Connecting Your Business with MCP Servers
The Problem With Today's AI
Most AI tools today live in silos. ChatGPT, CoPilot, Claude, or embedded AI in apps: each is valuable, but none connects seamlessly across a company's systems. Businesses don't work in silos; they operate across ERP, CRM, spreadsheets, documents, and email. Integrating AI into this landscape usually requires custom REST APIs, one-off scripts, or vendor lock-in. This creates friction, cost, and fragility.
Enter MCP Servers
Anthropic introduced the Model Context Protocol (MCP) in November 2024. It provides a standard way for AI models to access external tools, data, and actions. Think of MCP as the USB-C of AI: one plug that works across all systems. Instead of building custom integrations for each model or vendor, you define reusable "tools" (capabilities) once. Any MCP-compliant AI can then use them.
Unlike static APIs, MCP is flexible, lightweight, and designed specifically for AI workflows. It focuses on capabilities rather than rigid endpoints, letting you describe what can be done (e.g. searchOrders) and what is returned (e.g. delivery times).
Types of AI
- Embedded AI: ML in sensors, vision, etc. (outside today's focus)
- Chatbots: Generic assistants that answer based on inputs
- Integrated AI: AI built into apps to assist with tasks
- Agents: Autonomous workflows. Promising, but not universally reliable yet
MCP fits above these. It's not another assistant; it's the missing interoperability layer that connects them all.
How MCP Works
MCP defines a server that exposes tools, structured in JSON schemas, describing what actions an AI can take. The AI client (ChatGPT, Claude, etc.) queries the server for available tools, then executes them with arguments. Importantly, the model itself can reason about which tools to call, in which order, and how to combine outputs to solve multi-step business problems.
Example Toolset for ERP Orders
{
"tools": [
{
"name": "searchOrders",
"description": "Search orders by ID or name",
"input_schema": {
"type": "object",
"properties": {
"id": {"type": "string"},
"name": {"type": "string"}
},
"required": ["id"]
}
},
{
"name": "getOrderDetails",
"description": "Retrieve details for a given order ID",
"input_schema": {
"type": "object",
"properties": {"orderId": {"type": "string"}},
"required": ["orderId"]
}
}
]
}This schema describes actions available. An MCP client (like Claude) can immediately interpret, validate inputs, and call them without custom glue code. Crucially, the model doesn't just call one tool—it can chain them logically: searchOrders → getOrderDetails → getSubTreeOfOrder → getBatchDetails. The AI figures out the sequence, so humans don't have to hard-code workflows.
Case Study: German Clothing Manufacturer
At SFLOW, we applied MCP to a company struggling with data silos:
- Legacy ERP system (no API, only exports)
- Central Excel file
- Shared drive documents
- Shared mailbox
Business Questions
- "What are the delivery times of the sub-components of order XXXX?"
- "What did my supplier say about their next delivery?"
- "Are there shipping delays in the next 6 weeks?"
Previously, this required manual digging. No single AI or app could unify ERP exports, Excel, and emails.
Implementation
- Automated script exported ERP data to Excel
- MCP Server wrapped this export and exposed tools
- AI client (Claude with MCP support)
Result: a single chat interface could answer questions spanning ERP, spreadsheets, and email without centralizing everything into one monolith.
Demo
For this we used the MCP-server we set up for this German company and fed it with dummy data to see it in action. In this first phase it was decided we would fall back on an hourly incremental backup of the ERP system where the MCP-server could get its data.
For the AI agent we used Claude Desktop as they support MCP-servers in the best way at the time of writing.
Asking about the business
Imagine waking up and wanting to know about the business? Just ask your AI companion.
You can really "see" the reasoning steps the AI is taking to get to its goal. We gave a very vague description, but it tries to find its way around it. The arrows at the right side also allow you to directly check the data that is sent and received.
You can use the built-in text-generation capabilities of a model to generate a nice overview of the things you want to see.
In this case the AI provides a very verbose outcome, but it's perfectly possible to ask it to be more to-the-point, friendlier, etc.
Diving deeper
But the real power lies beneath—you're able to ask follow-up questions about things you see, and dive deeper into topics.
Here is where MCP really shines: you can directly "dive deeper" into a topic without needing to switch context. Ask questions, explore data...
The model will even propose actions that you could take as a next step based on the information it receives.
The core capability here is that you, as a person, are always in the driver's seat. You control, check and guide the AI on where to go.
Combining multiple tools
The MCP-model really shines when you combine multiple tools together. In this case, Claude goes above and beyond and starts searching for alternatives of one of your problematic suppliers. This is not always what you want, but it really lets you connect the "outside" world to your inner business operations.
In this case the AI went a bit further and tried to find alternatives. In a more realistic scenario you could ask how a real world event could impact your supply chain.
Taking action
Let's ask our AI to send a message to the client about the delay and apologize.
This is a very simple example, but it really shows the connectedness of everything. Instead of looking through the ERP, emails, contacts, etc., we just have one single interface from where we are in the cockpit.
We skipped some parts here—you can iterate as long as you want about the draft, cancel it, or just send it. As we have an "SMS-MCP" installed, we can send a message as easy as saying "send it".
And it gets sent as if it was you sending this message.
Your AI really becomes as powerful as the tools you give it. But MCP radically simplifies the way you can start small, test, iterate and grow from there.
Note: for the screenshots the information was obviously anonymized and scrambled, but the outcome is the same.
Why MCP Is Different From Just an API
- Standardized: Any MCP-compliant AI can consume the tools without custom adapters
- Self-describing: JSON schemas make tools discoverable to AI automatically
- Composable reasoning: The model can chain tools and plan next steps itself
- Open: MCP is open-source and free, avoiding vendor lock-in
Should We Care?
- Microsoft already started implementing MCP-servers inside their tools (Azure DevOps MCP Server)
- GitHub has their own (GitHub MCP Server)
- And many more
So we would say "yes". If you are a business using tools or a product supplier of software systems, it's your task to provide this access.
What About Security?
Some people joke that the "S" stands for security in MCP-servers. And that is to a great extent true. Especially if you go out and install external ones which you cannot verify, you expose yourself to great risks.
As with all new technology this is a learning curve, but most importantly, you do get better control by being able to audit what goes in and out of such MCP servers—opposed to employees throwing full exported Excels into ChatGPT.
The Takeaway
MCP is early but powerful. Instead of building brittle integrations for each AI product, you define your company's capabilities once, and future AI models can all plug in. For businesses, this means:
- Faster AI adoption across systems
- Lower integration costs
- Flexibility to evolve without rewriting connectors
And most importantly: you don't need to predefine workflows. With MCP, you leverage the model's reasoning power to decide which tools to use, in which order, and when—turning AI from siloed copilots into a true business companion.