PGDay Lowlands2026

Every team wants to connect an AI agent to their database. Nobody trusts it in production. The problem isn't text-to-SQL (that part is mostly solved). The problem is everything else: the agent doesn't know which tables are deprecated, which columns contain PII, or why three tables named users_* exist. And when something goes wrong, you have no idea which agent ran what query, or why. In this talk, I'll show you how to build a governed AI data layer for Postgres using pg-airman-mcp, an open-source MCP server. Instead of dumping schemas and hoping for the best, we'll use three capabilities that change the game: First, a semantic knowledge base that lets agents search your schema by meaning, not by name. The agent asks "where do we store customer emails?" and gets the right table, along with its PII annotations and deprecation status, without scanning every table in the catalog. Second, semantic aliases: named, reusable queries that your team curates and the agent discovers through natural language. Think of them as a library of blessed queries your agent can find and execute, with built-in role-switching for access control. Third, purpose-tagged connections that stamp every database session with the agent's intent (credit-scoring, fraud-detection, analytics-pipeline) so your DBA can see exactly who is asking what, directly in database logs. You'll walk out with a clear architecture for giving AI agents database access you'd actually approve for production.