What an AI agent actually costs to run

The first question every client asks about an AI agent is what it costs to run. The honest answer is that the model price is the least interesting part of the bill, and quoting it alone is how people end up surprised in month two.

Count conversations, not tokens

A support agent handling WhatsApp for a retailer runs perhaps three to six model calls per conversation: one to understand the message, one or two to look things up in your catalogue or order system, one to draft the reply, and sometimes one to check that reply against your rules before it goes out.

Multiply that out and a typical resolved conversation lands somewhere between one and three US cents on a mid-tier model. Ten thousand conversations a month is a few hundred dollars. That is the whole scary number, and it is smaller than one part-time salary.

The costs nobody quotes you

• Retrieval. Keeping your catalogue and policies searchable is cheap but not free, and it needs rebuilding whenever the catalogue changes.
• Escalation. Every conversation the agent hands to a human still costs a human. Budget for 10–20% handoff and be suspicious of anyone promising zero.
• Evaluation. You need a test set that catches a prompt change making things worse. This is the line item clients cut first and regret first.
• The retry tax. Failed calls, timeouts and rate limits all get retried, and retries bill.

Where the money is actually saved

The savings are rarely in headcount. They are in response time. An agent that answers at two in the morning captures orders that would otherwise go to whoever replied first at nine. In our own sample support desk, the figure that moves the business is not the cost per conversation — it is that 62% of them never reach a person at all.