest are seldom the ones burning the most hours.

One of our builds started as exactly this exercise. A Midlands manufacturing group had candidate use cases across procurement, finance, sales and operations and no way to pick between them. A two-week audit scored 16 workflows for payback and risk, identified £18k of monthly manual-process cost across the priority workflows, and put 3 builds on the first roadmap. Leadership received a first-build decision rather than a strategy deck.

What happens when a candidate has no number?

It waits. A workflow without a measurable cost gives you nothing to judge the build against after go-live, which means the project can neither succeed nor fail, only continue. Budgets attached to unfalsifiable projects grow quietly. Budgets attached to a payback number get renewed or cut on evidence, which is how it should be.

Why does one number keep a supplier honest?

A supplier scored on payback is rewarded for the shortest route to a working system. Analysis beyond what the decision needs, scope added for its own sake and technology chosen for novelty all show up as a worse ratio, so the incentive runs against padding.

The same arithmetic sometimes returns a small answer: the workflow fits an off-the-shelf tool, and the score says buy it. The score is only useful because it binds us too; a scoring system that gets overridden whenever it is inconvenient is a sales prop, and clients can tell.