AI agents can now access APIs directly via function calling and tool use. With reasoning, an OpenAPI spec, and a bit of hope, this approach can already solve a surprising amount. But how far does it really go when dealing with enterprise systems — and where does it hit its limits?

That’s exactly the question we explore. We start by looking at how an AI agent connects directly to an API: Does it find the right endpoints? Does it pass parameters correctly? How reliable are the results?

We then introduce the Model Context Protocol (MCP) as an integration layer and show what concretely changes: standardized discovery instead of trial and error, semantic tool descriptions that give the agent real context, and a protocol that addresses aspects like authentication and authorization.

MCP is no free lunch, though — the integration effort shifts rather than disappears. Implementing a good MCP server is an architectural challenge in its own right. Drawing on our experience from client projects and our own development work, we show when MCP delivers real value and where the direct API route remains the more pragmatic choice.