MCP: how to connect AI to the software you already use

You have been using ChatGPT or Claude for a while. For rewriting an email or summarizing a document it works fine. But the moment you ask "which quotes are still open with that client in Breda?", it gets stuck. The AI cannot look inside your CRM. So you copy data back and forth, paste an export into the chat window, and hope you did not miss anything. Is there another way?

There is. More and more often, it is MCP. In this piece we explain what it is, what an MCP integration can mean for your business, and what such a connection looks like in practice. Including where you should be careful.

What is MCP, actually?

MCP stands for Model Context Protocol. It is an open standard that Anthropic introduced at the end of 2024, and several AI tools have added support for it since. The idea is simple: one agreed way for an AI assistant to talk to other software.

Think of a plug. Every device used to come with its own charger. Now there is USB-C, and almost everything fits the same cable. MCP tries to do the same for AI and software. Instead of building a separate connection for every AI tool, you build one MCP server on top of your software. Any AI assistant that speaks MCP can then work with it.

That MCP server decides what the AI is allowed to do. It offers a few clearly defined actions, which MCP calls tools. For example "look up a customer" or "show open quotes". The AI sees which tools exist, picks one when it fits your question, and gets the result back. Anything not offered as a tool, the AI simply cannot do.

That last part matters. The AI does not get free access to your database. It gets a few doors, and you decide which ones.

Why this matters for small businesses

Most small and mid-sized businesses do not run on one system. There is a CRM, a planning tool, an accounting package, often a customer portal or web app of their own, and somewhere a shared folder full of documents. It all works, but the information is scattered.

AI only gets really useful once it can reach that information. An assistant with only general knowledge writes tidy text. An assistant that knows your customers, your planning, and your quotes can take actual work off your hands. What that looks like per process, we lay out in AI automation for SMEs.

Until recently, that meant building a separate connection for every combination of AI tool and software. Expensive, and it tied you to that one AI vendor. With MCP you build the connection once, on the side of your own software. If you want to switch to a different AI assistant next year, that connection does not need rebuilding.

There is another upside. Many packages you already use offer their own MCP server, or are working on one. For those tools you sometimes do not need to build anything at all. The real work then sits with the software only you have: your own web application, an older custom system, or a database that has held your customer data for years.

An MCP integration in practice: an example

Let us make it concrete. Say you have a web application where your team manages customers, planning, and quotes. It was custom-built at some point and runs fine. You want your account managers to be able to ask it questions through an AI assistant, without clicking through five screens every time.

We would then build an MCP server on top of that web application. It talks to the same API your application already uses, and offers a handful of tools:

• Look up a customer. By name, city, or customer number. Returns contact details and the most recent touchpoints.
• Show open quotes. Per customer or for the whole team, with amount and date.
• Check the planning. Who is scheduled with which customer, and when.
• Prepare a draft quote. Based on an earlier quote and what the account manager asks for.

An account manager now types into the assistant: "Which quotes are still open with customers in Breda, and when did we last visit them?" The assistant calls the right tools, combines the answers, and gives an overview. No export, no copy and paste.

Start with read-only

The first three tools on that list only read. They change nothing. That is almost always where we start. You quickly see whether the assistant fetches the right things, and if it gets something wrong, nothing breaks.

The fourth tool does write something. That one gets a brake: the assistant makes a draft, but a person confirms it before it lands in the system. Only once reading has gone well for a while do we carefully add more write actions.

Permissions and logging

An account manager should not suddenly see the numbers for the whole company just because the AI happens to have access. So the MCP server works with the same permissions as the user. If someone only sees their own customers in the web app, they only see their own customers through the assistant too.

We also log every call: which tool, by whom, with which question. If something odd happens, you can trace what went on. That also comes in handy when you later need to show how you handle AI.

Want the assistant to search loose documents as well, like manuals, contracts, or delivery terms? Then you add a piece that makes your own text searchable. How that works, you can read in RAG explained.

What does it cost and how long does it take?

That mostly depends on what your existing software looks like. If your web app already has a clean API, an MCP server on top of it is relatively quick to build. A first working version with a few read tools is often ready in 2 to 4 weeks. If an API has to be built first, or the MCP server needs to talk to an old system that was never meant for integrations, it takes longer.

On top of that, count on running costs for the AI model you use. At normal small-business volumes that usually lands somewhere between €50 and €500 a month, depending on how many people use it and how heavy their questions are. Curious how the costs of a bigger AI system add up? Have a look at what it costs to build an AI agent.

What to watch out for

MCP is useful, but it is not a magic wand. A few things we keep an eye on in every project:

Too many tools at once. The more actions you offer, the more often the assistant picks the wrong one. Start small. Five well-described tools work better than thirty vague ones.

Sensitive data. Everything a tool returns goes to the AI model. So think about which fields you send along. A customer name and a quote amount is something else than a social security number or medical details. Also check with your AI provider where that data is processed and what they are allowed to do with it.

Trusting the AI with write actions. An assistant can misread a question. When reading, that is annoying. When writing, it can do damage. Keep a person between the draft and the final action, especially at the start.

A standard that is still moving. MCP is young. The specification still changes, and not every AI tool supports everything equally well. So build it in a way that lets you move along with it, and do not rebuild your whole business around it in one go.

Where do you start?

Pick one question your team asks often and that makes them click through several screens. Build a few read tools for it. Let a small group of colleagues work with it for a few weeks and see what happens. Does it get used? Is what the assistant fetches correct? After that you know whether expanding makes sense.

Already have a web application and want to work AI into it step by step? Adding AI to your existing web application takes you further. And if you do not have software of your own yet but keep running into disconnected systems, take a look at web application development.

Want to talk through your systems?

Not sure whether an MCP integration pays off in your situation? Feel free to run it by us. Together we look at which software you use, where time is leaking away now, and whether a connection is worth it. Sometimes the answer is that a simple export does the job, and then we will tell you so. Book a call.

Frequently asked questions

Do I need my own AI model for MCP?

No. You use an existing AI assistant that supports MCP. What gets built is the MCP server on the side of your own software. That server decides what the assistant can look up or do.

Can the AI change anything it likes in my system through MCP?

Only what you offer as a tool. We usually start with tools that only read. Write actions come later, and then with a person confirming before anything actually changes.

Does MCP also work with software I did not have built myself?

Often, yes. More and more packages offer their own MCP server. If yours does not but it does have an API, an MCP server can be built around it.

How long does a first MCP integration take?

If your software already has a clean API, a first version with a few read tools is often ready in 2 to 4 weeks. Without a usable API it takes longer, because that has to come first.