Ask how an enterprise AI program is performing and the answer is often a usage chart: prompts served, active users, hours saved. These numbers are easy to collect and almost impossible to take to a finance committee.
Why activity metrics mislead
Usage shows that people tried a tool. It does not show that anything about the business changed. An hour saved is only a return if it is redeployed to something of value or removed from cost, and most programs track neither.
Activity metrics also flatter. A widely used assistant with no effect on outcomes will look healthier than a narrow agent that quietly removed a backlog.
Three places value shows up
Returns from AI land in the same places as returns from any other investment: lower cost to perform a process, more revenue from faster or better decisions, and reduced risk or loss. Every use case should name which of these it is meant to move, and by roughly how much, before it is built.
If a use case cannot name the line it moves, it is an experiment. That is fine, as long as it is funded as one.
Start from a baseline
The most common reason AI value cannot be demonstrated is that nobody measured the process beforehand. Cycle time, error rate, cost per transaction, and volume need to be captured before the first release. Without a baseline, every later number is an estimate that finance is entitled to discount.
Count the full cost
Model usage is the visible cost and rarely the largest. Integration, evaluation, monitoring, change management, and the team that runs the system all belong in the calculation. A use case with a modest benefit and a low run cost can outperform an impressive one that needs constant attention.
A practical framework
- 01Name the financial line each use case is expected to move, with an owner in the business.
- 02Record a baseline for the process before the first release.
- 03Track the outcome metric and the full cost to run, release by release.
- 04Agree with finance in advance how a benefit will be recognized.
- 05Review the portfolio quarterly and stop what is not earning its place.
Measured this way, some celebrated programs look smaller and some unremarkable ones look essential. Both are worth knowing.