AI agents

AI agents that act inside your software, within limits you set

An agent reads an enquiry, looks up the customer, checks the calendar and prepares the reply. It decides which of these steps a case needs. What it may change, and what waits for your team, is defined before it runs.

See how we build them →

30 minutes, no obligation, German or English. Bring one task where your team switches between several programs. Clicking loads Cal.com (USA); the call runs on Google Meet. Prefer to write first? Ask by email. We send the draft data processing agreement and our terms on request.

Limit
We define the task, the tools the agent may use and the actions that always need approval.How it works · step 1/3
Build
The agent runs in a workflow tool such as n8n, connected to your systems with narrow permissions.How it works · step 2/3
Check
We test it with real cases, including wrong inputs, timeouts and attempts to make it do too much.How it works · step 3/3

In short

An AI agent is a workflow that chooses between defined steps and tools, not an autonomous employee. We build agents inside your existing software, with clear limits on what they may read and do, and an approval step before anything leaves the company. Agents connect to your systems through interfaces or MCP servers, with the narrowest permissions that work. The written proposal and terms govern scope, credit and any refund.

What an agent is

An agent is a workflow with choices, not a digital colleague

Most useful agents in an SME do one job: they receive an input, choose which tools to use, prepare a result and hand it to a person. The value comes from the limits, not from autonomy.

One job, clearly bounded

Sorting enquiries, preparing quotes, capturing documents, answering internal questions. An agent that is supposed to do everything is hard to test and hard to trust.

  • Tools instead of free access

    The agent gets named tools: search the CRM, read the calendar, draft an email. Each tool has its own permission, and writing tools are separate from reading tools.

  • Approval before it matters

    Messages to customers and changes to customer data wait for a person by default. The agent prepares, your team decides.

  • MCP as the connector

    The Model Context Protocol is an open standard for connecting AI to tools and data. Since December 2025 it is developed under the Linux Foundation. We use it where your systems support it and the permissions can be limited.

  • When an agent is the wrong tool

    If every case follows the same steps, a fixed workflow is cheaper and easier to check than an agent. We tell you which one your task needs.

MCP servers for companies: what they are and what they may do →

Security

An agent can only do what you allow

An agent that reads emails and documents can be manipulated by their content. The protection lies in what it is allowed to do, not in what it is told.

Least privilege
Read access first, write access only where needed, and never more than the task requires.
Approval for write actions
Sending, deleting, paying and changing records go through an approval step unless you decide otherwise in writing.
Logged decisions
We set up and test which tool calls, inputs and results are logged, with access and retention rules, so a wrong result can be traced.
AI Act transparency
Where an agent talks to customers, it says that it is an AI. We document what it does for your records.
How draft plus sign-off works: our approach in detail →

How it starts

From one task to a running agent

  1. Intro call

    Free

    You describe the task and the programs involved. We tell you whether an agent or a fixed workflow fits.

  2. Trial run

    from €2,500 · credited against implementation; refunded if we advise against automation

    As an audit or a prototype: we assess the agent for your real cases and, in a prototype, build and test it, including the cases where it must stop. The written proposal and terms govern scope, credit and any refund.

  3. Implementation and care

    One workflow about €3,000 to €5,000 · care from €250 per month per workflow

    Fixed fee in your proposal before work starts. In monthly care we run the workflow, fix errors and adapt it as your processes change.

The implementation of one contained workflow costs about €3,000 to €5,000. Projects with several processes and systems cost more, and the fixed fee is in your proposal before work starts. Ongoing care starts at €250 per month per workflow. 30 min · remote · no obligation. And if AI isn't the right tool, we'll say so.

Before you build

Questions about AI agents

01What is the difference between an agent and an automation?

An automation runs the same steps every time. An agent chooses between defined steps and tools depending on the case. Both run inside limits; the agent needs more testing.

02Can an agent act on its own?

Only where you allow it. By default, anything that leaves the company or changes customer data waits for a person.

03What is an MCP server?

A small program that gives an AI access to one system, for example a calendar or a CRM, through an open standard. It decides which actions are possible, so its permissions matter.

04Which tools do you use?

Mostly n8n for the workflow, models in EU regions or on your own servers, and interfaces or MCP servers to your systems. The choice depends on your software.

05How do we know the agent works?

With a set of real test cases agreed before go-live, including wrong inputs and cases where it must stop. We repeat the tests after changes. The implementation of one contained workflow costs about €3,000 to €5,000. Projects with several processes and systems cost more, and the fixed fee is in your proposal before work starts. Ongoing care starts at €250 per month per workflow. Read the article with the cost examples. Calculate with your own numbers.

Robert Van Ysendyck

Robert Van YsendyckFounder & Managing Director

Bring that task to a 30-minute call. You will know afterwards whether an agent or a fixed workflow fits and what it would cost.