The AI projects that quietly get killed a year later almost always skipped one step: deciding, before building, what number would prove it was worth it.
Before you build anything
- Pick one existing metric — support resolution time, hours spent on manual entry, search-to-purchase conversion — not a new metric invented for the AI feature.
- Record the baseline before the feature ships, not after.
- Set a real target — “faster” isn't a target; “under 4 hours average resolution, down from 18” is.
After it ships
Measure against that same baseline on a fixed schedule — 30 and 90 days is typical. If it's not moving the number, that's useful information too: either the scope was wrong, or the task wasn't actually a good fit for automation.
We scope every AI engagement around one measurable outcome before writing any integration code — see AI Agent Development and AI Integration Services for how that works in practice.