Beyond model tokens and hosting, an AI agent's running costs include evaluation runs on every change to prompts, documents or models; monitoring and logging; a few hours of maintenance a month; and a model migration whenever the provider retires the model it runs on. Anthropic, for example, gives at least 60 days' notice before retiring a model. In our worked example, an agent with a $340 monthly infrastructure bill costs about $1,013 a month once those are included, mostly in people's time. Budget infrastructure, upkeep and change as three separate lines, and agree who owns each before launch.
Why does an AI agent cost more to run than the token bill suggests?
Because the token bill is the only running cost that arrives as an invoice. The rest - re-running tests, watching for failures, updating documents and prompts, and moving to a new model when the old one is retired - turns up as people's time, spread across months, and is easy to leave out of a budget that was approved on the model provider's price list.
None of these costs is large on its own. Together, in the worked example below, they come to roughly 3 times the visible infrastructure bill. Knowing that before launch is the difference between a system that is budgeted properly and one that gets switched off in its first cost review.
What does an AI agent really cost to run each month, with everything included?
Our cost guide works through an internal operations agent that answers around 8,000 questions a month for about $340 in model, embedding and hosting costs. Here is the same agent with the hidden costs added. The first three lines come from that example; the rest are assumptions, stated in the table, that you should replace with your own.
| Cost | Basis | A month |
|---|---|---|
| Model tokens | From the cost guide's worked example | $210 |
| Embeddings | From the cost guide's worked example | $25 |
| Vector store and hosting | From the cost guide's worked example | $105 |
| Evaluation runs | Assumed 8 runs at $6 each | $48 |
| Monitoring and logging | Assumed tooling and log storage | $50 |
| Maintenance | Assumed 6 hours at $75 | $450 |
| Model migration | Assumed 20 hours every 12 months, spread monthly | $125 |
| Total | $1,013 |
On these assumptions, the infrastructure bill is about $340 and the real cost is about $1,013. Most of the difference is people's time, which is why it rarely shows up in the original business case. Against the roughly 40 hours a month of staff time the agent in the cost guide replaced, it still pays for itself comfortably - but that is a conclusion you want to reach with the full number, not the partial one.
How should you budget for running an AI agent?
Budget three lines, not one, and agree who owns each before launch.
- Infrastructure - model, embeddings, hosting, monitoring. Scales with usage. Set an alert at the level that would surprise you.
- Upkeep - evaluation runs and a fixed number of maintenance hours a month, from your team or a support arrangement with whoever built it.
- Change - a model migration at least once a year, budgeted as a small project with its own evaluation and rollback plan.
If you are comparing this against a SaaS product that charges per resolution, include all three lines on the build side. Our three-year comparison of custom and SaaS agents does exactly that.
Who should carry the hidden running costs: you, the builder or a support plan?
Whoever can do the work cheapest and will still be there next year. There are three workable arrangements, and the wrong one is no arrangement at all - the default when a build contract ends at handover.
Your own team
Best when you have engineers who were involved in the build and will keep the agent as part of their remit. It needs a recorded handover, runbooks and the evaluation set in your repositories, so the knowledge does not leave with the builder.
A support arrangement with the builder
Best when you do not have AI engineering in-house. Agree a fixed number of hours a month, what they cover - re-indexing, evaluation runs, prompt fixes, model migrations - and how quickly someone responds when it breaks. A fixed monthly figure is easier to budget than ad-hoc invoices, and it makes the upkeep line visible.
An embedded engineer
Best when AI is becoming a regular part of the product and the agent is one of several systems. An engineer working in your team covers upkeep for all of them, alongside new work. Our dedicated AI engineers work this way.
Whichever you choose, write it down before launch. An agent nobody is paid to maintain is maintained by nobody, and degrades quietly until it is switched off.
What are the hidden running costs of an AI agent?
Beyond tokens and hosting, an AI agent costs evaluation runs on every change, monitoring and logging, a few hours of maintenance a month, and a model migration whenever the provider retires the model it runs on. In our worked example, those take a $340 monthly infrastructure bill to about $1,013. Budget for infrastructure, upkeep and change as three separate lines, and build so the last two are cheap.
For what an agent costs to build in the first place, see what AI agents actually cost, or AI agent development for what we include.