Hours saved is not cash: measuring AI value honestly
Twenty minutes saved per person per day is not a number a CFO can bank.
The arithmetic that does not survive scrutiny
The standard AI business case multiplies a per-task time saving by a headcount and a loaded hourly rate. The result is a large, confident number that finance will not accept, and they are right not to.
Time saved is only a cost saving if the cost actually leaves — a role not backfilled, contractor spend that stops, overtime that is no longer needed. Otherwise it is capacity, which is real but different.
Three kinds of saving, named honestly
Cost removed: spend that genuinely stops. The strongest claim, and the rarest.
Capacity released: the same people doing more valuable work. Real, but it needs a named answer to what they now do instead — otherwise it is unbanked.
Risk or quality improved: fewer errors, faster cycle time, better consistency. Often the most valuable of the three and the least likely to be measured, because nobody baselined the error rate.
Baseline first, or do not claim
Every honest number depends on a measurement taken before anything was built. Cycle time, error rate, volume, effort — captured while decomposing the work, not reconstructed afterwards from memory.
The discipline costs an hour at the start and is the difference between a defensible claim and an anecdote.