Most AI business cases quote a licence fee and hours saved. Neither is a unit cost. Our view: measure the cost of one accepted output, including failed attempts, review and rework. The Completed-Task Ledger explains how.
Most AI business cases contain two numbers: the licence fee and the hours the team says it saved. Neither is a unit cost. A licence fee tells you what you pay for access. A claim of hours saved tells you what people believe happened. Neither tells you what it costs your organisation to produce one finished, accepted piece of work with AI in the loop.
Finance leaders already know how to think about this in every other part of the operation. Nobody judges a print shop by the price of the printer. They ask what one usable page costs once the misprints, the reloads and the operator's time are counted. AI deserves the same discipline, and in our view the organisations that adopt it will be the ones that can state their number.
The honest unit of AI economics is not the prompt or the seat. It is the completed task that someone accountable was prepared to sign.
We use a simple ledger with four lines. It is deliberately boring, because the value is in having every line filled in, not in the sophistication of the arithmetic.
Cost per completed task = (direct cost of every attempt + review time + rework time + escalation time) divided by the number of outputs that met the quality bar.
Notice what sits in the denominator. It is not the number of outputs the tool produced. It is the number that were accepted. A tool that produces ten drafts and three usable ones has a different economic profile from a tool that produces four drafts and three usable ones, even if the invoice is identical.
A tool that saves ten minutes of drafting and adds twelve minutes of checking has not saved anything. It has moved the work to a place where nobody is measuring it.
Consider a programme team that uses AI to prepare first drafts of partner reports. The figures below are illustrative, chosen to show the mechanics, not to describe any real organisation.
Suppose the team attempts twenty reports in a month. Fourteen meet the quality bar after review, and six are abandoned and written by hand. Direct usage cost is modest. Review takes forty minutes per attempted report, because the reviewer reads every draft, including those that will be discarded. Rework on the fourteen accepted drafts averages twenty minutes.
The point of the ledger is what happens when you divide by fourteen rather than twenty. The six abandoned drafts still cost review time, and that cost is spread across the reports that survived. The number that emerges is higher than the team expected, and it tells them something specific: the failure is in report types where source material is thin. A narrower scope, not a better tool, would lower the unit cost.
First, it changes which tasks you choose. Tasks with low review risk, stable formats and clear quality criteria produce low unit costs. Tasks where every output needs expert judgement rarely do. This is consistent with the principle we return to often: the unit of AI value is a task, never a job.
Second, it changes the vendor conversation. If a supplier can only describe cost per seat, ask them for cost per accepted output in a comparable task. If they cannot answer, that is information about how well they understand your use.
Third, it changes what you report upward. A board does not need a count of prompts or a satisfaction score. It needs three figures per use case: cost per completed task, acceptance rate, and a comparison with the unit cost of the previous method. Those three figures survive scrutiny from a funder or an auditor because they are measured rather than asserted.
You do not need new software. For one use case, over four weeks, record five things on a shared sheet: attempts made, outputs accepted, minutes spent reviewing, minutes spent reworking, and outputs escalated. Convert minutes to cost using a loaded hourly rate agreed with finance. Then compute the ledger once, and repeat monthly.
Be careful about two temptations. The first is to count only the successes and quietly drop the abandoned attempts. The second is to price staff time at zero because it is salaried. Both make the number look better and both make it useless for decisions.
"Our staff are salaried, so their time is free." It is not free, it is committed. Every minute spent reviewing an AI draft is a minute not spent on something else, and the opportunity cost is real even when no invoice arrives. Use a loaded hourly rate agreed with finance, and use the same rate before and after.
"The tool gets better every month, so today's number is meaningless." The tool changes, but your tasks, your documents and your quality bar do not change at the same pace. A number recorded monthly shows whether improvements in the tool are reaching your work. Without the baseline, you cannot tell.
"Measuring it will slow us down." Four weeks of five simple columns is a small cost against a multi-year licence. It is also the fastest way to learn which use cases deserve more investment and which should quietly stop.
A single figure is a snapshot. The useful signal is the trend and the spread. If acceptance rate is climbing and review time is falling, the workflow is maturing. If acceptance rate is flat while volume rises, the organisation is generating more output without generating more value, and the review burden is likely growing unseen.
Watch for the divergence between the cost per completed task and the cost per attempted task. A widening gap means more work is being discarded. That is often the first visible symptom of a task that was a poor match for AI in the first place, or of instructions that have drifted out of date.
Finally, compare across use cases. Two or three figures side by side usually reveal that the organisation's real gains cluster in a small number of well-defined, low-risk tasks. That is a finding to act on, by investing in those tasks, not by spreading effort thinly across every possibility.
If your business case cannot state the cost of one accepted output, including the attempts that failed and the checking that followed, it is describing enthusiasm rather than economics. The organisations that will invest well in AI are the ones that can say what a finished task costs, and can show the number moving.
If this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.