What does it cost to build an AI agent?
2026 prices, from a studio that publishes its own.
At innops, building one AI agent for one workflow typically costs €5,000–15,000, fixed before work starts, and takes 4–12 weeks. That price covers the engineering, a test set built from your real cases, monitoring, a runbook, training for the person who will own it, and 90 days of follow-up. Running costs come on top: model usage, which scales with volume and is estimated in the scope, and hosting, which you do yourself or hand to us through Operate from €60 a month. Most projects start with an AI audit from €1,000, because the audit is where the scope, and so the price, gets fixed.
WHAT MOVES THE PRICE
- 01
Integrations
Every system the agent reads from or writes to — a CRM, a mailbox, WhatsApp, a booking or ERP system — needs its own connection, error handling and tests. One integration is quick; four is a different project. - 02
Human approval
When a person must approve output before a customer sees it, you need a review step and a way to hand cases back. Usually worth it, and it adds work. - 03
Data and test set
We test on real cases, not invented ones. If your examples are scattered across inboxes and PDFs, collecting a representative set is part of the build. - 04
Channels and languages
A web chat alone is simpler than web, WhatsApp and email together. Each extra language needs its own test cases. - 05
Risk
An agent that drafts costs less than one that commits. Prices, bookings, refunds or anything touching personal data need stricter guardrails and logging. - 06
Ownership after launch
A system your team runs needs documentation and training; a system we run needs monitoring and an on-call window. Both are priced in up front, not added later.
THREE EXAMPLE SCOPES
Illustrations, not quotes. The audit fixes the real scope and the fixed price for your workflow.
A website assistant that answers and hands over
Answers questions from your own content in one or two languages, and passes serious leads to your team by email or form. One channel, one integration, no commitments made on your behalf.
Similar in shape: the LG Builders assistantLead intake across several channels
Picks up enquiries from the website, WhatsApp and email, asks qualifying questions, scores them and routes each one to the right person, with every step logged.
The pattern behind the Toubkal Wanderers caseA document agent that updates your systems
Reads incoming orders, claims or quotes, pulls out the data, updates two systems and drafts a reply that a person approves. More integrations, more testing, more risk to manage.
WHAT'S NOT IN THE BUILD PRICE
- Your team's time: interviews at the start and a short test on real work every week.
- Licences for tools you already use, such as your CRM or mailbox.
- Model usage, billed by the provider on your account and estimated in the scope.
- Hosting and maintenance after the 90 days, unless you choose Operate.
- New features after handoff: that's a new, separately priced build.
HOW TO COMPARE QUOTES
- What exactly is live, tested and documented on the day you leave?
- Is the price fixed per deliverable, or an open-ended day rate?
- Whose name are the accounts and the code in?
- How is quality measured, and on which test set?
- What happens after launch, and can we leave without an exit fee?
More on choosing a partner: AI consultancy for SMEs. Subsidies and the steps before a build: implementing AI in an SME. Vision agents in the lab: pallet inspection with a vision LLM.
What buyers ask about AI agent prices.
Why do quotes for an AI agent differ so much?
Because "AI agent" covers very different things. A chatbot subscription you configure yourself, a no-code flow a freelancer wires up in a week, and a system that reads your documents, updates your CRM and is tested on hundreds of real cases are all sold under the same name. The price follows four things: how many systems the agent connects to, how much of its output a person must approve, how much testing it gets before customers see it, and who owns and maintains it afterwards. Cheap quotes usually leave out testing, monitoring and handoff, which is where most pilots die. Expensive quotes often include project management and long discovery phases. Ask every supplier the same question: what exactly is live, tested and documented on the day you leave?
Can I build an AI agent myself with no-code tools?
Often, yes, for a first version. Tools like n8n, Make or a chatbot builder let a handy person connect a mailbox, a model and a spreadsheet in an afternoon, and that's a good way to learn what you actually need. The trouble starts when the agent becomes something the business depends on. Who notices when it gives a wrong answer? What happens when the model provider changes a version, or a colleague edits the spreadsheet layout? Where are the logs when a customer complains? A Build adds the parts a weekend project skips: a test set from your real cases, monitoring, a runbook and a named owner. If you want to try it yourself first, the hands-on workshop is designed for exactly that, at €525 per track.
Is there a subsidy for building an AI agent in Belgium?
It depends on where your company is based. In Flanders, mostly not: since 1 February 2026 the VLAIO KMO-portefeuille only subsidises cybersecurity advice, IT development was never eligible, and VLAIO has announced the scheme stops on 1 January 2027. In Brussels, SMEs with an operating seat in the region can apply for the digitalisation premium: 25% to 70% of a consultancy mission that digitalises internal processes, up to €10,000. You apply via MijnBEW at the latest the day before the mission starts, and the consultant must meet conditions too, including at least two years of consultancy activity; innops started in 2025, so it doesn't meet that one yet. Brussels' general consultancy and training premiums are suspended since 12 August 2026. Small companies can sometimes use the investment deduction for software. Our guide to implementing AI in an SME lists the sources, checked in October 2026.