How much does it cost to build an AI agent? For a custom agent that does real work inside your business, expect €25,000 to €80,000 in most cases. A narrow agent inside one system can come in around €10,000 to €25,000. An agent that runs inside your own infrastructure, or several agents working together, starts around €100,000. Below roughly €10,000 you are usually not building an agent at all. You are configuring a tool someone else built, and that is often the right call.
Those ranges are wide because the word agent covers very different things. A bot that answers questions from twenty PDFs and an agent that reads invoices, checks them against your ERP and books them for approval share a name and almost nothing else. The price follows the responsibility you hand over, not the number of prompts.
AI agent cost in 2026, by bracket
These are the brackets we quote, and the same ones we publish on our pricing page. The off-the-shelf row is not something we sell. It is there because it is the honest comparison.
| Bracket | What you get | Systems touched | Quality control |
|---|---|---|---|
| €3,000 to €10,000 | Configured tool or no-code bot. FAQ assistant, lead form helper, simple document Q&A | One, often none | Manual spot checks |
| €10,000 to €25,000 | One focused agent inside software you already run | One system with a decent API | Small eval set, under 100 cases |
| €25,000 to €50,000 | Multi-step agent that uses tools and structured outputs | Two or three systems | Eval in CI, prompt versioning |
| €50,000 to €100,000 | Production agent that owns part of a real workflow | Three or more, including legacy | Eval, observability, audit log, human approval |
| €100,000 and up | Agent on your own infrastructure, or several agents coordinating | Inside your VPC or on premise | Custom eval harness and regression suite |
If you came here searching for custom AI agent development cost and the answer you need is a single number, the honest one is this: most of the agents we scope end up between €25,000 and €80,000.
The five inputs that decide the price
- 01Scope. One task with a clear right answer, or a chain of decisions where the agent chooses the next step.
- 02Integration surface. Every system the agent reads from or writes to adds authentication, permissions, error handling and testing. Write access costs more than read access, every time.
- 03Quality bar. How wrong is the agent allowed to be? An internal search helper can be wrong occasionally. An agent that changes customer records cannot, and proving that takes an eval suite.
- 04Where it runs. A hosted model through an API is the cheapest to start. EU region endpoints cost a little more. Open weights on your own hardware cost the most up front and can be cheapest at volume.
- 05Data sensitivity. Personal data, health data or financial records bring GDPR work, logging requirements and sometimes EU AI Act obligations, depending on the use case.
Anyone who quotes you a firm price in the first meeting is guessing at least three of these. That is why we run a fixed fee discovery week before we commit to a number.
Where the money goes inside a build
A typical €40,000 agent project does not spend most of its budget on prompts. Roughly, in the projects we run:
- Integration with your systems takes the largest share, often a third or more.
- Evaluation, meaning the test set and the harness that runs it on every change, is usually 20 to 30 percent. It is the line most cheap quotes leave out.
- The agent logic itself, prompts, tools and orchestration, is smaller than people expect.
- Hardening takes the rest: retries, fallbacks when a model provider has a bad day, cost caps, logging, a runbook for whoever is on call.
How a build is phased
Discovery week comes first, with a fixed fee and a written scope you keep either way. A common outcome is that the agent should not be built yet because the data is not ready. Then a prototype sprint of two to three weeks, with a working slice in your environment and the eval suite live by week two. Then two to six weeks of hardening before real users touch it.
Running costs people forget
The build is not the whole bill. Plan for 15 to 25 percent of the build cost per year to keep an agent healthy, plus usage.
- Model usage. Per token pricing from the provider, which depends heavily on volume and on which model handles which step.
- Retrieval and storage, if the agent searches your documents.
- Hosting and monitoring.
- Prompt and eval updates when your data, your process or the model changes. Providers retire model versions, and each change needs the eval suite run again.
- New integrations, because a working agent attracts requests.
Model costs are usually the smallest of these once the agent is designed sensibly. Routing simple steps to a cheaper model and caching repeated context can cut them by half or more.
Why cheap agents get expensive
A €6,000 agent can demo beautifully. The problems show up in month two. It answers confidently when it should say it does not know. It has no access control, so it shows one customer's data to another. Nobody can tell why it behaved differently on Tuesday. It was built around a prompt rather than around the workflow, so every edge case needs a manual fix.
When you should not build one
- The task is generic. Support macros, meeting notes and simple FAQ bots are well served by products that cost a few hundred euros a month.
- A rule or a query already gets the right answer. A model would add cost, latency and doubt.
- Nobody can define what a correct output looks like. Fix that first, or the project has no finish line.
- The whole budget is under €10,000. Spend it on a good tool and on cleaning the data you will need later.
If you are not sure which side of that line you are on, that is a reasonable place to stop and assess before spending. Our two week AI readiness assessment exists for exactly this question.
What an agent is worth
The cost question is only half of it. An agent that takes 20 hours a week of manual work off a team pays back a €40,000 build in under two years at typical EU labor costs, before counting faster response times or fewer errors. An agent that saves nobody any time is expensive at any price. If you cannot name the workflow it improves and how you will measure it, the project is not ready for a quote yet.