A model provider's invoice is the easiest cost of an AI agent to see and usually one of the smaller ones. The larger costs are people's time: reviewing held items, handling exceptions, keeping sources current, certifying agents and containing incidents. A business case built on the model bill alone will look excellent and prove wrong.
Why does the model bill dominate the conversation?
Because it is metered, itemised and arrives monthly. Usage-based pricing makes it easy to calculate a cost per case and compare it with the cost of a person doing the same case. The comparison is flattering, and it is the one that tends to reach the business case.
It is also the cost technology teams control, so it is the one they optimise: cheaper models, shorter prompts, fewer calls. Those savings are real but often small next to costs that sit in operations and never appear on a technology budget.
What does an agent actually cost to run?
Group the costs by where they fall, not by who sends the invoice.
| Cost | What drives it | Where it usually sits |
|---|---|---|
| Model usage | Volume of cases, length of context, number of steps per case | Technology budget |
| Platform and integration | Connecting to systems of record; maintaining connections as they change | Technology budget |
| Review of held items | Number of items held, minutes per item, approver seniority | Operations, often unbudgeted |
| Refused work and exceptions | Cases the agent cannot or may not handle, routed back to people | Operations |
| Source upkeep | Keeping procedures, terms and policies accurate for the agent | Process owners, rarely costed |
| Certification and renewal | Building case sets, running tests, reviewing results each cycle | Operations and risk |
| Agent operations | Watching queues, verdict patterns, widening and narrowing | A dedicated role, or hidden in other jobs |
| Incidents and remediation | Finding affected cases, putting customers right, re-testing | Unpredictable; often large when it happens |
Most of the rows below the second are people's time. That is the bulk of the cost for most agents in regulated work, and it is the part most business cases leave out.
Why does depth change the cost profile so much?
Because depth decides how much of the agent's work passes through a person. An agent at draft produces work that someone must read and send. An agent executing routine work within bounds sends most of it without review, holding only what exceeds a limit.
So the same agent, doing the same volume, can cost very different amounts depending on its depth. At draft, review time dominates. Widened to execute within limits, review time falls sharply, but certification and monitoring costs rise, because executing work needs stronger evidence and closer watching.
This gives cost a governance dimension. The cheapest way to run an agent is to widen it as far as evidence supports, and no further. Widening ahead of evidence looks cheaper until the first incident. Keeping an agent at draft long after the evidence would support execution wastes reviewer time indefinitely.
Where do hidden costs usually surprise people?
Three places stand out.
Reviewer time per item. Plans often assume a review takes a fraction of the time the original work took. For simple drafts that may hold. For items involving judgement, such as a refund above a limit or a payee from an unverified source, review takes nearly as long as doing the work, because the reviewer has to check the very thing that made the agent stop.
Exceptions concentrating. Agents take the easy cases. The cases left with people are harder on average, so the cost per human-handled case rises even as the number falls.
Renewal. Certification is not a one-off. If certificates lapse and the agent must be re-tested on fresh cases each quarter, that work recurs for the life of the agent.
How should you track cost per agent?
Per agent and per capability, including people's time. A cost tracking template for AI agents gives a structure; the essentials are to record review minutes against the capability that produced the held item, exception handling against the capability that refused the case, and certification effort against each certificate renewed. Our playbook on tracking the cost of each AI agent covers how to collect those figures without a heavy timekeeping burden.
Take an agent that explains customer bills line by line, as in explaining a bill line by line with an AI agent. The model cost per explanation may be trivial. If every explanation is drafted for a person to send, the real cost per case is mostly the reviewer's time. If the agent is certified to send standard explanations itself and holds only disputed bills, the cost shifts towards certification and dispute handling. Tracking both tells you which depth makes sense.
How do you reduce the real cost?
Not mainly by choosing a cheaper model. The levers that move total cost are operational: hold fewer items by widening on evidence, make review faster by showing approvers the deciding bound and the sources, cut exceptions by fixing the sources and routing that cause them, and automate the mechanics of re-certification. Our guide on reducing the running cost of AI agents goes through each lever, and our overview of what it costs to run an AI agent sets out the starting estimate.
What is the trade-off?
Costing agents fully makes business cases less impressive. Including reviewer time, certification and incident reserves can turn a striking saving into a modest one, and some proposed agents will not survive the arithmetic.
That is the purpose. An agent that only pays off if its review, exception and certification costs are ignored is an agent that will disappoint. Rebuild the business case for your next agent with every row of the table above, estimate how cost changes as depth widens, and decide what evidence would justify each step. The deployment approach this cost model follows is set out in our guide to deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Is the model bill ever the biggest cost?
It can be for agents that read or draft very high volumes of simple items with little review, or that use long contexts and many steps per case. For agents in regulated work with held items and certification, people's time usually outweighs it.
Should review time be charged to the agent's budget?
Yes, at least for tracking. If review time sits in an operations budget while the agent's savings are claimed by technology, nobody sees the true cost and nobody has a reason to reduce it.
How should incident costs be estimated before any incident happens?
Use a reserve based on the depth and volume of the agent's riskiest capabilities and the effort it would take to find and put right affected cases. A good decision record lowers that effort, which is one way governance reduces cost.