Anthropic Details How It Automates 95% of Analytics With Claude
Anthropic has published a new blog post outlining best practices for automating business analytics with Claude, revealing that internally it now automates 95% of its analytics queries via Claude with roughly 95% aggregation accuracy. Announced on June 3 in a post by Claude Developers (@ClaudeDevs), the article "How Anthropic enables self-service data analytics with Claude" explains data-analysis agent building around three pillars: skills, data foundations, and evaluations. Source
The core argument is that simply connecting an agent to a warehouse for natural-language querying does not deliver accuracy. Anthropic frames analytics automation as "not a code-generation problem, but a context and verification problem." Internally it identified three main failure modes: "Concept <> entity ambiguity," where a user's question fails to map to the right field in the data model; "Data staleness"; and "Retrieval failure." Source
To address these, Anthropic recommends a three-layer stack: "data foundations" with canonical datasets, dimensional modeling, and metadata as a first-class citizen; "sources of truth" built on a semantic layer, lineage, and business context; and "skills" that capture procedural knowledge and require calling the semantic layer. The impact of skills is striking: accuracy of 21% without skills rose to over 95% with them, reaching nearly 99% in certain domains. The approach runs on Claude Code (claude.ai/code), and the Developer docs publish skill authoring best practices, including a BigQuery data-analysis example. Source
Self-service analytics has long struggled with metric inconsistency from sprawling table creation. While LLM agents make natural-language querying feasible, accuracy can degrade due to concept and entity ambiguity or stale data. Unlike the general-purpose agents from OpenAI and Google, Anthropic's distinctive move is publishing concrete governance and evaluation methods grounded in its own internal practice. The full blog includes an appendix of skill templates, and related materials such as The Complete Guide to Building Skill for Claude are also available. Source
The official post drew 386 likes and 17 replies, with related discussion noting that "live data insights with Claude could threaten data-analytics SaaS" and praising "Anthropic's continuous improvements." Existing Claude Code users have called the agent-building approach highly reproducible. At the same time, few posts reporting specific stumbling points had surfaced immediately after publication, with external reproductions and discussion only beginning to spread.