Most AI Users Have Never Heard of MCP — and It's Not Their Fault

In developer circles, MCP is settled infrastructure. The Model Context Protocol — the standard way to let an AI assistant reach your files, your database, your ticket tracker, your calendar — went from roughly 100,000 SDK downloads at launch to about 97 million a month by March 2026.

Outside those circles, it is invisible. Not disputed, not resisted: unheard of. And there is no survey that can tell you by how much, because nobody has asked.

The half we can measure

Stacklok's State of MCP in Software 2026 surveyed software organisations and found adoption well past the experiment stage.

Within those organisations the user base is exactly who you'd guess: 80% software developers, 68% data analysts and scientists. And 72% of current adopters expect their MCP usage to grow over the coming year.

Anthropic's Economic Index points the same direction from usage data rather than surveys: agentic sessions — where the model is using tools rather than chatting — are on average more automated than chat sessions, and coding accounts for roughly half of API traffic. Tool-using AI is real, in production, and overwhelmingly technical.

The half nobody has measured

Now the other side. Around 52% of employees use AI in their role (Gallup, Q2 2026) and 49% of U.S. adults use a chatbot (Pew, February 2026). Software developers are on the order of 1–2% of the workforce.

So the overwhelming majority of people using AI at work are not developers, are not represented in any MCP survey, and are not counted in the 97 million downloads. What do they know about connecting AI to their own systems?

No major 2026 survey asks a non-technical worker whether they know their assistant can be connected to their own tools. The question has not been put.

We looked for it. Pew asks which chatbots people use and why non-users abstain. Gallup asks frequency. SHRM and the Census Bureau ask which tasks. McKinsey asks executives about functions and EBIT. Not one of them asks whether the respondent has ever connected a tool, an integration, or a data source to the assistant they use daily.

Why the gap is structural, not a failure of curiosity

It's tempting to read this as people not bothering to learn. The numbers say otherwise — the path to learning mostly doesn't exist.

• Most people were never taught anything. 35% of employees have had no AI training of any kind; 41% say their employer provided nothing at all. Of those who were trained, 85% say they can't apply it to their actual job.
• Self-teaching has a ceiling. People learned AI from the consumer product, in a chat box, on their own time. Nothing in that experience suggests the tool can reach outside the chat box.
• The vocabulary is engineering vocabulary. "Server", "protocol", "client", "transport". A marketing manager who would immediately understand "connect it to your CRM" bounces off the same idea labelled MCP.
• Setup is still developer-shaped for anything beyond the handful of one-click connectors shipped in consumer apps — config files, credentials, permissions.
• Their employer probably hasn't scaled it either. McKinsey found no business function above 10% scaled deployment of AI agents, and nearly two thirds of organisations not yet scaling at all.

What the gap actually costs

This is the mechanism behind the sector's most-repeated puzzle: adoption is near-universal and measurable business impact isn't. Only 39% of organisations report any EBIT impact from AI, and the small group that does is 2.8× more likely to have redesigned workflows around AI rather than bolting it onto existing ones.

You cannot redesign a workflow around a tool that can't see your work. A chatbot that knows nothing about your systems can draft, summarise and brainstorm — which is precisely the task list every 2026 survey reports. The task list isn't evidence of unambitious users. It's the complete set of things you can do with a disconnected model.

There are early signs of movement. Microsoft's Work Trend Index attributes 28% of workplace AI activity to decision-making rather than content generation, and Anthropic reports the top 10 tasks shrinking as a share of conversations. The ceiling is lifting. It is lifting slowly, and unevenly, and mostly where someone connected something.

So we're asking

The AI Usage Report survey asks the question the big trackers skip: not only which tools you use and for what, but whether you've ever connected one to your own systems, whether you've heard of the thing that makes that possible, and what stopped you if you tried. It branches by profession, so the answers can be read per job rather than averaged into a number nobody recognises.

If the result is that awareness outside engineering is near zero, that's worth publishing. If it's higher than we expect, that's worth publishing too — and it would be the first published figure of its kind. For the numbers that do exist today, see AI usage statistics 2026.