Headless 360: new label or new value?
BLOG
Multiple systems in one AI chat: how to build a scalable MCP architecture
You’ve connected Salesforce to your AI chat and you’re asking your first questions. An update on an account, logging a conversation, a chart for the management team: it works. Until you ask a question that goes beyond Salesforce: “Where do we stand with customer X?” That’s because the full answer is spread across several places. The active deals are in Salesforce, the orders and outstanding invoices in your ERP, and last week’s complaint in your ticketing system.
Prefer watching to reading? (The video is in Dutch.)
The more sources, the more complete the answer
For your AI to answer that question well, it needs access to all of those systems. Some systems make that easy: like Salesforce, they already have their own MCP server ready to go. Many other systems only have an API, a technical entry point your AI can’t simply talk to. For those, we build an MCP layer on top. That way, you can query that system from the same chat too.
Why separate connections are inefficient
The obvious step is to hook each MCP server up to your chat separately. That works, but for every question your AI is then handed a full toolbox: all the tools of all the systems. A tool is a single action your AI can perform through an MCP, such as looking up a record, retrieving an order or creating a ticket. Your AI has to work out for itself which tools it needs, in what order to use them and how to stitch the answers together. For the question about customer X, it does that from scratch every time.
That has three consequences. It takes longer. It uses more tokens (the units your AI usage is billed in), and the answer comes out slightly different every time, because the AI can pick a different route each time. For a single question, that makes little difference. When your whole team asks the same question dozens of times a week, you’ll notice.
All systems connected to the chat separately
One question, one tool
We take a different approach. Together with you, we map your system landscape and data models, and we look at which questions come up most often in your organisation. For those questions, we build the route into a custom tool. When you ask where things stand with customer X, that tool goes through Salesforce, your ERP and your ticketing system behind the scenes and returns the complete answer in one go.
Your AI no longer has to work anything out for itself. The answer arrives faster, it uses fewer tokens and you get the same reliable result every time.
One tool that touches three systems
Free to choose your AI vendor
This setup has another advantage: the layer of tools is separate from the chat you put in front of it. Do you work with Claude today and switch to Copilot next year, or to something that doesn’t exist yet? Then you only swap the front end, and everything underneath stays in place. You can even connect multiple AI vendors to the same layer at the same time, for example when teams work with different tools.
How we work with you
From the first connection to a complete architecture: we look at the bigger picture together with you. Which systems do you connect, and which questions do you want to answer with them? How do you set up the layer of tools, and who can use what? And which AI model and which subscription fit your questions, what you are allowed to do with your data and your budget?
Want to know what a scalable MCP architecture means for your organisation? Go to contact and leave your details, and we’ll get in touch with you.
