An AI assistant that answers "in general" is handy for writing an email. It becomes a working tool when it can answer questions like "which orders from this customer are still open?" or "who approved this accreditation?". To do that, it has to read your company's data. The Model Context Protocol, or MCP, is the standard way to connect it.
The problem it solves
Until recently, connecting an AI assistant to a management system meant a custom integration for every pair: one assistant, one piece of software. Change assistant, and you started over.
MCP is an open protocol, introduced by Anthropic at the end of 2024 and now supported by the main assistants. It works like a standard socket: your software exposes, once, through an MCP server, the operations you want to make available. Any compatible assistant can use them.
What an MCP server contains
An MCP server is a small program that sits next to your management software. It declares three things:
- tools: the actions the assistant can request, for example "find a customer", "list overdue invoices", "draft a quote";
- resources: the data it can read, such as documents, product sheets, regulations;
- prompts: ready-made instructions for recurring tasks.
The assistant only sees what the server declares. It does not enter the database or see the tables: it asks, and the server decides what to answer.
Why it pays off, in practice
- People ask instead of searching. Someone in accounts asks "how much does this customer still owe us?" in plain language, instead of opening three screens and cross-checking.
- One connection, many assistants. The server is written once. If you change assistant tomorrow, the connection stays.
- Permissions stay yours. Each tool can require the user's login, respect their roles and record who did what. A sales rep sees their own customers, not everyone's.
- Sensitive actions stay with people. A good MCP server prepares a draft or a preview; then a person confirms.
What we learned building one
We wrote an MCP server for Media Accreditation, our product for press accreditation, so organisers can query their event data straight from an AI assistant. Three lessons that apply to any management system:
- Few tools, well described. The assistant picks a tool by reading its description. Ten clear tools work better than fifty generic ones.
- Short, pre-filtered answers. Sending thousands of rows to the assistant is costly and confusing. Better for the server to filter and summarise.
- Design security from day one. Who can call which tool, what data leaves, where it stays. If data must remain in Europe, the assistant has to be chosen accordingly.
Where to start
You don't need to connect the whole system. Start from one question your team asks often that takes time today: the status of an order, this week's deadlines, stock availability. A first MCP server with two or three tools takes a few days to build, and tells you right away whether the idea works.
If you want to know whether it makes sense for your software, tell us how you work: we'll reply with a concrete idea of what to connect first.