Plenty of AI budget is spent on faith and produces nothing measurable. A lightweight ROI model fixes that, and doubles as a sharp qualifier for which use cases to fund first. Here's how to build one before you hire anyone.
The simple model
- 1Pick one use case and one primary metric (e.g. support tickets deflected).
- 2Estimate impact conservatively: units affected × value per unit × realistic adoption.
- 3Estimate cost: build (team or embed) + run (compute, tooling, maintenance).
- 4Payback = cost ÷ monthly impact. Under ~12 months is a strong signal.
- 5Sanity-check with a small pilot before scaling.
A worked example
AI support deflection: 10,000 tickets/month, 40% deflected, €4 handling cost each = ~€16k/month impact. If an embedded specialist pair costs ~€30k/month for three months to build (~€90k) plus modest run costs, payback lands well inside a year, then it's margin.
