What AI agent development costs

This page sets out what the firms in this guide publish about price, quoted exactly as printed, and what a buyer should ask for instead when a firm prints nothing. No amount on this page has been converted into another currency, and no amount has been estimated. Every page used was read on 21 October 2026.

An AI agent is a program that uses an AI model to finish a task in several steps by itself, such as reading an invoice, checking it and sending it on for approval. Its cost has two parts. The first is the work to build it: the people who connect it to your systems, write its rules and test it. The second is the cost to run it: the fees for the AI model each time the agent works, the servers it runs on, and the staff who review its work. A quote that covers only the first part leaves out a cost that continues every month.

What the firms publish

Firm	What its pages say about price	Amount printed
HAVIT	"1 day · on-site · your office From €3 900 · excl. VAT"	Yes, a starting price
HAVIT	"1 day · on-site · Prague or at your office (CZ/SK) From €4 900 · excl. VAT"	Yes, a starting price
Mactores	"Fixed date. Fixed fee. Customer-signed acceptance."	No
Mactores	"Mactores absorbs overage on delays inside our control."	No
Cay Digital	"Pricing is per cohort of up to 12 people" (training)	No
The other fourteen	No price found on the pages read	No

HAVIT's two workshops

HAVIT is a software firm in Prague, confirmed by the Czech ARES register as HAVIT, s.r.o. It prints two prices, both starting prices, shown by the word "From" [1] [2]. "excl. VAT" means the price excludes VAT, the value added tax, which is the sales tax charged on most purchases in the European Union. The tax is added to the printed amount.

The first is the AI Developer Onboarding Workshop. Its page reads "1 day · on-site · your office From €3 900 · excl. VAT" [1]. The page says the workshop is for "4 to 8 developers from one team" and an engineering lead, and that the session is hands-on, in the team's own code. It is a one-day version of a longer offer on the same page.

The second is the AI Adoption Workshop. Its page reads "1 day · on-site · Prague or at your office (CZ/SK) From €4 900 · excl. VAT" [2]. CZ/SK means the Czech Republic or Slovakia. The page says the session is for a company's leadership, product and engineering staff together, and that they leave with a ranked list of possible uses for AI and a small first project planned in detail.

Both prices buy one day of work, which makes them a fixed first step. Ask HAVIT, as you would ask any firm, for a separate written quote for building the agent itself.

Mactores' fixed fee

Mactores' commitment page has the heading "Fixed date. Fixed fee. Customer-signed acceptance." [3] Customer-signed acceptance means the client signs to confirm the work is done. Its homepage adds, "Mactores absorbs overage on delays inside our control." [4] In plain words, Mactores says it pays the extra cost when a delay is its own fault. No amount appears on either page. Ask which delays Mactores counts as inside its control, and write that list into the contract. Mactores also says its own platform, Aedeon, "builds the agent infrastructure" [14], so ask whether your agent depends on Aedeon after the project ends and what that costs.

How Cay Digital prices its training

Cay Digital's training page says, "Pricing is per cohort of up to 12 people" [5]. A cohort is a group of staff trained together. The page gives no amount and says, "We talk through the number on a quick call." That is a price for training staff to use AI. For agents, its agents page says client agents run on Agent Mission Control, its own product, where staff can see "what it costs and what it returns" for each agent [6]. Ask what a licence for that product costs each year.

A form that asks for a budget

Galadrim prints no price. Its contact form asks for a budget and offers bands from "0€ - 10K€" to "300K€+", with an "I don't know" choice [10]. These are answer choices on an enquiry form, and the page links no price or piece of work to any of them.

What changes the cost of an agent

Anthropic, which makes the Claude models, writes in its guide "Building effective agents" that "Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense." [7] Anthropic uses "agentic systems" for both agents and fixed workflows, and latency means waiting time. In plain words, a system built this way often takes longer and costs more per task than a simpler program, in return for better results. That is the model maker's own advice. If a simpler program does the job, it may cost less to run.

Four things drive the cost of building and running an agent.

1. The systems it connects to. Each system the agent reads or writes, such as an order system, an accounting system or a customer records system, needs its own connection and its own tests.
2. The tests. An agent chooses its own steps, so it has to be tested on many cases and many runs. How to test an AI agent before you buy explains why.
3. Model fees. AI model makers charge for the tokens a model reads and writes. A token is a small piece of text, often part of a word. Yellow Tech's AI Factory page says its agents process "5B+ Tokens" every month, its own figure [8]. Ask each firm how many tokens your agent will use per task and who pays the model maker.
4. People. Someone reviews the cases the agent passes on, and someone watches the agent after launch.

OpenAI's own guide to building agents advises, "Set limits on agent retries or actions." [11] A limit on retries also caps what one task can cost. Ask each firm what limits it sets.

How long the work takes

Some firms publish how long their work takes. Mactores' homepage gives "12 weeks Median time to production", a figure for all its work with no method shown [4]. A median is the middle value. First Line Software offers to embed with a client's engineers for 8 to 12 weeks [12]. Galadrim gives "4 operational AI agents in 6 months" for one client it names [13]. Each figure is the firm's own. Ask for the length of your own project in writing, with the date the agent is expected to start work on real requests.

How buyers check costs

KPMG's AI Quarterly Pulse Survey for the third quarter of 2026 says "74% incorporate formal cost reviews into their AI approval processes" [9]. Elsewhere on the same page it gives the figure as "74% of leaders require cost reviews during project approval, up from 61% last quarter." Both describe the leaders KPMG surveyed, and the page gives no sample size, so the figure says nothing certain about companies in general. A formal cost review means the cost is checked before a project is approved. A written quote that splits the build cost from the monthly running cost makes that review possible.

Reveneau and Fluxon

Reveneau says it uses AI in place of hiring more engineers and puts that saving into the prices its clients pay. By its own account, AI writes 100 percent of its code, and every change must pass a large eval suite, written from the specification, before it is released. Reveneau says Jev grades that suite and makes it run ten times faster than its previous grader, and that it takes responsibility for the whole project through production, the live system people use, and after release. An eval suite is a set of automated tests. You can ask Reveneau for a quote. Fluxon prints no price on the pages read for this guide.

What to ask for in a written quote

Ask every firm for the same items, so you can compare the answers:

• the price of the build, split into its steps, with what each step delivers;
• the monthly cost to run the agent, with the model fees shown on their own line;
• the fee for any product or platform the firm owns that the agent runs on, and what happens if you stop paying it;
• the limits the agent works within, such as the number of retries or actions per task;
• how many test cases and test runs are in the price, and who keeps the tests;
• the currency and whether tax is included.

The AI agent project checklist turns these questions into a form to print and send.