Solutions · AI agents
AI prepares, you decide
AI agents for business are software that carries out one defined task in your systems: reading an email, looking up the customer, preparing a draft. I build agents with fixed limits, a log for every action and a person involved in anything that touches a customer, an amount or a contract.
Free · 30 minutes online · you always get a short advice report
Sound familiar?
Where AI can already save time
An inbox full of similar questions
Looking up the same information and typing the same answer again and again, with small differences.
Documents in varying formats
Quotes, invoices and forms that are each laid out differently, so someone retypes them.
Trying AI, but not daring to let go
ChatGPT does help, but you don’t want it to send a customer something nobody has read.
How it runs
How an agent works with a person at the wheel
An agent chooses its route case by case. That is why I only use one where there is real variation, and the moment of approval is fixed.
- Email or document inA question, a quote, a form.
- Gather contextCustomer in the CRM, earlier agreements, your knowledge base.
- Draft or proposalAn answer, filled-in fields or a summary.
- Approveyou decideIs it right? Edit or send.
- Carry out and recordSend or save, with source and version in the log.
- LearnEvery correction becomes a new test case.
Beforehand I set down what the agent may do by itself, what always goes to a person and what is never allowed. An emergency stop blocks new actions without erasing the log of earlier cases.
What I build
AI agents for business that take over one defined task
A clearly defined AI agent costs €5,000 to €12,000 to build. On top of that there are usage costs for the model; I show those separately, with a budget and an alert. See the pricing page.
Support agent on your knowledge base
Answers questions from your own documents (RAG) in your tone, or passes them on to a person.
About RAG →Email triage
Every incoming email labelled, prioritised and, where possible, given a draft reply.
Reading documents
Data from quotes, invoices and forms in varying formats, to the right place in your systems.
Preparation before a meeting
What is known about the customer, on one page, with the source for each part.
Choosing
Agent, fixed flow or a person?
Use AI only for the part where the input really varies, such as free text or documents. If every case follows the same rules, a fixed workflow is cheaper, easier to explain and easier to test.
| Situation | What fits | Why |
|---|---|---|
| Fixed input and fixed rules | Workflow or API integration | Predictable and easy to test |
| Varying content, outcome can be checked | AI with human review | AI reads, a person approves |
| High impact or unclear standard | A person decides | No model carries that responsibility |
Systems
Models and systems I connect
In the cloud or on your own server. For each model I record which data leaves the environment and where it is processed.
- OpenAI
Anthropic (Claude)
Mistral
Ollama
Qwen
Qdrant
Supabase / pgvector
MCP-servers
Gmail- Microsoft 365 (Outlook)
Zendesk- Freshdesk
- Zammad
WhatsApp Business
Proof
AI with brakes, in my own systems
Own build · running in simulation
BH28: an autonomous system with a risk gate
32 n8n workflows around a Python core. ChatGPT and Claude help with review and analysis, but an order only passes through the fixed checks.
Own product · final test phase
Holos33: app, admin and CMS on one own API
278 API endpoints, an AI gateway to an open model on its own server, and privacy you can see in the screens.
I want to understand how something fits together before I say yes. With Stefan, that conversation is possible. He knows what he is talking about, explains why he builds something the way he does and looks at how we work.
And then he comes up with just the piece I had not thought of myself. When I send a question, I get an answer quickly.
Mike van EkelenburgDirector, Greven · translated from DutchWhen I advise against an agent
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, due to high costs, unclear value or insufficient risk control (forecast, June 2025). So I start small: one task, one owner, one test set. I advise against an agent when:
- there is no fixed set of test cases to measure quality against;
- it is meant to become a broad "digital employee" without a defined task;
- a mistake directly affects a customer, amount or contract and there is no person in between;
- a fixed workflow can do the same, cheaper and more predictably.
Frequently asked questions
About AI agents
What is the difference between Copilot and an AI agent?
Copilot works alongside you in your own programs and helps while you work. An AI agent completes a task in your systems, even when you are not there. Copilot makes you faster; an agent takes over a defined piece of work.
What does an AI agent cost?
Building a clearly defined agent costs €5,000 to €12,000. On top of that you pay for model usage, search and storage. I calculate with your volume, agree a budget and show build costs and usage costs separately.
Does my data go to an AI provider?
I set that down per model. With a cloud model, only the data needed for the task is sent. For sensitive data, an open model can run on a dedicated server in the EU, so business data stays in-house.
Do I need to understand AI myself?
No. You know how your work should go and when an answer is right. I build, test and maintain the agent. You approve the outcomes and stay in charge.
Next
Read on
Local AI on your own server
Qwen, Llama, Gemma and more, without data leaving the building.
MCP server and AI agents
One protocol between AI and your systems, with permissions per action.
Connecting ChatGPT and Claude
Language models in your processes, with a gateway in between.