How much does an AI agent cost to build and run?

An AI agent has two bills: a one-time build cost and a monthly running cost. Most quotes show you the first and stay quiet about the second, and the second is the one that decides whether the project makes sense over a year. This article explains what drives each, where costs quietly run away, and what to ask any vendor, including us. When you have both numbers, the payback calculator turns them into a payback period.

We will not print a single made-up number, because a price without your specifics is marketing. What we can do is show you every line item, so no bill surprises you.

What drives the build cost

Four things, in rough order of weight:

How wrong the agent is allowed to be. This is the biggest driver and the least discussed. A drafting tool a person reviews can be built quickly. An agent that talks to your customers needs rules it cannot break and a test suite that proves it. One of our production agents ships with a guardrail suite that tries seven different ways to trick it into quoting a price the business cannot fulfil, plus a live check on every reply that hands an ungrounded answer to a person. Writing those guardrails is real work, and skipping them is how agents end up in the news.

Integrations. Each system the agent touches, your CRM, your inbox, your calendar, your accounting tool, adds connection work and a failure mode to handle. Two integrations is a small project. Seven is not.

The messiness of the input. Extracting facts from a tidy web form is easy. Extracting them from rambling voice notes and forwarded email chains is where the effort goes.

Who has to be able to change it later. If your own team should be able to adjust rules and templates, that interface has to be built too. It is worth it, and it is a line item.

What the monthly running cost is made of

This is the bill that quotes hide. It has four parts:

Model usage. What the agent pays per call to the model provider. For most business agents this is genuinely small, and it is usually the only running cost a vendor mentions, precisely because it is small.

Hosting and infrastructure. The servers, the database, the scheduler. Modest, predictable, permanent.

Monitoring. Something has to watch the agent. The expensive failure is the quiet one: an API account runs dry, every run fails, and nothing looks different from a quiet day. Our systems page on the exact errors that mean an account is empty, and a watchdog confirms the critical jobs are still installed. That watching costs something to build and a little to run, and an agent without it is not cheaper, it is just unwatched.

Maintenance and change. Models get updated, APIs change, your business rules change. Budget for a few hours a month, or you are budgeting for a slow decay.

The costs nobody mentions

Evaluation. When you change the agent, how do you know it did not get worse? A proper evaluation suite is built once and run on every change. Ours grade our own products against live data. Without one, every change is a guess.

Incidents. Something will break: a provider outage, a price change, a weird input. The cost is small if the system logs what it does and a named person owns it, and large if neither is true.

The runaway loop. The classic agent failure: it retries, or calls itself, or loops on a bad input, and burns through credit at speed. The fix is a spending cap and a per-run cost log, wired in from day one. Ask any vendor where the cap lives. If there is no cap, your monthly cost has no ceiling.

Build, buy, or both?

The plain rule: buy the generic layer where an off-the-shelf product already does the job, and build the part that encodes how your business actually works, your rules, your tone, your prices, your boundaries. Most good outcomes blend both. Building everything from scratch for a standard problem wastes money; buying a generic tool for a specific problem wastes it slower.

And before either: check whether you need an agent at all. Many agent projects should have been simple automations, at a fraction of both bills.

What to ask any vendor, including us

1. What is the monthly running cost, all four parts, in writing?
2. Where is the spending cap, and what happens when it is hit?
3. Who owns the accounts? (Every account should be in your name. This is our standing rule and it should be your standing demand.)
4. What does the test suite check, and can we see it run?
5. What happened in your last incident, and what changed because of it?

A vendor who answers all five plainly is selling you a system. A vendor who cannot is selling you a demo with a subscription attached.

The honest summary

The build cost is what you negotiate. The running cost is what you live with. Insist on seeing both before you sign anything. Our two-week assessment prices both for your specific case, at a fixed fee, credited against any build. If the numbers do not work at that price, it shows you what would make them work.