Managing the Cost of Agentic AI

A single model call has a cost you can look up. An agent does not. Give a system the freedom to plan, call tools, read the result and decide whether to try again, and one user request becomes an unknown number of billed operations, decided at runtime by the agent rather than in advance by you. That is the whole promise of agentic AI, and it is also the reason finance teams open the invoice at the end of the month and cannot explain it.

One request, an unknown number of operations

A chat completion is a transaction: one input, one output, one line on the bill. An agent run is a loop. It plans, calls a tool, reads what came back, decides whether that was enough, and goes round again until it thinks it is finished or a limit stops it. Two users asking what looks like the same question can generate wildly different bills depending on how cleanly the first attempt worked out. The unit the customer sees, and usually the unit you priced, is the request. The unit you are billed for is the step, and nothing in the price list tells you how many steps a given request will take.

The context grows, and you pay for it every time round

The part that surprises people is not the number of calls but what each call carries. At every step the agent sends back the accumulated history: the original instruction, the tools it has already called, what each one returned, its own reasoning so far. A ten step run does not cost ten times a single call, it costs more, because the input grows with each pass. Add retrieved documents and tool outputs that are themselves long, and the context can become the dominant cost of the run. Caching and trimming help, and both are product decisions that have to be made deliberately rather than discovered in an invoice.

Failure is a cost line, not an exception

Agents retry. A tool times out, an API returns something malformed, a step produces output the next step cannot use, and the loop goes round again. Every one of those attempts is billed whether or not the run eventually succeeded, and the runs that fail completely are the most expensive of all: full cost, no delivered value. In a well instrumented product the failed and retried runs are a visible line that someone owns. In most products they are invisible, folded into a total that gets read as the cost of the successful work.

Cost per completed task is the number that matters

The useful measure is not cost per token or cost per call. It is the fully loaded cost of a completed task: every step, every tool call, every retry, every failed attempt that preceded the successful one, traced to the customer and the feature that caused it. That is an activity costing question, and it behaves like one. Define the activity the way the business defines it, a resolved ticket, a drafted contract, a reconciled invoice, then trace the resources one unit of it actually consumed. Once the cost of a completed task is known, everything downstream becomes answerable: what to charge, which customers are profitable, and whether the agent is cheaper than the process it replaced.

An agent does not consume a price per token. It consumes a price per attempt, and it decides how many attempts to make.

What to put in place before the next agent ships

Before an agentic feature goes to general availability, give it three things. A hard ceiling on steps and spend per run, enforced in the product rather than reviewed in the monthly bill. A record for every run of which customer, which feature and which outcome it belonged to, so cost can be attributed rather than averaged. And a cost per completed task reported next to the success rate, because the two only mean something together. An agent that succeeds ninety per cent of the time at four times the budgeted cost is not a working feature, and no invoice total will tell you that on its own.

Related reading: the cost of agentic AI and what it really costs to serve an AI feature.