Challenges in the AI-Driven Infrastructure Management Era

Abstract

Tool-augmented Large Language Models (LLMs) can support infrastructure tasks, but production adoption requires more than tool access. We report an industrial experience deploying Model Context Protocol (MCP) servers for Trento (SAP/HA monitoring) and Uyuni (fleet management), where infrastructure capabilities are exposed as typed tools for machine clients. From these deployments, we identify three governance challenges that become critical once LLMs interact with operational APIs at machine speed. Our main claim is practical: MCP is a useful integration layer, but production use depends on treating exposed tools as governed API products with explicit limits, reviewable actions, and enforceable control points.