Data Modeling for the Agentic Era: Semantics, Speed, and Stewardship (www.ssp.sh)

AI-driven analytics are reaching a turning point: LLMs can generate SQL and surface insights quickly, but without structured foundations they hallucinate, misinterpret business logic, or leak sensitive data. The article argues that reliable "agentic" data modeling rests on three pillars—semantics, speed, and stewardship. Semantics means a machine-readable metrics layer (e.g., One-Big-Table/OLAP targets, SQL-expressed "metrics SQL" and YAML semantic definitions) so agents operate against curated, unambiguous KPIs. Speed requires sub-second analytics (think ClickHouse, DuckDB) so humans can instantly verify outputs. Stewardship supplies configurable guardrails—ai_instructions.md, Model Context Protocol (MCP), platform integrations like Rill—to enforce type checks, date sanity, PII rules and human-in-the-loop confirmations. For the AI/ML community this reframes how LLMs should be used in analytics: not as autonomous replacements but as augmenting tools that extend declarative SQL/YAML semantics to create or query metrics safely and fast. Technically, it emphasizes making semantics portable and version-controlled, pushing SQL as the precise lingua franca for agents, and integrating governance directly into agent workflows. The result: faster iteration, fewer hallucinations, and a scalable model where more human stewards validate higher volumes of machine-assisted insights—transforming traditional BI practices for the agentic era.