How Much Does It Cost to Build an AI Agent for a Business?

It is the first question every buyer asks, and the one most vendors answer badly — usually by quoting a figure that was never going to be the figure.

This post will not hand you a number either, and that is deliberate. A price quoted before anyone has seen your data, your volume or your risk tolerance is a marketing number wearing an estimate's clothes. What is useful is the shape of the range and the handful of things that decide where in it you land. Know those and you can read any quote properly, including ours.

If you are still at the "does this pay for itself at all" stage, that is a different question and it has its own post: The real ROI of AI integration.

The three things that actually move the number

How messy the inputs are. This is the big one and it is almost invisible from the outside. An agent that reads clean structured records from one system is a different project from one that reads PDFs, scanned faxes, WhatsApp threads and three spreadsheets that disagree with each other. Same job description, completely different cost.

Most of the money in a hard build goes into the unglamorous work of making inputs usable: consolidating the knowledge base, deleting the duplicate policies, deciding which version of a document is authoritative. If your data is already clean, you are buying the cheap version of this project. If it is not, that work is unavoidable and no vendor can price around it — they can only decide whether to tell you about it.

What the agent is allowed to do. An agent that drafts and waits for a human is a substantially smaller build than one that acts. The moment it can send, refund, book or update a record, you need an audit trail, an escalation path, a reversibility story and usually a compliance conversation. None of that is optional and all of it is billable.

This is the clearest lever a buyer actually controls. Tighten the boundary and the project gets cheaper and safer at the same time, which is a rare combination.

How you will know it works. Evaluation is the line item everyone cuts and then pays for twice. If you want to know whether the agent is any good, someone has to build the test set — real questions, real expected answers — and run it before every change. Skipping it does not save the money. It moves the cost into the first month of production, where it arrives as rework.

The three costs most quotes leave out

The per-run cost is forever. A build is a one-off. Inference is a utility bill that scales with usage. A few cents per task looks different at ten thousand tasks a day, and the number that matters is cost per successful run, retries included.

Somebody has to own it. Every agent is a system with an on-call rotation whether or not anyone admits it. Budget for the person who reads the logs, updates the corpus when policy changes, and notices when quality drifts. Agents do not degrade in a way that announces itself. They degrade quietly.

The corpus does not stay clean. Whatever you tidied in week one starts drifting in month three: new policies, new products, new edge cases nobody predicted. Maintenance is not a flaw in the plan. It is the plan.

How to read any quote

Ask three questions of whatever number you are given.

What happens to the price if the data is worse than we think? A vendor who never asked about your data has priced the easy case.

What is excluded? Specifically: evaluation, monitoring, handover documentation, and the first three months of maintenance.

What does it cost to stop? If the engagement ends, what do you own, and can it run without them?

A quote that survives all three is worth comparing on price. One that does not is not expensive or cheap yet. It is incomplete.

The question underneath the question

"How much does it cost" is almost always standing in for "is this going to be an embarrassing amount of money for something that does not work?"

The honest answer to that one has very little to do with price. It depends mostly on whether you picked a job worth automating, which is why the first conversation worth having is not about the number. It is about whether the thing should be built at all.

That conversation is AI consulting, and it happens before the build rather than instead of it. When the answer turns out to be yes, the work lands under AI agents.