How much does an AI agent cost? Build, running and hidden costs

Part 01

How much does an AI agent cost? What the public data says

Few published price ranges have a method behind them. The most useful one comes from Clutch, which collects project budgets for AI development from verified client reviews. In its update of 21 September 2026, the most common budget is $10,000 to $49,999 and the average is $120,594.55. The average sits far above the common range because a few large projects pull it up.

Read those figures as an order of magnitude. They cover any AI project, not only agents. They are in US dollars, from providers worldwide, and they do not separate the cost of building from the cost of running. For your case, the real number comes from the five drivers in part 02.

Cost holds many companies back. According to Eurostat, among EU companies with 10 or more employees that once considered using AI and do not use it, 38.4% say the costs seem too high. In Spain the figure rises to 47.4%, in Germany it falls to 32.4%. Cost is not the main reason. A lack of relevant expertise comes up in 70.3% of cases.

Gartner points the same way. In June 2025 it predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.

Part 02

The five drivers that set the price

Two agents that look the same on a slide can cost very different amounts. An agent that sorts the invoice inbox of a distributor in Lyon and one that reviews lease contracts for a property manager in Rotterdam use the same kind of model. Everything around the model changes.

1. Integrations

Every system the agent reads or writes adds work. An ERP with a modern API connects quickly. An old accounting package that only exports files, or a web portal without an API that has to be read screen by screen, multiplies the effort and the upkeep.

2. Data quality

If the data is clean and in one place, the agent works on it from day one. If there are scanned PDFs, customers duplicated between a spreadsheet and the ERP, or fields everyone fills in their own way, someone has to sort them out first. Eurostat lists data availability or quality as a reason not to use AI for 43.5% of the companies that considered it.

3. Volume

Volume barely moves the build cost, but it moves the monthly cost. Each case means several model calls, and the model is paid per use. Two hundred emails a month and twenty thousand do not cost the same to process.

4. Human review

The more an error costs, the more work goes into the evaluation set, the confidence thresholds and the screen where a person approves. A wrong label is fixed in seconds. A wrong payment or a missed clause is not.

5. Hosting

A managed service is the simplest option. If the data has to stay in your cloud, in the EU or on your own servers, there is more infrastructure to set up and watch.

Part 03

Build cost and running cost: two separate bills

The most common mistake when budgeting an agent is to look only at the build. You pay for the build once. You pay for running it every month the agent works: model usage, hosting, maintenance and the time of the people who review uncertain cases.

Model cost, calculated with official prices

Of the monthly items, model usage is the one you can calculate yourself. Providers publish a price per million tokens (fragments of words), one for what the model reads and one for what it writes. In the models in figure 3, output costs five times as much as input. Here is an illustrative example with Anthropic and OpenAI prices from October 2026:

1. Volume.An accounts inbox receives 2,000 emails a month.
2. Calls.The agent makes 4 model calls per email: classify it, look up the ERP, draft the reply and check it.
3. Tokens.Each call reads about 3,000 tokens (instructions, the email and context) and writes about 400. Per email: 12,000 input and 1,600 output.
4. Price.With a mid-tier model at $2 per million input tokens and $10 per million output tokens: $0.024 plus $0.016, so $0.04 per email. For 2,000 emails, $80 a month.

The same work costs between $4 and $400 a month depending on the model. That makes the choice of model a cost decision, and the most capable model does not always win. Sorting emails usually needs only a small model. Drafting a sensitive reply may justify a larger one. A well-designed agent uses each model where it is needed.

There are two more levers. Prompt caching makes the instructions that repeat on every call cheaper: on Claude Sonnet 5.5, a cache read costs $0.10 per million tokens against $2 for normal input. And batch processing, for work that is not urgent, saves 50% according to Anthropic.

Part 04

Build or buy

You do not always need to build. If the process is the same in most companies, a product probably does it already. These prices come from the vendors' own pricing pages:

Buying brings the product, the hosting and the updates with it. What changes is how the bill grows. If a bought agent resolved 2,000 conversations a month at $0.99 each, you would pay $1,980 a month, and the figure rises with every conversation. A custom agent at the same volume spends a fraction of that on the model, but it has a build and maintenance to pay for. The sum depends on your volume and on how close your process is to the product's standard.

Part 05

The costs that are missing from the quote

These items are often missing from the first number you receive. None of them is large on its own. Together they decide whether the agent pays off.

• Preparing the data. Removing duplicates, granting read access and sorting documents before the agent touches them.
• The evaluation set. A collection of real cases with the right answer, run again after every change. Without it, nobody knows whether the agent is getting better or worse.
• Review time. Every case the agent passes to a person uses minutes of someone on your team. With badly set thresholds, the agent saves little.
• Model changes. Providers retire models and change prices (OpenAI and Anthropic publish their retirement schedules). Each change means running the evaluation set again.
• Integrations that change. Your ERP or CRM updates its API and the connection needs adjusting.
• Compliance. Since 2 August 2026, Article 50 of the EU AI Act requires that people talking to an AI system know it, unless that is obvious.
• The exit. If the code, prompts and evaluation set are not yours, switching provider means starting from zero.

Part 06

How Alpgency prices an AI agent

We do not publish rates because the price comes from your process. The method is fixed:

1. Free Audit.It starts with a 30-minute call. We map how the work is done today, rank the use cases by value, compare build and buy, and put a fixed price on the first one. The roadmap is yours to keep, whether you go ahead or not.
2. Demo and build.Before the build, you see a working demo on your real data. Then comes the build to production. Both are covered by one fixed price, agreed after the Audit.
3. Run: after go-live, you choose.A full handover to your team, a fixed monthly support fee, or us as your tech partner on a fixed schedule.

We never bill by the hour, by open-ended estimates or by hours saved. You see model usage measured from day one. That is how we worked on the deal-flow scoring system for a European pre-seed fund, live in production 9 days after kickoff (2026).

Ask for two numbers: what it costs to build, and what it costs every month to keep running.

Alpgency

FAQ

Frequently asked questions