How much does an AI agent cost?

The short answer

For a mid-market company automating one real decision, expect a functional pilot in the range of tens of thousands of dollars, and a first use case running in production somewhere in the mid five figures to low six figures, depending almost entirely on how deep the integration goes.

Any number quoted without knowing which systems it has to touch is a guess. The model is close to free. The integration is the project.

The four buckets nobody separates

Almost every disappointing AI budget comes from pricing one bucket and discovering the other three later. Build is the agent itself: the logic, the prompts, the tools, the interface. Integration is connecting it to the systems where the decision actually lives.

Run is what it costs per month once live: inference, infrastructure, observability. Maintain is keeping it correct as your business, your data, and the models underneath all change.

A proposal that only prices build is not cheaper. It is incomplete.

What drives the build number

Four things, in descending order of impact. How many distinct decisions the agent makes. How bad a wrong answer is, which sets how much evaluation and guardrail work is required. Whether a human stays in the loop, which adds an interface and a queue. And whether the output has to be auditable.

Notice that none of these are about the model. Choosing between the top models changes your cost less than adding one approval step to the workflow.

Integration is where the budget actually goes

The agent has to read from and write to systems that were not designed with it in mind. If your data is in a modern warehouse with clean APIs, integration is weeks. If it lives in an ERP from 2009 with an undocumented interface and a batch window, integration is months, and the agent is the easy part.

The questions that move the number: how many systems, do they have real APIs, is there a staging environment, who owns the credentials, and how long does a change take to get approved. That last one is a cost line even though it never appears in a quote.

When a project overruns, this is nearly always the reason. Not the AI.

What it costs to run each month

Inference is priced per token, so your monthly bill is volume multiplied by prompt size. Long retrieval prompts and multi step agent loops both multiply that. A single conversation that takes twelve model calls costs twelve times a conversation that takes one.

Then infrastructure, which is usually modest, and observability, which is not optional. If you cannot see what the agent did on a specific request, you cannot debug it, and you will pay that back in engineering hours.

For most mid-market first use cases the monthly run cost is far smaller than people fear, often a few thousand dollars. The fear is misplaced; it should be pointed at maintenance.

The line nobody budgets: evaluation and drift

Your agent will get worse without anyone changing it. Your product changes, your policies change, your customers ask new things, and the model provider ships an update. Systems without continuous evaluation degrade silently, and the first signal is usually a complaint.

Budget for a reference set of real cases, automated runs against it, and someone owning the number. This is a modest recurring cost that protects the entire investment. Skipping it is how a project that looked successful in month two becomes an embarrassment in month six.

Pilot pricing versus production pricing

A pilot exists to answer one question: is this decision worth automating, and can it be done in your environment. It should be small, time boxed, and produce a real answer including the production price. If a pilot proposal cannot tell you what production would cost, it is not a pilot, it is a paid exploration.

Production pricing should be per use case, not per hour. Per hour rewards slowness. Per use case aligns the vendor with getting it live.

In-house versus a partner, with the real math

Hiring is not cheaper, it is deferred. A senior AI engineer takes months to hire, costs meaningfully more than the salary once you add employment costs, and does not ship anything during ramp. Two of them for a year is a large number, and you still carry the learning curve on guardrails and evaluation that a team who has shipped before already paid for.

The honest comparison is not cost per year. It is time to the first working thing in production, and who carries the risk if it does not work. Building the team is the right answer when AI is core to your product and you need the capability permanently. Bringing in a squad is the right answer when you need the first case live now and your team trained along the way.

How to sanity check a quote

Five questions. Does it price all four buckets or only the build? Does it name the systems it will integrate with? Does it define what a wrong answer looks like and how that will be measured? Does the code belong to you? And is there a specific date when something is live in front of a real user?

If a proposal cannot answer those five, price is not your problem yet. This is how we structure engagements, and the pricing tiers are public rather than discovered in a negotiation; you can talk to sales or read how we scope production AI, end to end.