Ask ten teams what return they expect from an AI project and you will hear ten versions of 'faster' and 'better'. Those words hide four different kinds of value that behave differently, are measured differently and are believed differently by a finance team. Naming which one you are pursuing is the first step to measuring it.
Ledger one: revenue
More money coming in: faster responses that win enquiries, better follow-up that recovers lost deals, more relevant offers that increase conversion. Measure it against a control — the same period, segment or channel without the change. Revenue claims without a comparison are the easiest to inflate and the first a finance leader will question.
Ledger two: quality
Fewer errors, fewer reworks, fewer complaints, fewer compliance breaches. Quality is mostly avoided cost, which makes it invisible unless you count it. Track error rates and rework before and after, and put a cost on each incident type so the avoided cost shows up in the same units as everything else.
Ledger three: efficiency
The same output for less effort or spend: lower cost per processed document, shorter cycle time, fewer manual touches. This is the most measurable ledger and the one most AI projects start with. The trap is counting hours saved as money saved; hours only become money when work is redeployed, volume grows, or spending actually falls.
Ledger four: productivity
More valuable output from the same people: a team handles more cases, an analyst covers more accounts, a specialist spends time on judgment instead of data gathering. It is the ledger that matters most for growth and the hardest to prove, because output has to be defined and measured in a way that doesn't reward volume over value.
Choose one primary ledger per project and measure it properly. Report the others as secondary effects. Projects that claim all four usually evidence none.
How to avoid double counting
- Pick the primary metric before you build — and record its baseline on real data.
- Convert to money last — measure in the natural unit (hours, errors, days), then price it with agreed assumptions that finance has signed off.
- Net out the run cost — model usage, hosting, monitoring and the review time people still spend. A gross saving is not a return.
- Set a review date — compare the result with the baseline at a fixed point, in writing.
Why this makes better decisions
When the ledger is clear, the project shape follows. A revenue project needs a control group and attention to the customer journey. A quality project needs error taxonomies and a review loop. An efficiency project needs process mapping and honest volume. A productivity project needs a definition of valuable output. Different goals, different builds, different proof.