What is MCP — the protocol connecting AI to company data, without jargon

If you have read anything about AI in business over the past year, you have run into the acronym MCP. Usually presented as though everyone already knows what it means.

Below is the explanation without jargon — and, more usefully, the answer to what it actually changes for a company of 11 to 200 people.

The problem MCP solves#

A language model on its own knows nothing about your company. For it to know, someone has to hand it the data.

Until 2025 that worked like this: a developer wrote an integration connecting a specific model to a specific system. Want to connect a second model — write a second integration. Want to add a third system — two more.

With three models and four systems that is twelve integrations to write and maintain. Any change on either side means fixes.

For a large company with a development team that is a cost. For a company of 11–200 people with no IT department it is an insurmountable barrier.

What MCP does#

Model Context Protocol is an open standard describing how a language model asks an external system for data and how it calls that system's functions.

Instead of twelve integrations: each system exposes its data once in that standard, each model learns to understand it once. After that, everything connects to everything.

The closest analogy is the USB port. Before it existed, every device had its own connector and its own cable. Once it arrived, a manufacturer builds one socket and works with the rest of the world.

Why it became the standard now#

In 2026, MCP is supported by every major work-platform vendor. That means two things.

First: MCP itself has stopped being a differentiator. If a vendor leads with "we have MCP", they are telling you they have what everyone has. Worth pushing further — on permissions, on the scope of exposed functions, on whether MCP allows only reading or also acting.

Second: you can simply assume it. Choosing a tool in 2026, the absence of MCP is a warning sign; its presence is not a distinguishing one.

What it changes in practice#

Three concrete things.

You ask about company data where you already are#

Instead of logging into a system to check a status, you ask the assistant you already have open. The answer comes from current data, not the model's memory.

The effect is most visible outside the office: "what did we do at this customer last time" asked from a phone on the way to site, instead of scrolling a job list one-handed.

You stop being locked to one vendor#

The model market shifts every quarter. Without MCP, switching vendor means rewriting integrations, so in practice it does not happen — you stay with whatever you chose a year ago.

With MCP, switching model is a setting. That is insurance against a price rise and against your vendor falling behind the rest of the market.

Agents can act, not just answer#

MCP describes not only reading data but also invoking a system's functions. A model can therefore create a job or change a status — as far as the system permits.

And that is the point where you have to ask about control. We cover those questions in the piece on human approval for AI actions.

The security questions to ask#

The protocol itself describes how data is exchanged. It does not determine who can see what — that depends on how the system on the other side implemented it.

Four questions for your vendor:

Does the model inherit the permissions of the person asking? This is the most important one. If the integration runs with administrator permissions, every person asking gets answers from the full dataset regardless of what they should see.

Can MCP only read, or also write? And can that be separated per application?

How does authorisation work and can it be revoked? Connecting an external application should behave like signing in with a Google account: you see a list of what you have connected and disconnect in one click.

Is there a query log? Without a record of what was asked and which data was used, you cannot reconstruct events when a question arises.

What MCP does not solve#

Worth having realistic expectations.

It does not tidy your data. A model connected to a mess returns a mess, only faster and in a more confident tone.

It does not replace thought-through permissions. If everyone in the system sees everything, MCP will not fix that — it will expose that state to the model.

It does not stop models being wrong. It gives them access to current data, which limits guessing, but does not eliminate it.

Summary#

MCP is a common language between AI models and company systems. It removed the need for small companies to write integrations and freed them from lock-in to a single model vendor. In 2026 it is a standard, so the question is not "does the vendor have it" but "how exactly did they implement it" — and the answer comes down to permissions.

The wider context is in the guide on connecting AI to company data.