Why AI assistant schema catalogs break down weeks after launch

When teams build AI assistants to query their databases, they must create a schema catalog — a structured representation of table names, columns, types, and relationships — that sits between the assistant and the database. Most teams build this catalog quickly, often the same weekend they set up the assistant, and it works well initially. Rather than granting the AI direct production access, teams typically build an intermediate layer using approaches such as JSON files, nightly refresh scripts, custom MCP servers, semantic layers, or read-only replicas. While each method works at launch, the catalog tends to decay silently as the underlying database changes, causing quality issues that compound over time. The core challenge is that teams focus heavily on the connection layer — credentials and query costs — while neglecting the knowledge layer, which is where the assistant's actual usefulness lives.