Time saved is useful, but it is not automatically money saved. A credible AI business case explains what changed, what it cost and how the released capacity will actually be used.
Measure the current process first
Select a repeated task and record a representative sample. Include preparation, execution, checking and exception handling. Separate waiting time from active effort. A task that sits in an inbox for a day may require only five minutes of work, so automation affects those two measures differently.
Record quality at the same time. If the existing process regularly misses information, simply comparing speed can reward a system that repeats the problem faster. Define the minimum acceptable output and the consequence of a wrong one.
An illustrative calculation
This is a hypothetical example, not a client result or savings forecast. Suppose a team completes 200 internal briefs per month. Each currently takes 20 minutes, including checking. A pilot reduces total effort to 12 minutes after allowing for review and exceptions. That releases about 26.7 hours per month.
At an assumed loaded labour cost of £40 per hour, that is approximately £1,067 of capacity value. If model, hosting and support costs total £450 per month, the remaining capacity value is about £617. A £6,000 implementation would take roughly 9.7 months to recover on those assumptions, before tax, financing or other effects.
That is not a promise of £617 in extra cash. If salaries stay the same and the time cannot be redeployed, the benefit may instead be faster response or more capacity at the same headcount. State which benefit the business expects to realise.
Count the whole operating cost
- Implementation and integration effort.
- Model usage, retrieval and external tools.
- Hosting, monitoring and support.
- Human review, corrections and escalation.
- Training and maintenance when the process changes.
Use successful task completion as the denominator. A low-priced model that needs repeated attempts can cost more in practice. Likewise, a system with a high headline automation rate may still create a costly queue of difficult exceptions.
Test the assumptions that can break the case
Recalculate with fewer tasks, more review effort and higher running cost. If a modest change removes the benefit, the pilot needs better evidence before expansion. Include a stop condition as well as a target: for example, a maximum correction rate or an upper limit on reviewer effort.
Do not count the same benefit twice. Time freed in sales administration and extra sales revenue may be connected, but the revenue still needs evidence. Track it separately, account for other changes and avoid attributing every improvement to the new system.
Make the next investment conditional
A pilot report should explain scope, dates, task volume, quality, total effort and actual cost. Keep difficult cases visible. A small sample can reveal design problems without establishing a reliable annual forecast.
NIST’s measurement guidance supports evaluating systems in context. The financial example above is our own arithmetic, not a claim drawn from an industry benchmark. Use your numbers and have the budget owner review the assumptions.
Read our testing guide to build quality criteria alongside the commercial case.