Boards are asked to approve AI spending on the strength of category enthusiasm rather than commercial reasoning. The question that matters is narrower and more useful: which recurring activity in this business, at its current volume, costs enough that changing it would show up in the numbers?
Frequency beats sophistication
Return is a function of volume multiplied by unit time saved, minus the cost of building and maintaining the change. A task performed twice a month is almost never worth automating, however clever the solution. A task performed three hundred times a day usually is, even when the saving per instance is small. This is why the highest-return work in a mid-sized business tends to be unglamorous: triage, classification, data entry, routing, acknowledgement, retrieval and reporting.
The four conditions worth looking for
- High frequency: the activity recurs daily, not occasionally.
- Rule-describable: a competent colleague could explain the decision in a paragraph.
- Structured or semi-structured input: an email, a form, an invoice, a document.
- A measurable consequence of delay: revenue at risk, cost incurred, or a service level missed.
Where all four hold, the business case is usually straightforward. Where none hold, the honest answer is that AI is not the intervention — better management, clearer ownership or removing the process altogether may be.
Do the arithmetic before the demo
Take a worked illustration. Assume twelve people each lose six hours a week to repetitive handling, at a loaded cost of £32 an hour across forty-six working weeks. That is roughly 3,300 hours and approximately £106,000 of capacity annually. Assume conservatively that just over half of that time is genuinely addressable. The addressable capacity is then around £58,000 — before considering any revenue effect from faster response.
The numbers above are stated assumptions used to demonstrate the method. They are not benchmarks, survey findings or client results.
The point of the arithmetic is not precision; it is proportion. It tells you whether a project belongs in this quarter or next year, and it gives you a baseline to be judged against afterwards.
Two distinctions that prevent disappointment
First, released capacity is not the same as cash saved. Recovering 1,700 hours does not reduce payroll unless you intend it to; it creates capacity that must be deliberately redirected to something of value. Say which, in advance.
Second, revenue effects should be claimed cautiously and instrumented properly. Faster response times plausibly improve contact and conversion rates, but only measurement against a baseline turns that from a hypothesis into a result.
A practical sequence
- List the ten activities your teams repeat most often.
- Estimate volume and handling time for each; imprecise numbers are still decision-grade.
- Score each for rule-describability and input structure.
- Rank by annual value divided by implementation effort.
- Pilot the top item against a measured baseline before committing further.
Nothing in that sequence requires a technology decision. That is deliberate — and it is the main reason it works.