What is MCP, and why does it matter for your business?

MCP stands for Model Context Protocol. It is an open standard, released by Anthropic in November 2024, and it solves a dull but expensive problem: how an AI gets hold of what the business actually knows. Instead of building a new connector for every combination of AI tool and business system, you build once and plug in. The comparison people use is USB-C. Anthropic, OpenAI, Google and Microsoft all run on it today.

The bottleneck is not the model

The first thing we do at a new client is draw a map. Every system is on it: CRM, ERP, file storage, mail, and whatever industry software has piled up over the years. Then we draw a box for the AI. In most companies, no lines run from that box.

That gap is where the work sits. Every system the AI needs to see requires its own connection, and before MCP each connection was a piece of code someone had to write, maintain and debug when it stopped working.

Do the arithmetic. Five AI tools and ten internal systems means up to fifty connections to keep alive. BCG puts it as integration complexity rising quadratically without a shared standard. Every new tool makes the problem worse faster than you expect.

What MCP does

MCP puts one shared plug in between. The system gets an MCP server, and the AI is the client asking it questions. Build an MCP server for your CRM once, and any MCP compatible AI can use it afterwards. Same trade as USB-C: the standard handles the socket, so you stop solving the same job from scratch.

A system can expose three things through MCP:

• Data the AI can look up.
• Actions it is allowed to take, like updating a record or passing a request onward.
• Templates for the workflows that keep coming back.

The standard is open source, and since December 2025 it has been governed by the Linux Foundation, not by Anthropic alone. OpenAI, Google, AWS, Microsoft and Bloomberg are behind the Agentic AI Foundation that stewards it now. That makes it infrastructure rather than one company's product.

How fast it became the default

The timeline is worth knowing, because it tells you where we are.

MCP arrived in November 2024. By March 2025, OpenAI had taken it into ChatGPT Desktop, their Agents SDK and the Responses API. In April, Google confirmed Gemini support. In May, VS Code and GitHub Copilot shipped it natively.

Bloomberg adopted it across the house, and Amazon rolled MCP support out to most of its internal tools. Gartner expects 40 percent of enterprise applications to carry AI agents by the end of 2026, up from under 5 percent in 2025. MCP is the layer that lets those agents see real business data. At this point it is no longer a choice between standards. It is the thing you assume.

What it means if you are putting AI to work

MCP does not change what an AI can figure out. It changes whether the AI can see enough of your business to do it.

An agent with no access to your data stays in demo mode. Customer records, internal documents, order history: all invisible to it. It looks impressive until you ask it for something real. MCP is what moves it from demo to operation.

For us, MCP is the layer that connects our Coworkers to a client's systems. Without the standard, every setup has to be built from scratch. With it, the distance between "this AI understands our workflow" and "this AI runs inside our workflow" gets short.

If you are a company of 50 to 200 people, you are unlikely to build MCP servers yourself. The question is whether the tools and the people you buy from build on MCP. Ask the vendor straight out: do you support MCP? A no, or a blank look that says they do not know the word, tells you something you did not know before.

One thing not to skip. MCP is a connectivity standard, not a security model. Early implementations had real holes, and proper authorisation with OAuth 2.1 only entered the standard in March 2025, months after launch.

So give every connection the narrowest permissions you can, require logging of what the AI touches, and keep a human in the loop when an action has consequences. The specification says a human "should" be involved. Read that as must.

If you are still working out where AI makes sense for you at all, before anyone talks about plugs and standards, start here: Where should a small business start with AI?