What an AI workflow costs once review and rework are counted

Why the token price answers a narrower question

The FinOps Foundation describes how AI unit economics usually begins with cost per token and matures towards outcome measures such as cost per agent action or cost per case deflected. The token price is a resource metric. It tells an engineer whether a design is efficient, but not whether the process got cheaper.

Even the model line is harder to predict than a price list suggests. A case may take several calls, retries and long context, and an agent can loop. The AWS Generative AI Lens recommends stopping conditions for long-running agent workflows so that the maximum cost of a run can be predicted. Prices also move, and the FinOps Foundation's AI paper notes that they can change sharply in either direction.

The lines in a cost per handled case

• Model usage for every step of a case, including retries.
• A share of the platform: hosting, retrieval, logging and monitoring.
• Review time: minutes per reviewed case, times the share of cases reviewed, times the loaded hourly cost of the reviewer.
• Escalations: cases passed back to people and handled as they are today.
• Corrections: errors found after completion, including the extra work they cause further down the process.
• Fixed costs: the build spread over an agreed period, plus maintenance, evaluation runs and model upgrades.

Divide by cases completed, not cases attempted. A case that was escalated is still a handled case, but its cost is the manual cost plus whatever the system spent before giving up.

A worked example with its assumptions

Every number below is hypothetical and chosen for easy arithmetic. Assume 10,000 service requests a month. Today each takes 12 minutes at a loaded cost of 48 euros an hour, so 9.60 euros per case and 96,000 euros a month.

• Model and platform: 0.30 euros per case, 3,000 euros in total.
• The system completes 80 percent of cases, each reviewed in 2 minutes: 8,000 times 1.60 euros, 12,800 euros.
• 20 percent are escalated and handled manually: 2,000 times 9.60 euros, 19,200 euros.
• 3 percent of completed cases need a 20-minute correction later: 240 times 16 euros, 3,840 euros.
• Fixed: a 120,000 euro build spread over 24 months plus 4,000 euros of monthly upkeep, 9,000 euros.

The total is 47,840 euros, or 4.78 euros per handled case against 9.60 today. Model and platform account for about 6 percent of it. In this example, review and escalations together make up about two thirds of the cost.

Which assumptions move the result

Change one assumption at a time. If the escalation rate rises from 20 to 35 percent, the cost per case goes from 4.78 to 5.91 euros. If review takes 4 minutes instead of 2, it reaches 6.06 euros. Halving the model and platform cost lowers it only to 4.63 euros. In a design like this one, measured review minutes and escalation rates decide the business case. In a high-volume agent with long loops and little review, the model line can dominate instead, so run the calculation for your own design.

Collecting the numbers during the pilot

The FinOps Foundation calls planting instrumentation during development the cheapest early step: tag resources and emit a usage counter so that the denominator exists later. For a process, that means tagging every model call with a case ID, recording when a reviewer opens and closes a case, logging each escalation and correction with a reason, and taking fixed costs from contracts. Keep the figure as a trend for one process rather than comparing it with unrelated teams.

Comparing designs on the same formula

Run today's process, a bought product, your own build and a simpler rules-based workflow through the same calculation. Any saving stays conditional until the pilot has measured escalation, review and correction rates on real cases. Benefits that are not cost, such as faster turnaround, belong next to the figure rather than inside it. When we build process agents, we report this figure per handled case in euros, next to the share of cases escalated and the share the agent got wrong.

Budgeting for the people who review AI workWhat an AI pilot must prove