When a client asks what an AI agent will cost, they usually mean the model provider's bill. In our projects that is rarely more than a fifth of the running cost. The rest is the work around the model: hosting, evaluation, the people who review edge cases, and keeping the agent correct as your business changes.
This is the budget we walk through before any build starts. The numbers come from a worked example, an agent that triages around 400 inbound requests a working day. Your volumes will differ, but the shape of the bill rarely does.
The example agent
The agent reads each incoming email and attachment, classifies the request, pulls out about 12 fields, checks them against 2 internal systems and either completes the task or queues it for a person with a summary. It handles roughly 100,000 requests a year.
The 12-month budget
| Line | What it covers | Year-1 share |
|---|---|---|
| Build | Discovery, tools, integrations, evaluation set, launch | 45% |
| Human review | Staff time on the cases the agent sends to a person | 20% |
| Model usage | Tokens for reading, reasoning and writing | 12% |
| Support and changes | Fixes, new request types, prompt and tool updates | 13% |
| Hosting and monitoring | Servers, queues, storage, logging and alerts | 6% |
| Evaluation runs | Re-testing against the evaluation set on every change | 4% |
Build is a one-off, and it's mostly integration
Most of the build goes on the tools the agent uses: the narrow, permissioned actions such as "look up order" or "create ticket" that connect it to your systems. The prompt is a small part of the work. A well-built tool layer is also what lets you swap models later without a rebuild.
Human review is the line people forget
An agent that's right 92% of the time still sends 8,000 cases a year to a person in our example. If each takes 3 minutes, that's 400 hours of staff time. The fastest way to reduce this line is better confidence thresholds, not a bigger model.
The cheapest agent to run is the one that knows when to ask for help.
Model usage is volatile, but small
Token prices have fallen sharply every year since 2023, and the right model for a triage task is often a mid-sized one. We set a monthly ceiling and alert well before it's reached. Caching repeated instructions and documents typically cuts this line by a third.
3 ways to keep the bill down
- Start with one request type. Prove the economics on your highest-volume case before adding the long tail.
- Invest in the evaluation set. 200 real, labelled examples let you change prompts and models safely, and stop regressions reaching customers.
- Route by difficulty. Send easy cases to a small, fast model and only hard ones to a larger model.
What we need from you to estimate
Send us 4 numbers and we'll return a 12-month estimate within a week: requests per day, the share you'd want automated, the average staff time per request today and the systems the agent must read or write.
requests_per_day: 400
target_automation: 0.80
minutes_per_request: 6
systems: [email, crm, order_db]