REAL ROI · 5 MIN READ
Picking the one number your AI programme should be judged on
If you cannot name the metric before you build, you will not be able to prove the result after. Here is how to choose it.
LAUNCH AGENTS ·
Why most AI ROI claims fail
Ask a team how their AI pilot is doing and you will usually hear about usage: how many people tried it, how many queries it handled, how many hours it might have saved. These are signs of activity. They are not results. A CFO cannot find hours saved on the P&L.
The fix is simple and rarely done: agree one business number before anything is built, and judge the work on that number alone.
What makes a good number
- Leadership already watches it. If it appears in a board pack or a weekly operating review, it matters. Inventing a new KPI for the AI project makes the result easy to ignore.
- The constraint controls it. The number must move when the bottleneck moves. Quote turnaround for an engineering constraint, days to close for a reconciliation constraint, resolution time for a specialist support queue.
- It lives in your systems. The baseline and the result should come from the ERP, CRM or helpdesk, not from a survey or a vendor dashboard.
- It moves within a quarter. Annual metrics hide the effect. Choose something that responds within weeks.
Examples by function
Finance: days to close, DSO, days from invoice to payment. Customer service: resolution time, cost per ticket. Sales: proposal turnaround, win rate. Procurement: savings against a spend baseline. Operations: fill rate, inventory days, on-time in-full. People: time to shortlist, time to productivity.
Baseline, then cadence
Take the baseline from at least three recent periods so that one unusual month does not distort it. Write down how the number is calculated and where the data comes from. After go-live, review it on a fixed cadence with the sponsor, weekly or monthly depending on how fast it moves.
Supporting measures still matter. Adoption rates, exception volumes and agent accuracy explain why the number is or is not moving. But they serve the main number; they do not replace it.
When the number holds
Once the improvement is stable, the constraint has moved. That is the signal to start the next cycle, with a new bottleneck and a new number. Over time the organisation builds a track record of results it can see in its own systems.
Key takeaways
- Usage and hours saved are activity, not results.
- Pick a metric leadership already watches and the constraint directly controls.
- Baseline it from your own systems across at least three periods.
- Review it on a fixed cadence; supporting metrics explain it but do not replace it.
Frequently asked questions
How do you measure ROI from AI agents?
Agree one business metric before building, baseline it from your own systems, and review it on a fixed cadence after go-live. Examples include days to close, resolution time and proposal turnaround.
Is hours saved a good AI ROI metric?
Hours saved is a useful supporting signal but rarely shows up in financial results. A business metric that leadership already tracks is a stronger measure.