The real cost of running agents
Token spend is the line item everyone models and the smallest one that matters. The costs that decide the business case sit somewhere else entirely.
Every business case for an agentic system starts with a token cost estimate. It is the easiest number to compute and the least useful one to know, because inference is the one input whose price has moved consistently in the buyer's favour while every other cost line has held or grown.
Model the costs that do not fall. Here is the full picture, roughly in order of how badly each one is usually underestimated.
| Cost line | Why it is underestimated | What controls it |
|---|---|---|
| 01Inference | It is the only line most teams model, so it dominates the spreadsheet and not the budget. | Routing cheap work to small models, caching, and not retrying blindly on failure. |
| 02Human review | Reviewer time is charged to an existing team, so it never appears in the project cost. | Per-action handoff design and a plan to graduate categories out of review. |
| 03Data plumbing | Integrations are scoped as one-off build and behave as permanent maintenance. | Fewer integrations, owned interfaces, and refusing scope that needs a new connector. |
| 04Evaluation | Treated as a launch gate, so nobody budgets the standing cost of keeping it current. | A named owner, an automated suite, and a fixed cadence rather than ad-hoc panic runs. |
| 05Incident cost | Priced at zero until the first bad action, then priced emotionally rather than numerically. | Containment limits, reversal paths, and recovery time measured as a standing metric. |
Inference is the only cost in an agentic system that has reliably fallen. Budget as though every other line is permanent, because it is.
The failure-mode multiplier
Agentic systems have a cost characteristic that request-response software does not: a loop that fails badly can consume a multiple of its expected spend before anyone notices. A retry policy with no ceiling, a planning step that re-plans on every tool error, or two agents handing work back and forth will all produce a bill that looks like a bug because it is one.
Cap steps per task, cap total spend per task, and alert on the tail rather than the average. Mean cost per run is a comforting number that hides exactly the behaviour that will hurt you.
Model the saving honestly too
The other half of the business case is usually softer than it is presented. Time saved is only money saved if the time is redeployed or the headcount does not grow, and a system that removes forty percent of a task from ten people rarely removes four people. State the saving in the terms your finance team would accept: capacity absorbed without hiring, cycle time reduced, error rate reduced, or a queue that no longer needs a weekend.
A business case built on a defensible operational metric survives its first review. One built on notional hours saved gets challenged in month three, which is precisely when the system is working but has not yet compounded.
Want this graded for your own stack?
A systems audit runs your operation against exactly these dimensions and hands you the report.
Request a systems audit