What a custom AI agent costs in 2026, and what drives the price.

Start with the job, not the model

The cost of an agent is the cost of the system around the model: the data it reads, the tools it can call, the tests that prove it works, and the monitoring that tells you when it stops working. The model itself is a small line item that you can swap. When a quote is very low, one of those pieces is usually missing, and you find out which one in month three.

So the first question in scoping is not "which AI" but "which queue". A support inbox, a lead list, an intake form, a catalog: each has a volume, a set of systems it touches, and a cost when it goes wrong. Those three things set the price.

The five things that move the price

1. How clean the data is

An agent that answers from a tidy policy document and an orders API is a short build. An agent that has to reconcile three spreadsheets, a legacy database and a shared inbox is a longer one, because most of the work is making the data answerable. Ask any vendor how they will ground the agent in your data. If the answer is vague, the price is a guess.

2. How many systems it touches

Reading is cheap; acting is not. An agent that looks up an order is simpler than one that issues a refund in Stripe, edits the order in Shopify, and updates the CRM. Every action needs an integration, permission boundaries, logging and a way to reverse it. Count the tool calls in the workflow and you have a rough proxy for build effort.

3. What happens when it is wrong

A wrong answer about store hours is a nuisance. A wrong refund, a wrong dosage question or a wrong legal disclosure is a liability. The higher the cost of an error, the more the build spends on evaluation sets, confidence thresholds and human handoff. This is the part that separates a demo from something your operations team will trust.

4. Volume

Volume changes the running cost more than the build cost, but it also changes design. A thousand conversations a day needs caching, rate limits and model routing so a spike does not become a bill. A hundred a week does not.

5. Compliance

Wellness, telehealth, finance and anything touching personal data adds access controls, retention rules, disclosures and counsel review. It is not optional, and it belongs in the proposal rather than in a change order.

What a fixed-price proposal should include

We only work on fixed scope, and the proposal is where a client learns what they are buying. Whoever you hire, look for these items in writing:

• The exact workflow the agent handles, and the cases that are out of scope.
• The data sources it will read and the systems it will act on.
• An evaluation set: the questions and cases it must get right, agreed before tuning starts.
• The handoff rule: what happens when confidence is low, and who receives the conversation.
• Logging and reversibility for every action it takes.
• Monitoring after launch: resolution rate, response time, escalation rate, accuracy.
• Ownership: whose repository, whose accounts, whose data.
• A ship date and a support window.

If a proposal is a single number and a paragraph, you are being sold the demo.

What it costs to run

Running costs are model usage, hosting and whoever watches the dashboard. Model usage for a typical support agent is small relative to the labor it replaces, and it can be capped. Hosting is a few dollars to a few hundred a month depending on volume. The real running cost is attention: someone reviewing escalations, updating the knowledge base when policy changes, and re-running the evaluation set after a vendor updates an API. That can be your team or a retainer, but it should be somebody.

How to estimate payback before you call anyone

Take the hours per week your team spends on the queue, multiply by their loaded hourly cost, multiply by 52, and multiply by the share of the work an agent can realistically absorb. For a support queue that share is often 60 to 70 percent; for lead qualification it is higher; for anything clinical or legal it is lower by design. The calculator on our homepage does the arithmetic. One D2C support agent we run resolves 68 percent of tickets with a 1.4 second average response, which is the kind of number the proposal should be able to defend.

If the annual recoverable cost is several times the build price, the decision is easy. If it is close, start with a narrower agent and expand it after the first month of data.

The short version

A custom agent is priced by the data it needs, the systems it touches, the cost of being wrong, the volume, and the compliance load. Insist on a fixed scope with an evaluation set, a handoff rule and ownership in writing. Estimate payback from hours, not from hope. And treat month three, not launch day, as the test of whether you bought the right thing.

How we build AI agents · AI automation in Orange County · Get a fixed-price proposal