Every enterprise AI programme has a day 100. For the first hundred days the bills are absorbed by enthusiasm: the pilots are strategic, the tokens are cheap individually, the platform fee is an investment. Then a finance partner does what finance partners do, and asks two questions that end the honeymoon. What does this cost us per decision? And what does a decision return?
Finance teams are converging on a name for this discipline, AI unit economics, and it is overdue, because very few organisations can answer either question. The costs arrive as infrastructure line items: tokens, seats, GPU hours, vendor subscriptions. The returns arrive as anecdotes and screenshots. Between the two sits no ledger of any kind.
The unit is the decision
The instinctive unit of AI accounting is the token or the seat, because that is how the invoices are denominated. Both are the wrong unit, in the way that "cost per kilowatt" is the wrong unit for a factory. The thing the spend exists to produce is a changed decision: a forecast someone acted on, an exception someone resolved, an intervention chosen over a default.
Denominate in decisions and the questions become answerable, and uncomfortable:
- Cost per decision. Total programme cost divided by decisions actually influenced. A copilot with heavy usage and no decision attribution has an infinite cost per decision, which is the honest number for a great deal of current spend.
- Return per decision. What did acting on the system's call earn against not acting? This requires the outcome to be recorded against the decision, which is bookkeeping most programmes never built.
- The substitution line. When is an agent cheaper than an analyst hour, and when is it an expensive way to generate work the analyst must then check? Without a cost-per-decision figure, the debate runs on vibes in both directions.
Gartner's finding that 45% of CFOs' AI investments lean toward productivity while 20% lean toward decision quality is a unit-economics verdict: cost per task compressed, cost per decision unexamined.
Why the books were never kept
The missing ledger is not an oversight. It is a structural consequence of how programmes are set up. Pilots are scoped to demonstrate capability, not to attribute outcomes, so the measurement is bolted on afterwards, which means never. MIT NANDA's State of AI in Business research found 95% of pilots showing no measurable profit and loss impact, and the operative word is measurable: much of the missing value is not absent, it is unattributed, which for a CFO is the same thing.
The repair is unglamorous. Route the AI's outputs to named decisions with owners. Record what was recommended, what was decided, and what happened. Meter the costs against the same decision records. From that point the day-100 conversation changes character entirely: the programme has a cost curve, a return curve, and a defensible answer to "why should this budget survive?"
The organisations that build this ledger first will also, not coincidentally, be the ones whose AI improves fastest, because a system whose decisions are scored is a system that can learn which of its calls deserve trust.
Enthusiasm funds the first hundred days. Only unit economics funds the next thousand.
Prophesee keeps those books by construction. Every prediction is wired to an owned decision and every decision to an outcome, so cost and return per decision are queries, not archaeology. To start the ledger on your own data, start here.