What is an MCP (Model Context Protocol)? How can you prepare your company for custom agents with your own MCP?

If in my last article I talked about personalized AI agents as the new era of business automation, today it’s time to dive into the infrastructure that makes these agents actually work: the Model Context Protocol (MCP).

Because understanding what agents can do for our businesses is one thing, and quite another is knowing how to connect them securely, scalably, and efficiently to your systems. With full control over the information that is shared and what you want to operate within your own infrastructure. That’s where the MCP comes in.

MCP: The bridge between your data and artificial intelligence

The Model Context Protocol is an open standard developed by Anthropic that allows AI models (Claude, GPT, Gemini, etc.) to securely and standardizedly connect to your own data sources, such as databases, CRMs, ERPs, applications or internal systems, APIs, and any other information source your business wants to use as context for its agent, giving it the necessary framework for reasoning, decision-making, and thus improving efficiency and autonomy.

Think of it as the “standardized common language” that allows any AI agent to read, interpret, and act on your company’s information, no matter where it is stored or what format it’s in.

Why is MCP revolutionary?

Before MCP, connecting an LLM to enterprise systems involved custom developments for each case: specific APIs, proprietary connectors, ad-hoc integrations that needed updates with every change. It was costly, fragile, and hard to maintain. Not to mention having to dump sensitive information into external platforms.

With MCP, you standardize that connection. You build the connector once, and any compatible agent can use it. It’s like going from having 20 different chargers for your devices to having a universal USB-C.

Architecture and how an MCP works

A well-implemented MCP consists of several components, each fulfilling a specific function for generating business context:

1. MCP Servers

These are connectors specific to each data source. For example:

• An MCP server for your CRM (Salesforce, HubSpot, etc.)
• Another for your ERP (SAP, Oracle, Dynamics)
• One for your product catalog (PIM)
• Another for your internal documentation (Confluence, SharePoint)
• One for your ticketing systems (Jira, Zendesk)

Each MCP server “translates” that data source’s structure into a format any AI model can understand.

2. Security and permissions layer

Here is where you define who can access what information:

• Roles and permissions per user/agent
• Access policies for sensitive data (red lines or supervisor validation)
• Logging and auditing of all interactions
• Masking of critical data (PII, financial information)

This layer ensures that an AI agent never accesses information the user operating it shouldn’t see.

3. Context orchestrator

The component that decides what information the agent needs to complete a specific task:

• Prioritizes relevant data sources
• Optimizes the volume of information transferred (models have context limits)
• Caches frequently used information to save costs/tokens
• Manages dependencies between different sources

4. Private infrastructure

And this is critical: your MCP lives in your infrastructure. Not in the cloud of any model or commercial tool, not on OpenAI servers or any other model’s infrastructure, and not in shared infrastructure.

It is on your servers, under your control, complying with your compliance policies.

The data never leaves your security perimeter. The AI model accesses it through the MCP, but the information stays within your ecosystem.

This is why the importance lies, and why we at Azurally recommend creating our own MCP servers to maximize the potential of the agents you can/we can develop.

Real use cases where MCP makes a difference

Case 1: Commercial intelligence agent

Imagine an agent that needs to analyze sales opportunities. Without an MCP, it would have to:

• Make separate calls to Salesforce
• Another query to your pricing system or product catalog
• Search for competitor information in another system
• Check customer history in the ERP
• Access product documentation

With MCP, the agent makes a single request: “Give me the complete context for client X to evaluate opportunity Y.” The MCP orchestrates all those queries, consolidates the information, and presents it to the agent in a structured way.

Result: what previously took 15 minutes of a salesperson checking 5 different systems now takes 30 seconds.

Case 2: Technical support agent

A customer reports an error. The agent needs:

• History of previous tickets (Zendesk)
• System logs (internal databases)
• Product version the customer is using (CRM)
• Relevant technical documentation (Confluence)
• Similar resolved cases previously (Knowledge Base)

Without MCP: 30 minutes of an engineer jumping between systems. With MCP: 45 seconds where the agent has ALL necessary context and can propose a solution or informed escalation.

Case 3: Executive reporting agent

Every Monday, the management team needs a dashboard with:

• Weekly sales (ERP)
• Sales pipeline (CRM)
• Marketing metrics (Google Analytics + Meta Ads)
• Financial status (accounting system)
• Operational KPIs (multiple sources)

Without MCP: 4–6 hours of an analyst manually consolidating data, or using a pre-connected dashboard in a visualization tool like PowerBI or Looker. With MCP: The agent generates the complete report, with analysis and recommendations, not just displaying raw data, in under 5 minutes.

Common mistakes to avoid when setting up an MCP

Some of the most common mistakes observed in clients or MCP implementation projects are:

Error 1: Thinking MCP is “just a technical infrastructure” It isn’t. It is a transformation project that requires defining and understanding business processes, information flows, and the needs of end users who will operate the agent or use AI enriched with business context. If approached as just “set up servers,” the project will fail.

Error 2: Not defining governance from the start Who decides which data is accessible? How are new connectors approved? Who audits usage? Without clear governance, it ends up being a permissions mess or worse, with sensitive information exposed or accessible to anyone.

Error 3: Trying to connect everything at once and contextualize the entire company.

Start with 2–3 critical data sources for a specific use case. Validate that it works. Then scale. Projects trying to connect 15 systems simultaneously stall. Trying to contextualize the whole company makes governance cumbersome and ineffective—less is more; break the project into manageable, reviewable parts.

Error 4: Underestimating the data cleaning effort—this is where we always encounter issues, such as duplicate entries in the CRM, incorrectly labeled or filled fields in the ERP, or inconsistent legacy data. MCP will expose all these problems. It’s crucial to have sanitized and validated data before starting an MCP project.

Error 5: Not considering cultural change, and its impact on teams. With the speed of change and new ways of working, teams need to understand that agents can now access “their” information. They need training, and agents should complement, not disrupt, daily work. Clear documentation and seeing the value before adoption are key.

The ROI of having your own MCP

One of the questions we are always asked when implementing an MCP is the ROI it will have on operations or company strategy. Key points include:

Development time for new agents: From months/weeks to weeks/days.

Each new agent you deploy has immediate access to all connected data sources. No need to rebuild integrations, allowing context from the start.

Maintenance cost reduction: 60–70% of system updates are handled in the corresponding MCP server, not in each individual agent.

A Salesforce update doesn’t break 10 different agents. This is crucial for scalability.

Improved decision quality: Studies show that over 80% of agents report better-documented decisions, feeling better informed and equipped for decision-making.

Because agents have access to full context, not fragments or silos as is typical in companies—they see the full picture.

Regulatory compliance: Drastic reduction of risk. All interactions are logged, auditable, and subject to centralized policies. GDPR, HIPAA, SOC2… whatever the regulatory framework, MCP helps comply.

Why now is the time to build an MCP

The data is clear: according to Gartner, in 2025, 40% of AI initiatives will fail due to data integration and validity issues—not lack of good models, but inability to effectively connect them to clean real information.

MCP solves exactly this problem. Doing so gives you a clear competitive advantage:

1. Infrastructure advantage: While competitors struggle with custom integrations, you already have a scalable platform.
2. Speed advantage: Deploy agents in days, not months.
3. Learning advantage: The sooner you start, the sooner your teams enjoy an agentic ecosystem.
4. Data advantage: A well-built MCP is a golden opportunity to clean, structure, and enrich company data for future integrations.

How Azurally helps companies build their MCP

At Azurally, we have developed a methodology applied with our clients or when a company wants to start their MCP:

Phase 1: Discovery & Design

We conduct deep analysis workshops where we:

• Map all relevant data sources
• Identify priority use cases for agents
• Define governance and security policies
• Design the specific architecture of your MCP
• Establish an implementation roadmap

This is not generic consulting. We use Design Thinking applied to AI infrastructure to understand not only which data is available but how it is used in daily operations.

Phase 2: MVP Implementation

We build a functional MCP with:

• 2–3 connectors to the most critical systems
• A basic but robust security layer
• A pilot agent demonstrating immediate value
• Complete documentation and training for the different teams involved

The goal: within 2 months, you have your first agent running in production, generating real value.

Phase 3: Scale & Optimize (continuous)

Once the MVP is validated:

• Add new connectors progressively
• Refine access policies based on real usage
• Optimize performance based on observed patterns
• Deploy additional agents on the same infrastructure

This phase is where ROI skyrockets: each new agent added has minimal marginal cost because the infrastructure already exists.

Our differentiating advantage

1. Multisector experience
2. Hybrid approach: Combining technical depth (system architecture, security, DevOps) with business understanding (processes, KPIs, adoption).
3. Vendor-agnostic: We work with different models depending on the problem… MCP doesn’t tie you to a specific LLM provider.
4. Sovereign infrastructure: Everything on client servers and infrastructure, under your control, complying with European data protection regulations.
5. Knowledge transfer: We don’t create dependency. We train your team to manage and evolve the MCP autonomously.

Is your company ready for an MCP?

Before starting this project, consider some questions or points that condition the launch:

• Multiple data sources that need connecting (if everything is in one system, maybe you don’t need an MCP)
• Identified specific use cases where agents would generate clear value (measurable ROI)
• Executive sponsor who understands strategic importance and can provide access/resources
• IT team willing to learn new paradigms
• Resources (time and money) available for the initial 2–3 months
• Compliance and industry security requirements identified

If you check at least 4 of 6, your company is ready. An MCP can truly provide real value in daily operations and strategy.

Companies without an MCP will fall behind

It sounds dramatic, but Forrester predicts that by 2027, companies without agentic infrastructure will have a 30–40% competitive disadvantage in operational efficiency.

The difference is not having access to GPT, Claude, or Gemini (everyone has that). The difference is having agents that deeply understand the business because they have structured, secure, real-time access to all your company’s information.

And this can only be achieved with a well-built and structured MCP.

Ready to take the step?

If after reading this you think “my company needs this, but I don’t know where to start,” good news: this is exactly why our Think Tank Workshops exist.

In a half-day intensive session:

• We audit your current tech stack
• Identify high-impact opportunities for agents
• Design the blueprint for your specific MCP
• Estimate roadmap, resources, and expected ROI
• Provide actionable documentation to present to your executive committee

You’ll leave with a clear plan, not generic AI theory.

Contact us at Azurally and let’s start building the infrastructure that will differentiate your company over the next 3 years.

Because MCP is not “just another tech project.” It’s the foundation for competitive advantage for the next decade.

Is your company already exploring MCPs? Or are you still figuring out what they are and what they do?

I’d love to know your current state and specific challenges. Comments are open.

#MCP #ModelContextProtocol #ArtificialIntelligence #AIAgent #DigitalTransformation #Automation #DataIntegration #AIInfrastructure #Azurally #ROIUP #ROIUPGroup