Corahvel
AI Solutions

How Much Does an AI Agent Cost to Build and Run?

Corahvel Team3 min read

The question gets asked early and answered badly. Most quotes cover the build and go quiet about what happens afterwards, which is how a business ends up with a working agent and an unexpected monthly bill.

There are two numbers. Both are knowable before you start.

The build

What drives the build cost is not the model — everyone uses the same handful — but how much of your business the agent has to understand and reach into.

A single-channel agent answering questions from a fixed set of documents is the small end. The work is assembling the content, grounding the agent in it, writing the escalation rules, and testing. Scope creeps when the agent needs to read from live systems, when several teams each want different behaviour, or when the source material is scattered across inboxes and nobody owns it.

That last one is the most common cause of a build costing more than quoted, and it has nothing to do with AI. If your policies live in six versions across three drives, the first job is deciding which version is correct — and that is your decision to make, not ours.

The running cost

This is the part that gets skipped. Every conversation sends text to a model provider and you are billed for it, usually per million units of text processed.

The practical way to forecast it: take your actual monthly enquiry volume, assume a realistic conversation length, and multiply. For most small and mid-sized businesses handling hundreds rather than millions of conversations a month, the figure lands somewhere between a modest subscription and a junior salary — but the point is that it is arithmetic, not a mystery. Anyone proposing an agent should hand you that calculation before the build, and we do.

Two things move it. Grounding the agent in a large document set means more text per conversation, so retrieval should fetch the relevant passages rather than everything. And a chattier agent costs more than a concise one, which is one of several reasons to keep it concise.

What makes it cheaper over time

The first agent is the expensive one, because that is when you build the knowledge base, the escalation rules, and the test set. A second agent on a different channel reuses most of that.

Costs also fall on their own. Model pricing has dropped repeatedly, and because we keep the model behind an interface with accounts in your name, moving to a cheaper or better provider is a configuration change rather than a rebuild. An agent that locks you to one vendor will quietly become the most expensive option available.

When the honest answer is "don't"

If the task is deterministic — the same input always produces the same output — a plain automation does it for a fraction of the cost, runs faster, and is never wrong. A fair number of the AI projects we're asked to quote are really automation projects, and saying so is part of the job.

If you want the arithmetic for your own volumes, that is what the first conversation covers. See AI agents and AI solutions, or get in touch with a rough monthly enquiry count and we'll work it through with you.