Skip to content

How we automate

AI agents that work inside your software, not just talk.

Everyone has seen a chatbot. The step that makes the difference is a different one: we build an MCP server connected to the software you already use, and AI agents work through it. Without that server an agent can only talk. With it, it can look up an order, prepare a quote, book an appointment and write to the CRM, with rules, permissions and a log.

People and agents

  • Assistant on a colleague’s computer
  • Process agent, runs on its own
  • Voice agent, on the phone

The MCP server

built by us, installed with you

  • fetch balance and overdue invoices
  • check free time slots
  • prepare a quote from the catalogue
  • write the note to the CRM
  • permissions per tool
  • a log entry per call
  • answers with their source

Your systems

  • ERP / invoicing
  • Email
  • Calendar
  • CRM
  • Drive / files
  • Telephony
The simplified picture: agents never touch your software directly. Everything goes through the MCP server built for your company. The server lives on your server, in your cloud or on a machine in your office. Your data never comes to us.

What an MCP server is

A short list of tools, defined by us, in plain business terms.

MCP (Model Context Protocol) is a standard that lets an AI agent use tools. An MCP server is the list of tools we build for your company, one by one: “find the customer by tax ID”, “fetch the balance and overdue invoices”, “read quote requests from the inbox”, “check free time slots”, “create an appointment”, “prepare a quote from the catalogue”, “write the note to the CRM”.

It does one thing

Each tool has one clearly defined job, in one system: ERP, invoicing, email, Drive, CRM, calendar, telephony or a database.

It has explicit permissions

Read or write, which fields, which customers, which amounts. The agent never gets the ERP password; it gets a short list of tools, with limits.

It leaves a trace

Who asked, what ran, when, with what result. The log is yours and can be checked at any time.

It returns the source

The agent shows where the answer comes from, not just the answer. A colleague can verify it in seconds.

The engineering part, connectors, access rules, logging and tests, is what we have been building since 2008 under other names: integrations, APIs, migrations. What changed is that these connectors can now be used by an agent that understands plain language.

Who works through the server

Three kinds of agent, chosen by the process.

Agent on a person’s computer

An assistant installed for the accountant, the sales team or the director. People ask in their own words; the agent searches the permitted systems, runs the steps and shows the source. The AI model can be swapped at any time; the server stays.

Process agent

Runs on its own, on the server, for a repeating process: reads requests, extracts the data, prepares the document, asks a person to approve before anything irreversible, then executes.

Voice agent

Answers the phone at any hour, checks the real calendar through the MCP server, books the appointment, writes to the CRM and hands over to a person, with context, when it is out of its depth.

A working day

What it looks like in practice

Three illustrative situations with fictional data.

  1. 01

    The quote request

    Today

    A distributor receives requests by email every day. Someone reads the message, looks up prices in three files and copies everything into a quote.

    With automation

    The process agent reads the request, identifies products and quantities through the “search the catalogue” tool, prepares the quote from a template and submits it for approval. A person checks and presses “send”.

    What changes: The quote no longer waits in the inbox. The person checks instead of typing.

  2. 02

    The balance question

    Today

    A sales colleague needs to know whether they can deliver to a customer with overdue invoices. They ask finance, who searches the ERP and a spreadsheet.

    With automation

    The colleague asks the agent on their own computer. It uses the “balance and overdue invoices” tool and replies with the figures, the last payment date and the company’s credit-limit rule, with the source.

    What changes: The answer arrives in seconds and nobody is interrupted.

  3. 03

    The 7:40 pm call

    Today

    A clinic gets a call after hours. It rings out.

    With automation

    The voice agent takes the call, checks free slots through the MCP server, proposes two options, confirms the booking and writes it to the calendar and the CRM. For an emergency, or when the caller asks for a person, it records the request and passes it to the team.

    What changes: The call is no longer lost.

What stays with people

The agent applies rules. People decide.

Approvals

Every irreversible action (sending, paying, deleting, a firm confirmation) goes through a person, until the company decides otherwise based on the log.

Permissions

The agent sees and does only what it has been allowed to, like a new employee.

Rules

We write them together with the people who do the work today. The agent applies rules; it does not invent them.

Verification

Every answer comes with its source. Every action is in the log.

Why MCP

Why an MCP server, and not “a chatbot”

A standard, not a lock-in

The MCP server works with several AI models and applications. When a better or cheaper model appears, the model changes, not the integration.

Reusable connectors

The same tools serve the desktop assistant, the process agent and the voice agent. They are built once.

Security per tool

We never hand the agent the ERP password. It gets a short list of tools, with permissions and limits, and every call is logged.

No data with us

The server stays with the client. We build, install, test and train the team.

How a project runs

One process first. A fixed price. A result you can check.

01

A 30-minute call

free

We pick one process: the one that repeats most often and hurts the most.

02

A two-day analysis

€800

We check system access, data quality and the rules of the process. A written plan, success criteria and a fixed-price quote. The analysis is credited against the first project.

03

The first working process

from €2,500

Usually one to two weeks after access is confirmed. A workflow from €2,500, a voice agent from €3,000, an assistant connected to the first system from €4,000.

04

Extension

step by step

Once the first process runs and the log shows what it does, we add the next tool or the next process. Maintenance is optional, from €250/month.

Running costs (AI models, telephony, hosting) are paid directly to the providers; we estimate them before launch, based on volume.

See the full pricing

What we do not promise

We state the limits before, not after.

  • We do not promise savings percentages. We say what changes concretely and measure after launch, against a baseline agreed together.
  • We do not put AI where a simple rule or an integration will do. Sometimes the analysis concludes “a project does not make sense”.
  • Messy data stays messy until it is cleaned. The hard part of these projects is not the AI; it is access, data and the people who have to come to trust the system. That part we have been doing since 2008.

Questions

Frequently asked questions

No. The MCP server connects to what you have: ERP, invoicing, email, calendar, CRM, files. If a program has no API, we find another way or say clearly that it cannot be done.

With you: on the company server, in your cloud or on a dedicated machine. The agents on people’s computers install like any application.

The one that fits the process: for speed, for precision or for cost. The model can change without rebuilding the integrations.

The AI model receives only what the tool sends it, for that request. We agree together which fields may leave and what stays inside the company.

The first working process usually within one to two weeks after access and scope are confirmed.

That is why irreversible actions go through a person, and why we start with the analysis rather than the implementation. The log shows exactly what it did and why.

In short

If you remember one sentence

The difference from a chatbot is that we connect the agent to your systems through an MCP server built for your company. That is how it can look up an order, prepare a quote or book an appointment, rather than just talk. We start with a single process, the one that costs you the most time, and you see it running within one to two weeks.

Which process would you start with?

In 30 minutes, free of charge, we describe it together and tell you which tools your company’s MCP server would need.

We reply the same business day.