The AI unit economics trap
Most AI SaaS teams can see token spend. Far fewer can answer the executive question: which AI workflows create margin, and which ones quietly destroy it?
Cost per request is a weak metric
Cost per request is useful for engineering, but it does not map cleanly to business value. A support triage request, a legal document review, a sales research task, and an internal coding agent can all have similar token costs and radically different economics.
The better unit is the business outcome: cost per resolved ticket, cost per qualified account, cost per document processed, cost per customer served, or cost per successful agent run.
The margin leak usually has three sources
- Model mismatch: expensive models are used for low-risk classification, extraction, or routing tasks.
- Workflow rework: failed generations, escalations, QA correction, and retries are excluded from the cost model.
- Shared spend: cloud, vector databases, observability, SaaS tools, and LLM APIs are not allocated back to product lines or customers.
What executives need to see
A CFO or CTO does not need another provider dashboard. They need a decision table: keep, optimize, route, cap, reprice, or shut down. Each AI workflow should have a cost baseline, a value assumption, an owner, and a control plan.
The Olive-One view
The first diagnostic should not be a full platform migration. It should normalize the available spend data, model cost per outcome, identify margin exposure, and produce a 30-day action plan. That is enough to make the first economic decisions before building heavier infrastructure.