AI has unit economics: price the workflow, not the licence
Executive capstone: the money case a CFO will actually sign
The question: What does this cost when it works — and what does it cost when it is used a lot?
Most AI business cases price the licence and forget the meter. Seat costs are fixed and visible; inference, review time, evaluation and rework scale with volume, and they are where budgets quietly break. A leader who can price a workflow per completed task can defend it — and can tell an expensive habit from an expensive capability.
What the lesson covers
A licence is a price. A workflow has an economics. The unit that matters is **cost per completed task** — one resolved ticket, one signed-off commentary, one approved campaign asset — measured end to end, including the human minutes it still consumes. If you cannot state that number, you do not yet have a business case; you have a quote.
Four cost lines, and only the first is obvious. **Licences and seats**: fixed, visible, easy to negotiate. **Inference**: variable, volume-driven, and the one that surprises people — a per-call cost that looks trivial in a pilot of 200 documents and does not look trivial at 40,000 a month. **Human-in-the-loop**: review, approval and escalation minutes, which are a real cost even though they arrive as salary rather than invoice. **Assurance**: evaluation sets, monitoring, incident response, documentation and periodic re-testing — the costs that keep the thing trustworthy, and the first ones cut by teams who mistake them for optional.
Then the offsetting side, and it must be as honest. **Time released** is only a saving if it is redeployed to something valued, or if headcount genuinely changes; otherwise it is a nicer working day, which is worth having but is not a P&L line and should not be presented as one. **Quality effects** — fewer errors, faster cycle time, higher conversion — are usually the larger prize and the harder measurement, which is precisely why they must be instrumented before launch rather than estimated after.
Payback discipline keeps you out of theatre. Compute payback on the **fully loaded** cost, state the period, and then run a sensitivity: what happens at 3× the assumed volume, at half the assumed quality gain, at a 30% price change from the model provider? A case that only works at one exact set of assumptions is not a case, it is a hope. Model prices have moved dramatically in both directions over short periods, so treating today's price as permanent is a modelling error, not a forecast.
The questions a CFO will ask, in the order they ask them: what is the unit, and how did you measure it? What is the volume, and what happens if it triples? What is the payback, and on whose numbers? What is the downside case, and what is the kill criterion? Who owns the budget line when the pilot ends? A team that has answers to those five has usually built the right thing; a team that has not usually built a demo.
Key points
- The unit is **cost per completed task**, measured end to end — not the licence price.
- Inference is variable and volume-driven; a pilot systematically understates it.
- Review minutes and assurance work are real costs even when they never appear on an invoice.
- Released time is a saving only if it is redeployed or headcount actually changes. Say which.
- Any case that survives only one set of assumptions is a hope; sensitivity-test volume, benefit and price.
Framework — AI Unit Economics: cost per completed task
Cost per completed task = (licences + inference + human minutes + assurance) ÷ tasks completed. Compare it against the same task done the old way, then pressure-test at 3× volume, half the assumed benefit, and a ±30% model price move. Fund it only if it survives all three.
The lab
Price one real workflow per completed task, then find the assumption it depends on most.
Open this lesson, its lab and its quiz
Sources and further reading
- Total cost of ownership — Reference
- Payback period — Reference
- Sensitivity analysis — pressure-testing assumptions — Reference
- AI Index — capability and cost trends — Stanford HAI