How to Measure AI Pilot ROI Without Inventing Savings
Measure the whole task, include the cost of checking and separate capacity released from cash saved. A faster draft alone is not a business case.
Record a baseline before the pilot
Choose a defined task and observe how it is done today. Record volume, end-to-end handling time, corrections and exceptions over a period that reflects the work. Note unusual events and distinguish observations from estimates. If work varies by complexity, group comparable cases rather than hiding the difference in one average.
Agree the minimum acceptable quality before measuring speed. A process that becomes faster by omitting an important check has changed the task, not necessarily improved it.
Include review and rework
Measure the pilot from the arrival of the input to an acceptable final result. Include preparing information, writing instructions, checking outputs, correcting errors and handling exceptions. Keep initial configuration and training effort separate from recurring operating time so the ongoing comparison is clear.
Where practical, use comparable work and record who performed it. A test completed by the most enthusiastic user may not represent the whole team. Include people who will run the workflow after handover and ask where it creates friction.
Use a transparent calculation
Illustrative example only — not a WaypointAI client result. Assume 100 tasks a month. The measured manual process takes 12 minutes per task; the pilot takes 8 minutes including checking. The difference is 400 minutes, or about 6.7 hours a month.
At an assumed fully loaded staff cost of £25 an hour, that represents about £167 of monthly capacity value before software, support and implementation costs. The volume, times and rate are hypothetical. Replace every assumption with evidence from your business.
Released capacity is not automatically a cash saving. It becomes a cash saving only when an actual cost is avoided or reduced. Otherwise, explain how the available time will be used: reducing backlog, improving service or taking on additional work. Do not count the same benefit twice.
Include the full cost
- Initial assessment, configuration, testing and training.
- Incremental workspace, software or API usage charges.
- Ongoing review, maintenance and support.
- Time spent on failures and the manual fallback.
Use costs in GBP and apply VAT consistently with the business's accounting treatment. Compare incremental costs with incremental benefits; do not charge an entire existing software contract to a pilot if only a small additional cost is attributable to it.
Track quality and adoption alongside time
Useful measures include the proportion accepted after review, material error rate, exception rate, completion rate and whether intended users continue using the workflow. A low-volume pilot can give useful operational evidence, but it cannot justify sweeping claims of reliability or return.
Keep a simple evidence log: date, task type, manual or pilot, handling time, correction required, exception and outcome. Avoid collecting unnecessary personal information merely to make the spreadsheet look thorough.
Agree the decision the evidence will support
Before the pilot, decide what would justify expansion, another iteration or stopping. Review unresolved risks as well as averages. If the benefit disappears when realistic checking and support are included, revise the approach rather than polishing the forecast.
WaypointAI's workflow pilots use agreed measures and representative testing. For help selecting the right task, read our pilot selection checklist or discuss an AI readiness assessment.