What does an AI agent cost? The five cost drivers

"How much does an AI agent cost?" has the same answer as "how much does a house cost?" — it depends, but not mysteriously. Five things decide almost all of it. Understanding them lets you scope a first version sensibly and compare quotes on equal terms.

1. How many systems the agent touches

Every tool an agent can use — the CRM, the order system, the calendar, the ticketing tool — is an integration to build, secure and test. One system with a good API is quick. Five systems, one of them an old internal database with no API, is a different project.

To keep it lean: start with the one or two systems where the repetitive work actually happens.

2. How much the agent is allowed to do on its own

An agent that only suggests actions for a person to approve is far simpler to make safe than one that acts alone. Autonomy needs more evaluation, more guardrails and more monitoring.

To keep it lean: launch with approvals on every action, and remove them one at a time as the logs prove they are safe.

3. The state of your data

Agents are only as good as the data they read. If customer records are duplicated, documents are out of date or key fields are empty, some of the budget will go into cleaning data rather than building the agent. That is not wasted money — it would have hurt you anyway — but it should be planned.

4. How you measure "good enough"

A serious agent needs an evaluation set: real past requests with the correct outcome, run automatically on every change. Building that set takes effort from people who know the work. Skipping it is the most common reason agents look impressive in a demo and disappoint in production.

5. Running costs after launch

Budget for more than the build:

• Model usage: you pay the AI provider per use, so cost grows with volume. Choosing the right model for each step — a smaller, cheaper model for simple classification, a stronger one only where reasoning is needed — keeps this under control.
• Hosting and monitoring: logs, dashboards and alerts for errors and unusual spending.
• Maintenance: your processes, tools and AI models all change. Plan for regular updates.

How to compare quotes

When you get proposals, ask each vendor the same questions:

• Which systems will the agent connect to in the first version?
• Which actions need human approval at launch?
• How will you measure accuracy before go-live, and on what data?
• What will monthly running costs look like at our volume?
• Who owns the prompts, code and evaluation set at the end?

A quote that cannot answer these is a quote for a demo.

The cheapest way to start

Scope one workflow, with approvals on, measured against how long it takes people today. That first version tells you — with your own numbers — whether to expand. We give a written estimate with scope, timeline and cost after a 30-minute discovery call; see how we approach AI agent development.