AI agent development
We build AI agents that take over your team's groundwork. A person approves what matters. For SMEs and start-ups, from Tilburg.
Want an AI agent built? Then you're in the right place. Together we look at which work comes back every week and what an agent could take over. Sometimes a subscription to an existing tool turns out to be enough. We'll just tell you that.
Agent, chatbot or workflow tool?
What is an agent, really? Software that carries out a task, with a language model as a thinking step in between. It gets input and works out what needs to happen. To do that it uses your systems, and it delivers something. Sometimes on its own. More often with a person who gives the final click.
A chatbot is something else. It answers questions from a fixed set of texts. Handy, but it doesn't do anything in your systems. A workflow tool follows fixed rules. If this, then that. For work without doubt, that's often cheaper and more reliable than an agent.
An agent only adds value where something has to be read or written. Understanding an email, for example. Or turning a messy enquiry into something structured.
Which work an agent can take over
Which work qualifies? Rule of thumb: if messy text comes in and something structured has to come out, it's a candidate. In AI automation for SMEs we go through six processes.
Lead qualification is one. The agent reads an enquiry and scores it against your criteria. Then it prepares a draft reply. Quotes and intake lend themselves well to it too. So does follow-up, where the agent tracks what's still open and lines up a follow-up email in time.
What should you not hand over? Decisions with big consequences for people stay with people, and so does anything where money leaves the business directly. A process that isn't clear in itself is no good candidate either. Does your team disagree on what a good lead is? Then a model won't solve that for you. It just makes the mess faster.
How we build an agent
How do you keep an agent under control? With a person who approves what matters. Is something going to a customer, or is money leaving the business? Then someone on your team sees the draft first, with the reasoning next to it. They approve it or change it. Only then does anything happen.
We treat the model's instructions as code. They live in version control, next to the rest of the software. You can go back to an earlier version and see who changed what. A change gets reviewed before it goes live. Why that matters so much is in prompts are code too.
We log every decision the agent makes, with the input and the reasoning. And with a fixed set of test examples we check that a change doesn't make things worse somewhere else.
What drives the price of an agent
What will an agent like that cost? Mostly it depends on what surrounds it. Calling the model is a few lines of code. The work sits in the connections to your systems and in your data. Checks and monitoring count too. An old system without an API can easily become the most expensive part. We set out honest ranges per level in what it costs to build an AI agent.
Sometimes a subscription is smarter. Does your task look like what thousands of other businesses do, and does an existing tool connect to your systems? Then just take that tool. Building pays off when the agent has to follow your own process.
Is an AI agent the same as a chatbot?
No, an agent does more. A chatbot answers questions from a fixed set of texts. An agent carries out a task. It reads the input, uses your systems and delivers something, like a draft quote or a filled-in line in your CRM. Often a person approves the result before anything happens.
Can an agent you build send emails on its own?
Only if you want that after a while. We start with a person who approves everything. Once you see a simple step running reliably, that step can go automatically. Anything that goes to customers or involves money, we keep under approval. That's cheaper than fixing mistakes afterwards.
Which processes lend themselves to an AI agent?
Processes where messy text comes in and something structured has to come out. Lead qualification and intake are often good first candidates, because the outcome is easy to check. Quotes and follow-up work too. Does your process follow fixed rules without doubt? Then ordinary automation is usually cheaper and more reliable.
Why do prompts belong in version control?
Because a prompt running in production is part of your product. In version control you can go back to an earlier version and see who changed what. You can also review a change before it goes live. Don't know which version is running? Then you also don't know why yesterday's output was better.
How soon is a first version of an agent ready?
Usually within 2 to 4 weeks. That shows you quickly whether the idea holds up. An agent that really runs in your daily work, with connections and logging, takes longer. How much longer depends on how many systems are involved. We map that out together with you up front.
Which work would you like to hand over?
Tell us which task comes back every week. Then we look together at whether an agent fits. Would a simple tool do the job? Then you'll hear that from us.