Calculating AI ROI — what a twelve-month payback actually requires

The basic model

Annual return = (hours saved per person per month × realisation rate × loaded hourly cost × 12 × people) − (licence cost + rollout cost)

The term that does the work is the realisation rate: the proportion of saved time that converts into other output rather than evaporating. Twenty minutes saved across a day often becomes twenty minutes of something else, not twenty minutes of additional billable work. Assume 30 to 50 per cent unless you can show otherwise.

Three assumptions that break the case

1. Everyone uses it

Actual sustained usage after three months is frequently under half of licensed seats. Model on measured adoption, not on seats purchased, and buy seats in waves rather than all at once.

2. Saved time is fungible

It is fungible when the saved task is a bottleneck for something queued behind it. It is not when the saved time is scattered across a day. The first case supports a real ROI claim; the second supports a quality-of-work claim, which is legitimate but should not be dressed as a financial one.

3. Rollout is free

Training, policy writing, the security review and the person who owns the numbers all cost time. Put a figure on them. A deployment that needs forty hours of internal effort has a real cost that belongs in the denominator.

A worked example

Note that the case survives even though only 60 per cent of seats are used and only 40 per cent of saved time is realised. A business case that needs optimistic values for both is not a business case.