Custom AI or off-the-shelf models?
Off-the-shelf models are ready to use and powerful; a dedicated AI knows your company’s language better. How to choose, and why the best answer is often a combination of the two.
Technology · 3 min read
What an MCP server is, and why it matters
The Model Context Protocol is the standard way to connect an AI assistant to business tools and data. What it means in practice, and how to plan your first connections without mistakes.
The connection problem
An AI assistant is only as useful as the information it can reach. However powerful it is, a language model knows nothing about your customers, your orders or your procedures. To answer real questions, it needs to read data where it lives: in your management software, your email, your document archive, your CRM.
Until recently, every connection was a project in its own right. Letting AI read ERP data required a dedicated integration, then another for the document archive, and yet another for email. Every change of model or vendor meant redoing part of the work.
Without MCP
- One integration per system
- Rework every time the model changes
- Unclear exposed functions
With MCP
- One server per system
- Connections that work across models
- Declared, documented operations
What the Model Context Protocol does
The Model Context Protocol, or MCP for short, is an open standard that defines how an AI assistant can discover and use external tools. An MCP server is a small component that sits in front of a system, such as your management software, and exposes some of its functions in a format any compatible assistant can understand: finding a customer, checking the status of an order, retrieving a document.
The benefit for the business is clear. The connection is built once and can be used by different assistants and models. If you switch models tomorrow, the connections still work. If you add a new system, you simply add a new server without touching the others.
There is also a gain in clarity: each server declares exactly which operations it makes available. That makes it easier to understand, and to document, what the AI can and cannot do.
Where to start
Your first connection should be the one that answers the most frequent questions. Ask people what information they look for most often and where they find it, and it will quickly become clear which system to connect first. Often it is the management software for orders and invoices, or the document archive for procedures and contracts.
Starting in read-only mode is advisable. An assistant that looks up data and answers with sources already delivers a lot of value at limited risk. Write operations, such as creating a task or updating a record, can be added later, once people trust the tool and the confirmation rules are clear.
- 01 Collect the most common questions
- 02 Pick the most requested system
- 03 Connect it read-only
- 04 Add write actions with confirmation
Security and permissions
The most sensitive point is identity. An MCP server should act with the permissions of the person asking the question, not with a technical account that can see everything. Otherwise the assistant becomes a shortcut to data the person would not be able to access directly.
The operations you expose also need to be chosen carefully. A few well-defined functions are better than full access to a database. Every request should be logged, with the user, the tool used and the data read, so you can reconstruct at any time what happened. It is good security practice and, in effect, a requirement for GDPR compliance whenever personal data is processed.
A final tip concerns credentials: servers that access critical systems should use encrypted connections and expiring keys, managed by the system administrators rather than by individual users.
MCP server security
- Permissions of the person asking
- A few well-defined functions
- Every request logged
- Encrypted connections
- Expiring keys
MCP in the Innovaq suite
Innovaq’s AI assistant uses MCP both internally and externally. The suite’s modules, from CRM to Finance and from People to Regulate, are already accessible to the assistant, always within the user’s permissions. For external systems, dedicated MCP servers or API connections are added, configured by choosing which tools to expose.
Every request goes into the log, and write operations always require the user’s confirmation. The result is an assistant that knows the business as a whole, without the business losing control over what it sees and what it does.
Keep reading
