How to Connect Your Business System to AI in 4 Weeks with MCP

Most companies have now tried AI. Perhaps ChatGPT for writing content, Copilot for summarizing meetings, or some generative AI service to reduce a bit of manual work from everyday life.

But there's one thing that type of AI tool cannot do: it doesn't know what's happening in your systems right now.

It doesn't know that you have 47 unresolved support tickets in the queue. It doesn't know that a customer just placed an order worth 380,000 SEK. It doesn't know that your next event starts in three days and you've only sold 40% of the tickets.

That's the difference between AI as a conversation tool and AI as an integrated part of your operations.

Model Context Protocol – MCP – is the open standard that makes it possible to take the step from the former to the latter. Without needing to switch systems. Without needing to rewrite code. And, if you do it right, without it taking more than four weeks from idea to working integration.

At Smart Source, we've done it four times – for customers and for ourselves. In this article, we explain exactly how.

The Problem with AI Adoption in 2026

Everyone is buying AI. Almost no one is integrating it.

We talk with CTOs and IT managers every week. They rarely lack interest in AI. Often they've already purchased some form of AI service. The problem is different: the AI lives outside the systems.

It's a paradox that has become clearer the longer the AI hype continues. On one hand, we have powerful AI models that can reason, plan, and communicate in a way that feels genuinely intelligent. On the other hand, they can't answer a simple question like "How many active customers do we have?" – because they don't know. Nobody has told them.

Traditional system integration is the answer most people think of. Connect the ERP system to the AI tool via an API. Let data flow. But that's easier said than done.

Standard integrations have three classic problems:

Complexity. Each integration is its own project. An ERP integration can take six months and cost 500,000–1,000,000 SEK. That's hard to justify for an AI use case that hasn't yet been proven.

Maintenance. Integrations break when systems are updated. You get a new version of the ERP – and suddenly nothing works. Someone needs to fix it, and it's rarely someone who planned for that.

Lock-in. Traditional integrations are point-specific. You build a connection between System A and System B. If you then want to add an AI assistant, you need another integration. And another. And another.

The result is that most organizations end up in one of two camps:

1. They do nothing. AI projects stay in the pilot phase, consuming calendar time but delivering nothing to production.

2. They build expensive, complex integrations that take time and resources – and ultimately deliver limited value relative to the investment.

That's the gap where MCP exists.

What is Model Context Protocol – and Why is it Different?

Model Context Protocol (MCP) is an open standard for connecting AI models to external data sources and systems.

It's important to note: MCP is not a Smart Source tool, not a startup concept, and not yet another SaaS subscription you have to pay for. It's an open standard published by Anthropic and now supported by most major AI players – including Claude, GPT-4, Gemini, and open-source models like Llama and Mistral.

The core principle is elegant:

An MCP server exposes your existing systems as a number of "tools" – simple, well-defined functions that an AI model can call. The tool could be "Fetch customer order" or "Create new ticket" or "Show available events". The AI model reads which tools are available, and when a user asks a question, it can call the right tool, get back the answer, and present it in a meaningful way.

That's all. No large platform needed, no new interface, no rewriting of existing systems.

Technically, an MCP server is a relatively thin wrapper around your existing APIs. If your business system already has a REST API – and almost all modern systems do – a large part of the work is already done. You don't need to open the system externally, nor give the AI model access to the entire database. You expose exactly the functions you want the AI to be able to use.

That makes MCP one of the safest paths into AI integration available today.

Another major advantage is model-agnosticism. Since MCP is an open standard, the same MCP server works with whichever AI model you choose. You can start with Claude, switch to GPT-5 when it arrives, or run a local Llama model for GDPR reasons – and the MCP server doesn't need to change.

In short:

• Builds on existing APIs – no rewriting of your systems

• Works with all major AI models

• Open standard – you own the code, no vendor lock-in

• You control exactly what data the AI can access

RAG vs MCP – What's the Difference?

Before we move on to concrete implementations, it's worth clarifying a common confusion: the difference between RAG and MCP.

RAG – Retrieval-Augmented Generation is a technique where the AI is fed documents and can then search through them. You upload your manuals, policy documents, FAQ pages – and the AI can answer questions based on that content.

RAG is excellent for:

• Internal knowledge bases and handbooks

• Customer service bots that answer common questions

• Search through large amounts of static text

But RAG has a fundamental limitation: it's passive. The AI reads. It cannot act. It always works against a snapshot of data, never against real-time information.

MCP solves the other problem: making the AI capable of doing things – not just reading. Create an order. Update a ticket. Fetch real-time data. Book a meeting. Log time.

They're not competitors – they complement each other. You can have a RAG solution for your product documentation and an MCP integration against your CRM. But if you want AI to be an active part of your operational business, MCP is the right path.

Four MCP Implementations – What We Built and What We Learned

We believe more in showing than telling. Here are the four MCP servers we've built and what they achieve.

1. Eventry – AI Assistant for Ticketing System

Eventry is an event platform where organizers manage events, tickets, and attendees. The assignment was to connect an AI assistant to the platform so organizers can interact with their events through natural language.

We built an MCP server with 21 endpoints – tools for searching events, managing tickets, viewing attendee lists, creating campaigns, and more.

The result:

• Implementation: 4 weeks

• Lines of existing code rewritten: zero

• Exposed functions: 21

An organizer can now ask: "How many tickets have we sold for next weekend's event?" and get a direct answer. Or: "Create a 15% discount code for VIP guests" – and it happens directly in the system, without navigating the interface.

2. Umbraco – AI-Assisted Content Management

Umbraco is one of the most widely used CMS systems in Sweden and the Nordics. We've built an MCP server that exposes Umbraco's Content Delivery API as tools for AI models.

What the AI can do:

• Search and retrieve existing content from Umbraco

• Create new pages and content objects

• Update texts and publish changes

• Navigate and analyze the content tree

An editor can now give an AI assistant the instruction: "Find all product pages that haven't been updated in six months" or "Create a draft landing page for our new service based on our existing tone" – and the work happens directly in the CMS.

For organizations running Umbraco, it means that AI integration is possible without buying a new CMS or a separate AI tool. The MCP server acts as the bridge.

3. Kimai – AI for Time Tracking

Kimai is an open-source time tracking tool, and we use it internally at Smart Source. We built an MCP server for Kimai for a simple reason: we wanted to be able to log time via natural language.

Instead of navigating to the interface and finding the right project and activity, we can now tell our internal AI assistant: "Log 2 hours on the Gothenburg City project, activity 'tendering process', for today" – and it's done.

It's a small example – but it illustrates the core principle: MCP eliminates the manual intermediary step. The AI knows which tool to use, when to use it, and how.

4. Krayin CRM – AI-Assisted Sales Work

Krayin is our CRM system. With an MCP server, we can now interact with the CRM directly through the AI assistant: create leads, update contact status, search among opportunities, and log activities.

"Create a new lead for Malmö City with contact Anna Lindberg, IT Manager, mark it as warm and assign to Mattias" – and the lead is entered with the right fields, right status, and right assignment. That's the same work that takes 3–4 clicks in the interface. With MCP, it takes a second.

What We Learned from the Four Implementations

There's a clear pattern across the four projects:

1. Implementation takes 2–4 weeks if the system has an existing API

2. No existing code needs to be rewritten

3. The interface disappears – conversation takes over

4. Security is manageable – we expose exactly the endpoints we choose

5. Value is immediately measurable – time saved, errors reduced, throughput increased

Want to know if your systems are ready for MCP? Contact us and we'll take a look together – free of charge.

What an MCP Implementation Looks Like in Practice

We divide a typical MCP implementation into four phases.

Phase 1: Mapping (1–3 days)

Before we write a single line of code, we map which use cases create the most value. What questions does your staff ask the systems every day? Which tasks take the longest? Which processes are most repetitive?

We also look at existing API documentation. Most modern systems – ERP, CRM, CMS, service management – have REST APIs. That's the starting point. If documentation is missing, we help create it.

Phase 2: MCP Server (1–2 weeks)

We build the MCP server in .NET/C# – our primary stack – or whatever language best fits the project. Each "tool" is defined with a name, a description, and parameters.

The description is central – it's what the AI model reads to understand when and how the tool should be used. A well-written tool description is the difference between an AI system that works well and one that guesses incorrectly.

Phase 3: Integration and Testing (1–2 weeks)

The MCP server is connected to an AI model and tested against real scenarios. We test security boundaries – what should the AI be allowed and not allowed to do? We test edge cases – what happens if a customer doesn't exist? We fine-tune the tool descriptions so the AI model makes correct choices.

It's also in this phase that we decide which AI model to use: cloud-based (Claude, GPT-4) or local (Llama, Mistral via Ollama). The decision significantly affects the GDPR picture.

Phase 4: Deployment and Handover (2–3 days)

The MCP server is deployed in your infrastructure. We document what was built – how tools are defined, how the server is maintained, how you add new tools going forward. You own the code and can manage it entirely on your own terms.

Total time: 2–4 weeks. Cost: 100,000–300,000 SEK depending on the system's complexity and number of use cases.

MCP, GDPR, and the EU AI Act

One of the most common questions we get is: "But where does our data end up?"

It's an important question – and with MCP, the answer is clearer than with most AI services.

The data never leaves your systems – unless you want it to.

The MCP server runs in your infrastructure. If you choose a local AI model (like Llama or Mistral via Ollama), all data stays within your organization. Even if you choose a cloud-based model like Claude or GPT-4, you control the flow: you decide which endpoints are exposed, what data the AI has access to, and where the MCP server runs (Sweden, EU, or on-premise).

EU AI Act and traceability. The EU AI regulation requires traceability, documentation, and human oversight for AI systems in sensitive contexts. MCP architecture is fundamentally compliance-friendly: every tool is documented, every call is traceable, and you retain full control. That's a sharp contrast to sending sensitive business data to an external AI service and hoping for GDPR compliance.

For the public sector – where requirements for data locality, transparency, and security are high – MCP is one of the better choices for AI integration available today.

Open standard – you own everything. Since MCP is an open standard, you own the code completely. There's no monthly subscription, no vendor lock-in, no risk that the service shuts down. You can build further, adapt, and manage entirely on your own terms.

Common Objections – and Honest Answers

"We have legacy systems that don't have modern APIs."

It depends on what you mean by legacy. Systems from the 2000s often have SOAP services or direct database access. It's possible to build MCP wrappers around those too – albeit with a bit more work. An initial mapping quickly shows whether it's feasible. We do that mapping at no cost.

"We don't know which use cases provide the most value."

That's actually a great place to start. We typically lead a half-day workshop with the daily system users – they know exactly which tasks are most repetitive and most frustrating. From there, it's easy to prioritize the three to five tools that create the greatest immediate value.

"Yet another AI tool to maintain."

The MCP server isn't a tool for your users – it's infrastructure. The interface to the AI can be exactly the same tool you already use (Teams, Slack, an internal web app). There's no new application to train people on. Users notice that the AI has become smarter – not that a new system has arrived.

"What happens if the AI makes a mistake?"

You define which actions the AI can perform. Read operations are always risk-free. Write operations can be limited to specific values, require confirmation from the user, or logged for audit. An MCP server can be designed with multiple layers of security and human oversight – exactly what the EU AI Act asks for in high-risk systems.

"We don't have budget for a large AI project."

A basic implementation is not a large project. 100,000–150,000 SEK and four weeks gives a proof of concept with real results in a production environment. If it delivers value – and it usually does – it's easy to justify the next step.

The Future: From AI Tools to AI Agents

We're in a phase where most organizations are testing AI as a tool. You paste in text, ask for suggestions, and use it as an advanced search tool. That's a reasonable first step.

The next phase is different.

AI agents – systems that can plan and execute tasks independently, without human input for each step – are on their way to becoming commonplace. And MCP is the infrastructure that makes it possible.

Imagine an AI agent that automatically every Monday morning:

1. Fetches all open tickets in your support system

2. Prioritizes them based on customer SLA and ticket age

3. Sends a summary to the responsible team lead

4. Flags the tickets at risk of breaking SLA before the day ends

That's not science fiction. That's what MCP enables with the systems you already have.

Organizations building MCP infrastructure today aren't just building a point integration. They're building a platform for future AI automation – a connection point that can be coupled to new AI capabilities as they develop. That's the difference between buying an AI license and building AI capability.

Ready to Connect Your Systems to AI?

We've completed four MCP implementations and can list exactly what each week delivered. It's not a concept we sell – it's something we've actually done.

If you're curious about whether MCP fits your systems, we offer a free mapping session:

• A one-hour review of your existing APIs and systems

• An assessment of which use cases provide the fastest value

• A rough timeline and cost estimate

No obligations. No sales pitch.

Get in touch via smartsource.eu or contact Mattias Jacobsson directly on LinkedIn.

Smart Source AB – We build AI integrations and digital solutions for Swedish companies and organizations. Malmö, Sweden.