A monthly AI invoice tells a finance team what was spent. It does not say what one resolved claim, one qualified lead or one drafted memo cost to produce. As AI moves from pilots into production, that blind spot decides which AI products scale profitably and which get shut down.
Token costs compound where nobody is looking.
The same patterns appear in almost every system we review. Full documents are re-sent on every call. Retrieval pulls whole files when a paragraph would do. Failed calls retry without limits. Agents call other agents in loops. A premium model classifies emails that a far smaller model could sort.
None of this is visible on an invoice grouped by vendor and project. Usage grows faster than value because each new workflow copies the habits of the last one.
The unit that matters is the cost of each useful result.
Define the useful result first: a resolved request, an approved draft, a correct extraction. Then add every cost behind it, including tokens, model fees, compute and the human review it required, and divide by the number of useful results. That figure can be compared with the cost of doing the work manually and tracked as volume grows.
We apply this to our own systems. One of them cost $17.60 a day, and we could not say which tasks drove the spend. Tracing showed repeated context and oversized models doing routine steps. After routing, caching and trimming context, with quality checked after every change, the same system costs between $4 and $5 a day, a reduction of about 70 to 75 percent.
We also learned a discipline lesson. Our own presentation claimed that input tokens drive 70 percent of cost. On inspection, that ratio depends heavily on the workload, so we corrected it. Savings figures rarely transfer from one system to another; they have to be measured.
If you cannot say what one good answer costs, you cannot say whether AI is paying off.
Five levers lower the cost of each result without lowering quality.
Route routine tasks to smaller models and reserve the largest for the hard cases. Cache context and results that repeat. Send only the text a task needs rather than the whole document. Cap retries and agent loops so a failure cannot run up a bill overnight. Give every workload a budget, an alert and an owner.
Unit economics will decide which AI products survive.
Before a pilot expands, forecast its cost per result at production volume. Put finance and engineering on the same number and review it monthly. Products whose unit cost falls as they scale deserve more investment. Products whose unit cost rises with usage need redesign before they grow.