What is MCP?

The Model Context Protocol is an open standard (launched by Anthropic in late 2024) that lets AI models directly access tools and data in your systems, without intermediaries.

Think of it this way:

Before: AI asks → Your application interprets → Your application talks to the ERP → Your application answers the AI
Now with MCP: AI asks → AI accesses the ERP directly through the protocol → AI answers

It's safer, faster, and more reliable.

Why MCP changes the game for SMEs

1. Direct connection = Trust

Without MCP, you have to:

Write "glue" code that translates between the AI and your ERP
Keep it updated every time the ERP changes
Debug when the AI misunderstands the data structure

With MCP, the protocol handles it. Your ERP exposes its capabilities, the AI understands them, and that's it.

2. Security by default

MCP includes:

Authentication and authorization — the AI only accesses what you allowed
Audit logs — every action the AI takes is recorded
Rate limiting and sandboxing — protection against bad decisions or infinite loops

You don't need complex firewalls. The protocol handles it.

3. Model-agnostic

MCP doesn't depend on OpenAI, Anthropic, or any vendor. It works with any model that implements the standard.

Implication: If prices change, policies change, or you feel another model is better, you switch. Your data architecture stays the same.

Concrete use cases

Case 1: Purchase order automation

Before (without MCP):

Customer: "I need 200 units of product X"
↓
Custom API → interprets → updates ERP
↓
20 minutes, risk of manual errors

With MCP:

AI reads the customer's order
↓
AI directly accesses:
  - Inventory (in real time)
  - Suppliers and prices
  - Your SME's purchasing policies
↓
AI generates the PO, updates status, sends confirmation
↓
2 minutes, no manual touches

Case 2: Profitability analysis

Your AI can now:

Query revenue per customer (from your ERP)
Query operating costs (from your accounting system)
Calculate real margin per business line
Automatically identify unprofitable customers
Suggest actions (raise prices, renegotiate terms, reduce service)

All in 10 seconds. Today this takes hours.

Case 3: Customer service

Your AI can:

See the customer's complete history (orders, payments, previous support)
Access return policies, warranties, current promotions
Answer specific questions without saying "I don't know"
Propose solutions within the limits you define

Result: Happier customers, fewer "please hold to speak with a human".

How to implement MCP

Step 1: Systems audit (1–2 weeks)

Map out:

What data is the AI using today?
What data would it need to be more useful?
What systems do you have (ERP, CRM, accounting, HR)?

Step 2: Choose an MCP tool (1 week)

Current options:

Anthropic Claude API — native MCP, best integration
OpenAI — MCP available in their API (Responses API / Agents SDK)
Orchestration tools — LangChain, LlamaIndex, etc.

Step 3: Connect your ERP (2–4 weeks)

Document the data you will expose
Configure authentication
Write the "handlers" that translate AI queries into ERP actions
Test and debug

Step 4: Monitoring and adjustment (ongoing)

Review what's working
Refine access policies
Expand to more processes

Important cautions

1. Not everything should be automated

MCP lets the AI access your data. It doesn't mean it should make decisions on everything.

Good candidates: Reading data, action proposals, information consolidation

Bad candidates: Price changes without approval, critical financial decisions, customer communications without review

2. Documentation is critical

The AI accesses complex data. If the meaning of each field isn't well documented, it will make mistakes.

3. Governance

You need someone responsible for:

What data the AI exposes
Who can see the logs
How you escalate when something goes wrong

The fundamental shift

Before MCP, every AI was an "isolated box" that did specific things. You had to write manual integrations for every use case.

With MCP, your AI becomes a business layer. It can talk to any system you have, respecting security and policies.

That means:

Less glue code
Fewer bugs
Less maintenance
More value

Next step

If your SME:

Has a structured ERP (SAP, Odoo, NetSuite, Navasoft, etc.)
Is open to exploring AI agents
Wants to understand how MCP could apply to your specific business

Contact us for an evaluation: 90 minutes, technical mapping, no commitment.

We don't promise "AI that replaces your team". We promise tools that expand what your team can do.

About the author

Teo García Pascual is CEO and CTO of Nodalix Consulting. As the company's sole founder, he designs the integration architecture (MCP Protocol, ERP/CRM) of every AI portal he deploys. LinkedIn · contacto@nodalix.es

Model Context Protocol: how to connect AI directly to your ERP