There’s a distinction in professional AI tooling that’s easy to miss: the difference between the data layer and the knowledge layer.

The data layer is raw facts. Property records. Market transactions. Public filings. Interest rates. It’s what the system knows about the world — stored, indexed, retrievable. The data layer is increasingly served by MCP servers because that’s the natural shape of the problem: an external system has data, and an AI agent needs to query it on demand.

The knowledge layer is different. It’s not facts — it’s how to reason about facts. How to model a deal. What variables matter in a given analysis. What a result means in context. This is domain expertise encoded into something an AI can use.

When a domain expert builds an AI tool, they’re often building the knowledge layer without naming it. The value isn’t that their tool knows more facts — it’s that it knows how to think about the facts correctly. It applies the right frameworks. It flags the right exceptions. It knows which numbers to stress-test.

The data layer and the knowledge layer feel similar from the outside. Both are things you query. Both return information. But they solve different problems.

If an AI agent lacks data, it gives you wrong answers because it doesn’t know what happened. If an AI agent lacks knowledge, it gives you wrong answers because it doesn’t know what to do with what happened.

Most professional domains need both. The data layer without the knowledge layer means you have numbers you don’t know how to interpret. The knowledge layer without data means you have frameworks but nothing to run them on.

The interesting design question for professional AI tools is: which layer are you building? And which layers are you assuming the user will get elsewhere?

Tools that try to own both layers tend to become platforms. Tools that own one layer and integrate cleanly with the other tend to stay composable. The MCP protocol is built for the latter model — each server owns a layer, agents compose them.

The question “what layer does this tool sit in?” is more useful than “what does this tool do?” because it tells you what it doesn’t do — and what the user still needs to bring. +++