Anthropic's Internal Claude Tag Deployment Answers 75% of Slack Data Questions Unprompted

Anthropic published an account of a year running Claude Tag — its Slack-based agent — as a self-service data-analytics tool internally, and the numbers are concrete rather than aspirational: a governed semantic layer, skill files, and evaluation suites brought the system to roughly 95% accuracy on data questions, and in one internal data channel Claude Tag now answers more than 75% of questions posted, typically within a minute or two, often without being explicitly tagged.

The five learnings Anthropic shares read like an operating manual for anyone building an internal analytics agent: treat skill files as continuously refreshed content rather than static documentation; give the agent skills beyond raw querying — forecasting, cohort analysis, funnel analysis, charting, analytical writing; connect it to internal knowledge indexes and business context, not just the data warehouse; permission its service account deliberately through a five-layered security approach; and instrument every answer with structured logging and telemetry. The recommended rollout sequence is permissions first, distribution second, telemetry from day one, then knowledge indexing, then analytics skills.

The deployment detail lands the same week Anthropic shipped a Claude Tag update that gives the agent broader channel context before deciding whether to jump into a conversation — reading the surrounding thread, memory, and standing instructions rather than evaluating messages in isolation. Anthropic says that added context alone made Claude Tag roughly 30% better at judging when, and when not, to respond — a workplace-agent problem that’s arguably harder than the accuracy one, since an agent that answers correctly but interrupts constantly gets muted by its users.

Together the two posts are a rare paired data point: not just a vendor’s feature announcement, but the vendor’s own year-long usage numbers for that feature category, plus a same-week fix for the most obvious failure mode. For consulting engagements evaluating conversational agents for internal analytics, Anthropic’s own 95%-accuracy, 75%-answered figures are a real benchmark to hold other vendors’ claims against.