How to measure AI ROI before expanding a pilot

Ask a reviewer who did not produce the output to check a sample from both processes using the same rubric. Hide which process produced it where practical. Keep the rubric short enough to apply consistently: correctness, required information, unsupported additions and readiness for the next person.

For a cash-only business case, replace the planning value with an expense that will actually disappear, such as approved overtime or an avoided temporary contract. If no expense changes, describe the result as capacity. The organisation can still choose to buy capacity, but finance should know what decision it is making.

The simple first-year return on the initial setup investment in this example is (12 × HK$1,400 − HK$14,000) ÷ HK$14,000 = 20%. Its denominator is the initial setup investment, so label it explicitly when comparing other ROI definitions. It excludes discounting and tax. It is a screening calculation, not a valuation model, and it is only useful if the assumptions remain visible.

Repeat the calculation at lower adoption and longer review time. At 50% adoption, the same example produces about 13.3 hours, worth HK$2,400 under the assumed rate. Net monthly value falls to HK$200 before setup recovery. If review grows enough to remove the two-minute saving, the time-based business case disappears.

List costs that change with volume separately from fixed costs. Include licences, model usage, document storage, integration support, access reviews, training, source updates and someone monitoring failed tasks. Ask the supplier which line items increase if staff repeatedly regenerate an answer or process an unusually long document.

An illustrative gate is positive net value at 50% adoption, no increase in material correction errors and a payback period within the company’s own purchasing limit. Use the limit finance already applies to similar projects rather than inventing a special rule for AI. Keep high-consequence failures as separate stop conditions even if their average cost appears small.

Stop or redesign if the result depends on excluding review time, if the team cannot reproduce the baseline, or if the benefit is counted twice. Schedule a review after a full operating cycle. A successful pilot earns the next measured step; it does not establish permanent savings across every department.