Approving the case is the start. Most AI value is lost after go-live: adoption stalls, costs creep, and nobody compares results with the case.
Set up before launch
- Baseline: measure volume, time, cost, error rate and cycle time before the change.
- Holdout: where possible, keep a team, region or period without the tool to compare.
- Owner: one person accountable for realized value, not just delivery.
Track monthly
| Measure | Why |
|---|---|
| Adoption (active users, share of volume through AI) | Benefits follow use |
| Unit outcome (time or cost per item, error rate) | The real effect |
| Run cost (seats, tokens, people) | Costs that scale with use |
| Realized benefit vs case | The honest scorecard |
| Incidents and overrides | Quality and control |
Review quarterly
Compare realized value with the approved case and decide: scale, fix or stop. Re-forecast benefits and costs with what you now know. Retire tools whose value never arrived.
Why this is the CFO's job
Esker's 2026 Global Finance AI Trust Index found that 65% of CFOs report difficulty connecting AI usage to specific business outcomes, compared with 40% of other finance leaders. The function that owns the numbers has to own the method.
Signs a use case needs attention
- Adoption flat after the ramp period.
- Unit cost rising faster than volume.
- Override rates high or rising: people don't trust the output.
- Benefits reported only as anecdotes.