AI agent price: the build cost, the running cost, and what quotes leave out

The short version

• An AI agent has two costs, a one-off build cost and a recurring running cost, and a quote that prices only the first is incomplete.
• Off-the-shelf AI agents are usually priced per user, per conversation, per action or per resolved outcome.
• The build cost of a custom AI agent is driven by the systems it connects to and the testing it needs, not by the model.
• An AI agent's running cost scales with volume, because one task can involve several model calls and each call is billed.
• Spending benchmarks built from payments to paid AI tools exclude custom builds, so they cannot tell you what a custom agent will cost.

What are you paying for when you buy an AI agent?

You pay for two separate things: building the agent and running it. The build is a one-off cost for the work that connects a model to your systems, sets its rules and proves it behaves. The running cost arrives every month the agent is in use: model usage, hosting, monitoring and the people who keep it working.

Most price comparisons go wrong by setting one vendor's build quote against another vendor's subscription. A subscription folds build and run into one recurring number. A custom quote separates them, and the second number is often missing from the proposal. Before comparing anything, put both options on the same footing: what you pay in month one, and what you pay in month twelve.

The running cost is also the one that grows. A build price is fixed by scope. A running cost is set by volume, and volume is the thing you are hoping will go up.

How are AI agents priced?

Most AI agents are priced in one of five ways: per user, per conversation, per action, per resolved outcome, or as a custom build with a project fee followed by running costs. Each model moves risk between you and the vendor in a different direction, and that matters more than the headline rate.

Per-user pricing is predictable but charges for seats rather than work. Per-conversation and per-action pricing follow usage, which is fairer, but you cannot forecast anything until you know what the vendor counts as one conversation or one action. Per-outcome pricing sounds safest, and it is only as good as the definition of a resolved outcome. A custom build costs more up front and usually leaves you paying the model provider directly, which can be cheaper at volume and is more work to manage.

Vendor pricing for agents has not settled. Salesforce announced new, flexible pricing for Agentforce in May 2025, and the same release describes the list pricing it already had. The unit you are billed in today may not be the unit you are billed in next year, so be wary of long contracts priced on it.

Which AI agent pricing model suits which situation?

What drives the cost of building a custom AI agent?

The build cost of a custom agent is driven by the systems it has to touch and the testing it needs before it can be trusted. The model is the easiest part to change and rarely the largest part of the bill. An agent that answers questions from one document set is a small build. An agent that reads a CRM, writes to a ticketing system and issues refunds is a different project, even if the two demos look alike.

Each integration brings authentication, error handling, rate limits and a set of failure cases somebody has to think through. Each action the agent can take brings a permission decision: should it do this alone, or ask a person first? Reading data is cheap. Writing to a system of record is where the cost and the care go.

Testing is the line most often under-quoted. A dependable agent needs a set of real examples with known right answers, run every time the prompt, the model or the tools change. Building that set means someone from your side sitting with the engineers and saying what a correct answer looks like. That time is part of the cost even when it is not on the invoice.

Then there is everything the demo never showed: a connected system that is down, a request that is ambiguous, a user asking for something the agent must refuse. Handling those cases is most of the distance between a prototype and something you can leave running.

What does an AI agent cost to run each month?

An AI agent's running cost is model usage multiplied by volume, plus hosting, monitoring and maintenance. Model usage is billed per token at published rates. Anthropic's API pricing, read in November 2026, lists one model's output at $10 per million tokens.

At that rate, a 1,000-token reply costs one US cent in output. That figure misleads on its own, for three reasons. An agent rarely finishes in one call: it plans, calls a tool, reads the result and calls the model again, and every step is billed. Each step also re-sends the conversation so far as input, which is charged separately and grows as the task goes on. And failed attempts, retries and tool errors are billed exactly like successful ones.

So the honest unit is cost per completed task, not cost per token. Measure it on real traffic during a pilot, then multiply by the volume you expect. That number, not the per-token rate, belongs in the business case.

The rest of the running bill is less visible. Hosting and logs are usually modest. Maintenance is not. Connected systems change their APIs, your policies change, and model versions are retired. Somebody has to rerun the tests and fix what moved, so budget for upkeep as a standing cost rather than an occasional surprise.

Why don't published AI spending figures tell you what an agent will cost?

Most published AI spending figures measure payments for paid AI tools, not agents built for one business. They answer 'what do firms spend on AI software', which is a different question from 'what will this agent cost us'.

The JPMorganChase Institute's study of small-business AI use is a careful example. It uses de-identified Chase Business Banking transaction data from 4.6 million small businesses, covering 2019 to 2025, and counts payments to more than 500 paid AI services. By design it excludes free tools and custom builds, and its median describes firms that pay for AI, not all small businesses.

That is a sound method for what it measures. It also means its figures say nothing about a custom agent, because custom builds sit outside its scope. If an average spending figure is used to anchor a price, check what was counted before letting it shape your judgement of the quote.

When is an AI agent not worth the price?

An AI agent is not worth the price when volume is low, the process keeps changing, or a simpler tool would do the job. An agent earns its running cost by handling enough work that people stop doing it by hand. If the task comes up a few times a week, a person with a good checklist is cheaper and easier to correct.

If the process changes every month, the agent's tests and rules change with it, and you pay for the rework each time. Settle the process first. If the steps are fixed and the inputs are structured, ordinary workflow automation without a language model costs less to run, is easier to test and does not guess. Agents justify their cost where inputs are messy: free text, varied documents, requests that need judgement to route.

And if nobody in the business will own the agent after launch, do not build it yet. An agent without an owner degrades quietly as the systems around it change, which costs more than never having built it.

What should you ask before accepting an AI agent quote?

• •Which costs are one-off, and which recur every month?
• •Who pays for model usage, and is it passed through at cost or marked up?
• •What is the expected cost per completed task, and how was it measured?
• •What happens to the price at ten times today's volume?
• •What does the vendor count as a conversation, an action or a resolved outcome?
• •Who owns the prompts, the test set and the integration code if you leave?
• •What does maintenance cover when a connected system or model version changes?