Anthropic’s data team hits 95% automation with Claude

The company’s data team reports that 95% of business queries are now handled by Claude agents with around 95% accuracy. The main error drivers are term ambiguity, data staleness, and information retrieval issues.

To tackle these issues, Anthropic built an agentic stack centered on skills. These are Markdown instruction sets guiding the agent on what to search, which filters to apply, and how to validate results. Without these skills, model accuracy dropped to 21%.

A key finding: simply granting the model access to the full SQL query history didn’t improve accuracy. Curation works better: creating canonical datasets and encoding business logic as code that updates alongside data models.