Anthropic automates internal business analytics with Claude

Self-service analytics automation is not a code generation problem, but a database governance and context curation challenge that raw LLM retrieval cannot solve alone.

• –**Retrieval failure**: Giving agents raw search access to thousands of prior query logs yields under a 1% improvement in accuracy, proving that unstructured history is no substitute for curated, human-managed domain documentation.
• –**Quality at a cost**: Stacking adversarial review sub-agents raises accuracy by 6% but incurs a massive 72% increase in latency and 32% higher token usage.
• –**Preventing documentation decay**: To stop knowledge going stale as databases change, teams must colocate skill documentation in the data transform repository, enforcing doc updates as part of the CI pipeline.
• –**Semantic layer as a blocker**: Restricting agents to a semantic metrics layer first, rather than raw table querying, drastically limits wrong-answer modes caused by complex table joins.