# anthropic-self-service-data-analytics-claude-agentic-stack-2026-06-03

## Veille

Engineering write-up from Anthropic's **Data Science & Data Engineering** team (Chen Chang, Clement Peng, Justin Leder, Johanne Jiao, Josh Cherry) published on **June 3, 2026** on the Anthropic blog (*Enterprise AI* category, focus on **Claude Code**). **Headline result**: ***"95% of business analytics queries are automated by Claude, with ~95% accuracy in aggregate"*** (up to **~99%** in certain domains). **Core problem**: analytics is **not** code — *"there's often only a single correct answer using a single correct source"* — it requires **mapping a user question to precise, up-to-date entities** in the data model. Three **failure modes**: (1) **concept↔entity ambiguity** (e.g. *"active users"*: which actions? exclude fraudsters? which window?); (2) **staleness** (assets and the agent's knowledge become *"subtly wrong"*); (3) **retrieval failure** (*"80% of failed queries had the information present in the corpus"* but unfindable). **Solution = a 4-layer "agentic analytics stack"**: (L1) **Data foundations** — dimensional modeling, **canonical datasets** *"single source-of-truth"*, metadata *"as a first-class product"*, integrity via CI/CD; (L2) **Sources of truth** in decreasing order of trust — **semantic layer** (the agent is *"structurally required (by skill instruction) to leverage the semantic layer first"*), lineage graph, **query corpus** (distilled into structured docs, **not** raw retrieval), business context (knowledge graph: roadmaps, decision logs, org); (L3) **Skills** — the decisive lever: ***"without skills … didn't exceed 21% … Adding skills gets these numbers consistently above 95%"***; structured **in pairs** (*Knowledge skill* = router to ~30 reference files; *Unbook skill* = senior analyst workflow: clarify → find sources → execute → **adversarial review**); **colocated** maintenance (*"a code-review hook flags any reporting-model change that doesn't touch a skill file"* → **~90% of data PRs include a skill change**); (L4) **Validation** — offline evals (threshold ~90% to launch an agent, target ~100%), **ablation testing** (notable negative result: raw grep across thousands of SQL files → accuracy moves *"less than a point"*), online (adversarial review: **+6% accuracy, +32% tokens, +72% latency**), **provenance footers** (source tier + freshness + ownership), **active correction harvesting** (scheduled agents scanning channels to draft markdown fixes). **Strategic insight**: *"documentation generated, definitions owned by humans"* — letting the LLM **define** metrics was *"net-negative"*. **Minimal starting point**: a handful of canonical datasets + a few dozen evals + a *thin knowledge skill* capture *"most of the upside"*. Strongly converges with [[shihipar-claude-code-lessons-building-skills-2026-06-03]] (skills = folders, Gotchas, hooks), the *systems around the model* doctrine of [[dropbox-okumura-beyond-code-generation-engineering-productivity-ai-agents-2026-05-28]], the **semantic layer / ontology** of talisman-modern-data-101-ontology-pipeline-refresh-2026-05-04 and seale-semantic-agent-model-harness-ontology-data-2026-04-17, the *context development lifecycle* of debois-tessl-context-development-lifecycle-ai-coding-agents-2026-02-19, and the UDA/knowledge graph of netflix-uda-unified-data-architecture-knowledge-graph-2025-06-12.

## Titre Article

How Anthropic enables self-service data analytics with Claude

## Date

2026-06-03

## URL

https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude

## Keywords

self-service analytics, agentic data analytics, Claude Code, single correct answer, question-to-entity mapping, data model, failure modes, concept-entity ambiguity, active users, asset staleness, data staleness, retrieval failure, agentic analytics stack, data foundations, dimensional modeling, canonical datasets, single source of truth, metadata as first-class product, lineage, semantic layer, structurally required, query corpus, business knowledge graph, decision logs, skills, knowledge skill, unbook skill, router, 30 reference files, adversarial review, retention curves, funnel analysis, rate decomposition, Gotchas, skill-model colocation, code-review hook, 90% of PRs with a skill, validation, offline evals, ground truth snapshot, ablation testing, raw grep negative result, provenance footers, source tiers, correction harvesting, scheduled agents, definitions owned by humans, documentation generated, 21% without skills, 95% accuracy, 99%, 80% info present, minimal starting point, data governance, Anthropic Data Science

## Authors

**Chen Chang, Clement Peng, Justin Leder, Johanne Jiao, Josh Cherry** — équipe **Data Science & Data Engineering d'Anthropic**. Article publié le **3 juin 2026** sur le blog Anthropic (claude.com/blog), catégorie *Enterprise AI*, ~5 min de lecture.

## Ton

**Profile**: internal engineering write-up (*engineering blog post*), **technical-pedagogical and empirical** register, high level (data engineering, semantic layer, evals, ablations), aimed at **data/analytics and platform teams**. Posture: *"here's what worked and what didn't work for us"*, backed by **numbers** and owned **negative results**.

**Style**: Structured in **layers** (a 4-tier stack) with a clear throughline (collapse ambiguity → make discoverable → flag staleness). Scientific honesty: **ablations** and **negative results** are foregrounded (raw grep <1%, LLM definitions net-negative), lending credibility to the piece. Dense with precise **metrics** (21%, 95%, 99%, +6%/+32%/+72%, 80%, 90%). Data-governance vocabulary (canonical, lineage, ownership, tiering).

**Key aphorisms**:
- ***"For analytics use cases, there's often only a single correct answer using a single correct source."***
- ***"Without skills … didn't exceed 21% … Adding skills gets these numbers consistently above 95% in aggregate."***
- ***"Agents are structurally required (by skill instruction) to leverage the semantic layer first."***
- ***"Treat metadata as a first-class product."***
- ***"Roughly 90% of our data-model PRs now include a skill change in the same diff."***
- (silent failure) *"The answer is wrong, but looks plausible and is used without objection."*

**Elaborated metaphors / frames**:
- ***Single correct answer*** — analytics ≠ code: the model's creativity isn't what's valued but **deterministic accuracy**; infrastructure takes priority over the generative layer.
- ***Agentic analytics stack (4 layers)*** — foundations → sources of truth → skills → validation.
- ***Skills in pairs*** — *Knowledge* (router/search) + *Unbook* (senior analyst workflow).
- ***Documentation generated, definitions owned by humans*** — the LLM drafts, the human **decides** the definitions.
- ***Provenance footers*** — traceability of the answer (source tier / freshness / ownership) as an anti-staleness guardrail.
- ***The bottleneck is structure, not access*** — the raw-grep ablation proves that more access ≠ more accuracy.

**Epistemic stance**: an **empirical, measured** write-up, with reproducible evals and ablations — one of Anthropic's most rigorous posts on **context engineering** applied to data. Reads as a production **architecture blueprint** for analytics agents, transferable outside Anthropic.

**Authority**: (a) **the model's own producer** (Anthropic) applying Claude to its own back-office data; (b) **numbers and ablations** (negative results included); (c) directly actionable **layered architecture**; (d) consistency with the *skills / systems around the model* doctrine found across the rest of the ecosystem.

## Pense-betes

- **Date / source**: **June 3, 2026**, Anthropic blog (*claude.com/blog*, Enterprise AI). Authors: Anthropic's **Data Science & Data Engineering** team (Chang, Peng, Leder, Jiao, Cherry).
- **Headline result**: **95% of business analytics queries automated**, **~95% accuracy** in aggregate (up to **~99%** in certain domains).
- **Thesis**: analytics ≠ code — *"only a single correct answer using a single correct source"* → the challenge is **mapping the question to the right, current entities**, not generating. ### 3 failure modes 1. **Concept↔entity ambiguity** — hundreds of options; e.g. *"active users"* (which actions? exclude fraudsters? which lookback window?). 2. **Obsolescence (staleness)** — assets and the agent's knowledge become *"subtly wrong"* (schemas/definitions change continuously). 3. **Retrieval failure** — **80%** of failed queries had the info **present** in the corpus regardless.
- Ultimate danger = **silent failure**: *"the answer is wrong, but looks plausible and is used without objection"*. ### 4-layer agentic stack
- **L1 — Data foundations**: dimensional modeling, **canonical datasets** (single source-of-truth, owned, consumption-ready), metadata as a *"first-class product"*, cross-layer integrity via CI/CD.
- **L2 — Sources of truth** (trust ↓): **semantic layer** (mandatory first) → **lineage/transformation graph** → **query corpus** (distilled into docs, not raw retrieval) → **business context** (knowledge graph: roadmaps, decision logs, org).
- **L3 — Skills** (the lever): **21% → 95%+**. Pairs: **Knowledge skill** (router → ~**30** reference files) + **Unbook skill** (senior analyst workflow: clarify → sources → execute → **adversarial review**; reusable patterns: retention curves, rate decomposition, funnel). Typical doc skeleton: Quick reference / Dimensions & key tables / **Gotchas** / Best practices & query patterns / Cross-domain refs.
- **L4 — Validation**: offline evals (threshold **~90%** to authorize an agent, target ~100%, ground truth fixed on a snapshot, *"store results like telemetry"*), **ablation testing**, online (adversarial review, provenance footers, data quality checks, **correction harvesting** by scheduled agents). ### Notable results (numbers & ablations)
- **Without skills ≤ 21%**; **with skills > 95%** (≈99% in certain domains).
- **Adversarial review**: **+6%** accuracy, but **+32%** tokens and **+72%** latency (a trade-off to accept).
- **Raw grep ablation** (thousands of accessible SQL files): accuracy moves ***"less than a point"*** → **the bottleneck is structure, not access**.
- **LLM-generated definitions** = *"net-negative"* (they encode the very ambiguities meant to be eliminated) → **definitions owned by humans**.
- **Unstructured retrieval** across thousands of queries: gain **< 1 point** → corpus should be **distilled**, not searched raw.
- **Maintenance**: code-review hook → **~90% of data PRs** touch a skill file in the same diff. ### To use in engagements / presentations
- **Production analytics agent blueprint** that transfers elsewhere: data governance + **mandatory** semantic layer + skills + threshold evals.
- **Minimal starting point** (reusable as-is): *"a handful of canonical datasets, a few dozen offline evals, and a thin knowledge skill"* capture *"most of the upside"*.
- **Anti-naive-RAG arguments**: the two negative results (raw grep, unstructured retrieval) are ammunition against *"just give the agent everything"*.
- **Living governance metric**: *"90% of PRs include a skill"* = proof that docs and code evolve together (anti-staleness).
- Connects to: skills (Shihipar/Anthropic), semantic layer/ontology (Talisman, Seale), context engineering (Debois/Tessl), *systems around the model* (Dropbox), knowledge graph data (Netflix UDA).

## RésuméDe400mots

Published on **June 3, 2026** on the Anthropic blog, this write-up from the **Data Science & Data Engineering** team (Chen Chang, Clement Peng, Justin Leder, Johanne Jiao, Josh Cherry) describes 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 isn't 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 counts as *"active users"*? do fraudsters get excluded? what window?), **staleness** of assets and the agent's knowledge, and **retrieval failure** — **80%** of failed queries had the information present in the corpus regardless. The worst case is **silent failure**: a wrong but plausible answer used without objection.

The response 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** 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 without a matching skill change — **~90% of data PRs** now include a skill in the same diff. (4) **Validation**: threshold-based offline evals (~90% to authorize an agent), **ablation testing**, and online guardrails (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 to 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 start: a handful of canonical datasets, a few dozen evals, a thin knowledge skill.

## GrapheDeConnaissance

- équipe Data Science & Data Engineering Anthropic —publie→ How Anthropic enables self-service data analytics with Claude (DOCUMENT, 0.97)
- Anthropic —mesure→ 95% des requêtes analytics métier automatisées via Claude (MESURE, 0.95)
- analytics agentique —est_basé_sur→ une seule bonne réponse depuis une seule bonne source (CONCEPT, 0.94)
- skills (analytics) —améliore→ la précision de 21% à plus de 95% (CONCEPT, 0.95)
- agent analytics —utilise→ semantic layer (CONCEPT, 0.93)
- équipe Data Science & Data Engineering Anthropic —recommande→ distiller le query corpus en docs structurées (pas de retrieval brut) (METHODOLOGIE, 0.9)
- retrieval non structuré —améliore→ la précision de moins d'un point (CONCEPT, 0.9)
- équipe Data Science & Data Engineering Anthropic —mesure→ accès grep brut au SQL : précision quasi inchangée (goulot = structure, pas accès) (MESURE, 0.92)
- revue adversariale —améliore→ la précision de 6% (mais +32% tokens, +72% latence) (CONCEPT, 0.9)
- équipe Data Science & Data Engineering Anthropic —affirme_que→ les définitions de métriques générées par LLM sont net-negative sur les evals (AFFIRMATION, 0.88)
- équipe Data Science & Data Engineering Anthropic —recommande→ définitions de métriques détenues par des humains (METHODOLOGIE, 0.92)
- hook de code-review —permet→ colocalisation modèle/skill (signale tout changement de modèle sans fichier skill) (CONCEPT, 0.9)
- équipe Data Science & Data Engineering Anthropic —mesure→ ~90% des PR data incluent un changement de skill dans le même diff (MESURE, 0.9)
- échec silencieux —permet→ une réponse fausse, plausible, utilisée sans objection (CONCEPT, 0.9)
- équipe Data Science & Data Engineering Anthropic —recommande→ seuil d'eval (~90%) requis avant qu'un domain owner lance un agent (METHODOLOGIE, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/anthropic-self-service-data-analytics-claude-agentic-stack-2026-06-03/