Published on June 3, 2026 on the Anthropic blog, this engineering retrospective from the Data Science & Data Engineering team (Chen Chang, Clement Peng, Justin Leder, Johanne Jiao, Josh Cherry) recounts how Anthropic made its analytics self-service with Claude: 95% of business queries automated, ~95% accuracy in aggregate (up to ~99% in certain domains).

The starting point is that analytics is not code: "there's often only a single correct answer using a single correct source". The challenge is not generative creativity but the ability to map a question to precise, up-to-date entities in the data model. Three failure modes threaten this: concept↔entity ambiguity (what are "active users"? do you exclude fraudsters? which window?), staleness of assets and the agent's knowledge, and retrieval failure — 80% of failed queries nonetheless had the information present in the corpus. Worst of all is silent failure: a wrong, plausible answer used without objection.

The answer is a four-layer "agentic analytics stack". (1) Data foundations: dimensional modeling, canonical datasets "single source-of-truth", metadata treated "as a first-class product", integrity via CI/CD. (2) Sources of truth in decreasing order of trust: a semantic layer that the agent is "structurally required (by skill instruction) to leverage first", then lineage, a query corpus distilled into documents (not raw retrieval), and a business knowledge graph (roadmaps, decision logs, org). (3) Skills — the decisive lever: "without skills … didn't exceed 21% … Adding skills gets these numbers consistently above 95%". They are organized in pairs: a router Knowledge skill (~30 reference files) and an Unbook skill encoding the senior analyst's workflow (clarify, find sources, execute, adversarial review). Maintenance is colocated: a review hook flags any model change that doesn't include a skill modification — ~90% of data PRs now include a skill in the same diff. (4) Validation: threshold-gated offline evals (~90% to authorize an agent), ablation testing, and online safeguards (adversarial review: +6% accuracy but +32% tokens and +72% latency; provenance footers; correction harvesting by scheduled agents).

Two negative results shape the doctrine: giving raw grep access across thousands of SQL files moves accuracy "less than a point" (the bottleneck is structure, not access), and letting the LLM define metrics was "net-negative" — hence the rule: documentation generated, definitions owned by humans. To get started: a handful of canonical datasets, a few dozen evals, a thin knowledge skill.