Anthropic's data team describes how it answers most internal business analytics questions with Claude Code, using skills that define how each headline metric is calculated, with human sign-off on anything going to leadership.
- First reported
- claude.com · 3 Jun 2026
- Added
- 13 Oct 2026
- Source checked
- 13 Oct 2026
What they did
Each data domain has skills, folders of Markdown that Claude Code reads on demand, describing tables, definitions and how headline KPIs are calculated. A standing check compares each domain's top KPIs daily against the approved dashboard, and every answer carries a footer with its source, data freshness and owner. Skill changes are now part of most data model updates.
What happened
The company says about 95% of its business analytics queries are now handled this way at roughly 95% accuracy overall, versus no more than 21% on its tests before skills were added.
Keep in mind
Accuracy drifted from about 95% to about 65% in a month when skills were not maintained, and the team says plausible but wrong answers are not fully solved. Leadership-bound answers still need explicit human sign-off.
Try it with your agent
Read the metric definitions in <your metrics skill folder> before answering. Then answer: <your business question>, using only the tables named there. Show the query you ran, the result, and a footer with the data source, how fresh the data is and who owns the metric. If the definition is missing or ambiguous, stop and tell me instead of guessing.
Replace the <angle brackets> with your own details.
Original source
claude.comVendor case studyPublished by the company that makes the agent. Read it as a claim.
