Skip to content

Agentic Analytics Engineering — Anthropic self service analytics pattern (June 2026)

Anthropic self-service analytics pattern (June 2026)

Anthropic published a production account of using Claude for self-service business analytics: How Anthropic enables self-service data analytics with Claude. The headline claim is stark: 95% of business analytics queries automated via Claude, with ~95% aggregate accuracy.

The durable point is not "point Claude at the warehouse". Anthropic says that creates a false sense of precision. Their framing is that analytics accuracy is a context and verification problem, not a code-generation problem.

Failure modes

  1. Concept/entity ambiguity — user language like "active users" must map to one governed definition, grain, population, and exclusion set.
  2. Staleness — schemas, definitions, source tables, and business context rot quickly.
  3. Retrieval failure — the answer may exist somewhere, but unstructured search over dashboards/SQL/query history does not reliably route the agent to it.

Stack pattern

  • Data foundations: dimensional modelling, tests, freshness/completeness checks, governed model tiers, and metadata treated as a first-class product.
  • Sources of truth: semantic layer first; lineage/transformation graph second; query corpus distilled into curated docs rather than dumped raw into retrieval; business context/knowledge graph for ambiguous references.
  • Skills: pairwise skill design — a knowledge/router skill narrows the search space, while a procedural analysis skill encodes the senior analyst workflow.
  • Maintenance: skill docs colocated with transformation code; PR hooks flag model changes that do not update related skill/reference docs. Anthropic reports offline accuracy drifting from ~95% to ~65% over a month before they treated skill maintenance as engineering work.
  • Validation: offline evals, fixed ablation tests, adversarial review, provenance footers, passive monitoring, and correction harvesting.

Implications for Adam

This is directly aligned with the Personal Context Management shift. The winning system is not "more stored context"; it is curated, routed, maintained context with provenance and evals.

For LocalStack/Claude usage, the pack strategy should copy the pattern: compact router instructions, domain-specific reference docs, explicit gotchas, freshness/provenance footers, and correction harvesting from real work sessions.

For analytics engineering positioning, this becomes a strong narrative: the future AE value is building governed agentic analytics systems — semantic layers, docs, skills, evals, provenance, and correction loops — rather than just writing SQL faster.

Back to [[Agentic-Analytics-Engineering]].