AI & Automation

MCP for Companies: What the Model Context Protocol Is and When It Is Worth Adopting

November 11, 2026 · 7 min read · Etzal Group Team

A company connects its AI assistant to the ERP so staff can ask about stock. Three months later a second team wants the same data in a different AI tool, and the integration is built again. Then a third. Each connection is custom, each has its own way of handling permissions, and nobody can say which AI tools can reach which systems. The Model Context Protocol, usually shortened to MCP, addresses this problem. It is worth understanding before deciding whether your company needs it.

The problem it addresses

An AI model on its own knows only what it was trained on and what is in the conversation. To be useful at work it needs to read company data and take actions in company systems. Each such link has traditionally been built by hand: this assistant to that system, with its own code.

With several AI applications and several systems, the number of custom links multiplies. Every link has to be maintained, secured and rebuilt when either side changes.

What MCP is

MCP is an open standard, introduced by Anthropic in late 2024 and since supported by other major AI vendors and tool makers, that defines how an AI application talks to an external system.

Two roles:

  • An MCP server wraps a system, such as a database, an ERP or a document store, and describes what it offers.
  • An MCP client sits inside an AI application and connects to servers.

A server can offer three kinds of thing:

  • Tools. Actions the AI can ask to perform, each with a name, a description and defined inputs. For example, "get stock level for item" or "create purchase request".
  • Resources. Data the application can read, such as a document or a record.
  • Prompts. Ready-made instructions for common tasks.

Build the server once and any compatible AI application can use it. People compare it to a standard plug: the appliance and the socket no longer need to be made for each other.

What changes in practice

  • Reuse. One server for the ERP serves the chat assistant, the coding assistant and the internal agent.
  • Portability. Changing the AI model or vendor does not mean rebuilding every integration.
  • A catalogue. The list of servers becomes a list of what AI can reach in your company, which is a useful thing to have in one place.
  • Ready-made servers. Many software vendors now publish MCP servers for their products.

What does not change: the hard part of integration is still deciding what to expose, to whom, with what limits. MCP standardises the plumbing. The design is still yours.

When it is worth adopting

Good reasons:

  • Several AI applications need the same internal systems.
  • You use AI tools that accept MCP servers, and want them to work with your data.
  • You expect to change models or vendors.
  • You want one governed place where AI access to systems is defined.

Not yet necessary:

  • One agent, one system, built by one team. A direct integration is simpler.
  • A system with no interface to wrap. The work is building the interface first.
  • A process that does not need AI at all.

Six security questions

A server that lets an AI act in your systems deserves the scrutiny of any privileged integration.

1. What does it expose? List every tool. A server that can "run any query" or "send any email" is too broad. Prefer narrow tools that do one thing with validated inputs.

2. Whose permissions does it use? A server running with one powerful service account lets every user do everything through the AI. Better: the server acts with the permissions of the user making the request, so the AI can never reach more than that person could.

3. Which actions need a person's approval? Reading is lower risk than writing. For actions that send, change, delete or pay, require confirmation, and enforce it in code, not only in instructions to the model.

4. Where did the server come from? A server downloaded from the internet is code running with access to your systems. Review it as you would any dependency, pin its version and prefer official ones.

5. What can it be tricked into doing? Content the AI reads, such as an email or a web page, may contain instructions. If the same session holds a tool that can send data out, hidden instructions could cause a leak. Be careful about combining tools that read untrusted content with tools that act or export.

6. Is everything logged? Each tool call, with who asked, what inputs and what result. Without logs there is no investigation after an incident.

Designing a good server

The quality of a server is mostly the quality of its tools.

  • Design for the task, not the database. "Find open invoices for customer" is more useful and safer than a general query tool.
  • Describe tools clearly. The model chooses tools by reading their descriptions. Vague descriptions lead to wrong choices.
  • Return only what is needed. Large responses cost money and carry more data than the task requires.
  • Validate inputs. Treat every input from the model as untrusted.
  • Keep the number of tools moderate. Dozens of overlapping tools confuse the model. A few well-named ones work better.
  • Fail with useful messages. "Customer not found, check the customer code" helps the model recover.

Governance

When more than one team builds servers, agree a few rules.

  • A register of approved servers: owner, what it exposes, data classification.
  • A review before a server is used in production, covering the six questions.
  • Versioning, so a changed tool does not silently break the applications that use it.
  • An owner for each server, responsible for its upkeep.
  • A way to switch one off quickly.

A sensible first project

Choose something read-only, valuable and low risk.

  1. Pick one system and one question people ask often. For example, stock availability or order status.
  2. Build a server with two or three read-only tools.
  3. Authenticate as the user. Use the user's own access rights.
  4. Connect it to one AI application used by a small group.
  5. Log every call and review the logs weekly.
  6. Measure: how often it is used, how often the answer was right, how much time it saved.

After a month you will know whether the approach suits your systems, and you will have a pattern for the second server. Add tools that write only after the read-only ones have earned trust, with approval steps.

Common mistakes

  • Wrapping an entire system interface as tools, one tool per endpoint.
  • One all-powerful service account.
  • Installing third-party servers without review.
  • No approval step on actions that change things.
  • No logs.
  • Adopting MCP because it is new, without a second consumer in sight.

The standard is still moving

MCP is young. The specification, the tooling and the security practices around it continue to develop. Check the current documentation before designing, keep servers thin so they are easy to update, and avoid depending on features that are still experimental.

We design and build integrations of this kind, with and without MCP, in our AI Consulting and Enterprise Software work. If you are deciding whether a protocol layer is worth it for your systems, contact us.

Summary

MCP is an open standard that lets AI applications use company systems through servers exposing tools, resources and prompts. It pays off when several AI applications share systems or when you want portability. Treat each server as a privileged integration: narrow tools, the user's own permissions, approval on actions, reviewed sources, care with untrusted content, and full logs. Start with a read-only server for one system and learn from a month of use.

Frequently asked questions

What is the Model Context Protocol in simple terms?

It is an open standard for connecting AI applications to external systems. A system is wrapped in an MCP server that describes what it offers, such as tools to call and data to read. Any AI application that speaks the protocol can then use it, without a custom connection built for that pair.

Do we need MCP to build an AI agent?

No. An agent can call your systems through ordinary interfaces written into its own code. MCP becomes worthwhile when several AI applications need the same systems, when you want to change AI vendors without rebuilding integrations, or when you use off-the-shelf AI tools that accept MCP servers.

Is MCP secure?

The protocol is a way of connecting, and security depends on how each server is built and governed. A server that exposes too much, runs with broad credentials or comes from an unverified source is a risk. Treat every MCP server like any integration with access to your systems: least privilege, authentication per user, logging and review.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.