Semantic Layers Lift AI Agent Analytics Accuracy to 95%

My notes

Summary

A semantic layer is a translation layer sitting between a data warehouse and everything that reads from it: dashboards, APIs, and AI agents. It defines each business metric exactly once. Anthropic’s data science team reported that adding one raised the accuracy of AI-agent-generated analytics answers from 21% to 95%.

Key Insight

• Without a semantic layer, every AI agent (and every human team) independently interprets what a metric like “active users” or “revenue” means, producing inconsistent, often wrong answers.
• The fix is not a smarter model or better prompting, it is structural: a single, human-curated definition of each metric that the agent is required to reference before writing any SQL or generating an answer.
• The headline number (21% to 95%) comes from Anthropic’s own internal self-service analytics system, where the agent was structurally required to check the semantic layer first rather than freelancing a SQL query.

Three concrete benefits fall out of centralizing definitions:

1. Everyone (dashboards, APIs, agents) works off the same definition, so there is no drift between teams.
2. Changing a metric’s definition is a one-time edit in the semantic layer, not a hunt-and-fix across every consumer.
3. The agent does not have to guess. It looks up a single source of truth instead of inferring meaning from column names or past examples.