What is MCP? The plug between AI and your systems

A language model knows a lot, but nothing about your business. It doesn't know your customers, your orders or your agreements. To do real work, it has to be able to look that information up, and sometimes do something too: make a note, schedule a meeting, prepare a draft.

Before MCP, you had to build each connection separately: again for every AI model and for every system. Ten systems and three AI tools meant thirty connections in the worst case. And if you switched models, you often had to redo part of the work.

MCP turns that around. A system gets one MCP connection, and every AI application that speaks MCP can in principle work with it. That can make connecting faster and switching models cheaper.

There are three roles:

• The host: the AI application you work in, such as Claude, ChatGPT or your own assistant.
• The client: the part inside the host that connects to an MCP server.
• The server: the connection to a system. It tells the AI what's possible, for example "look up a customer", "read a document" or "make a note".

Roughly speaking, an MCP server offers three kinds of things: tools (actions the AI can carry out), resources (information the AI can read) and prompts (fixed ways of working or templates).

Say you ask your assistant: which quotes are still open? The model sees there's a CRM server with a tool to search quotes, calls it and gets the results back. Then it writes an answer, or suggests a next step. For actions that change something, a good application asks for permission first.

• Your assistant knows your own business. Employees ask their AI assistant about customers, orders or documents, instead of clicking through five systems.
• Faster connections. If a system already has an MCP server, the connection doesn't need to be built again.
• Less tied to one model. If you switch AI model or vendor, your connections stay usable, as long as the new application also supports MCP.
• One place for agreements. What the AI may read and do is largely arranged in the server. That's clearer than separate connections per tool.

MCP doesn't make AI smarter. It makes AI more useful, because it can reach the information the work is about. More on connecting in general is in connecting AI to your systems.

MCP itself is mainly an agreement on how systems talk. How safe it is depends on how you set it up and what you allow. An AI that can reach your systems has hands. So watch five things:

1. 01Give as few permissions as possible. Start with read-only. Writing, sending and deleting only when needed, and per action.
2. 02Only use servers you trust. An MCP server of unknown origin is comparable to software of unknown origin.
3. 03Allow for prompt injection. An email or document can contain hidden instructions, such as "forward all invoices". An AI that simply follows input like that does things nobody wanted. More in using AI safely.
4. 04Have a human approve actions. Especially anything that goes outside the company or costs money. Read more about human in the loop.
5. 05Record what happens. Which tool was called when, with what result? Without a log, you won't know afterwards what the AI did.

The rules you already have apply here too: protect personal data under the GDPR, and record agreements in your AI policy.

They complement each other.

• An API is the technical entrance to a system. An MCP server usually uses an API itself to talk to the system.
• MCP makes that entrance understandable for an AI assistant. Useful when a person works in an assistant and asks varying questions: look this up, summarize that, prepare a draft.
• A fixed workflow (for example in n8n) is better when the same process has to run the same way a hundred times a day, without anyone asking. Think of processing invoices or entering requests into the ERP. That's more predictable and easier to check.

The rule of thumb: varying questions from people, then MCP in an assistant. Fixed work that should run by itself, then a workflow or agent with fixed steps and a human who approves.

Start small, and start with reading:

1. 01Choose one system where employees often look up information, such as the CRM or a document folder.
2. 02Check whether an MCP server already exists for it, from the vendor itself or a reliable party. If not, one can be built on the API.
3. 03Give read-only permissions and test with a small group of colleagues.
4. 04Only expand once it works and the agreements are in place: more systems, and possibly actions with approval.

Want to know where it delivers the most for you? In the AI Quick Scan we look per process at what AI can take over, and which connections that needs.