Why Claude Handles 95% of Anthropic’s Internal Analysis Queries

Anthropic’s recent report states that Claude now handles approximately 95% of internal analysis queries, allowing employees to retrieve business data independently without relying on the data team.

In Anthropic, 95% of business analysis queries are automated through Claude, achieving an overall accuracy of about 95%.

The company emphasizes that this achievement stems more from data governance, semantic definitions, and operational standards than from advances in large‑model technology.

The report notes that the reliability of AI‑driven analysis depends on the quality of the underlying data platform, making data modeling, testing, metadata management, and quality checks essential for accuracy.

Without relevant skills, Claude answered only 21% of analysis questions correctly. After encoding analysis workflows and business context as "skills," overall accuracy rose above 95% and approached 99% in certain domains.

Anthropic’s approach addresses a common analytics challenge: self‑service access can cause dataset overlap and metric definition conflicts, while narrow reporting environments fail to support long‑tail business questions, leading to dashboard sprawl. The five‑person data science and engineering team explains that the data foundation is a dimensional model accessed through a semantic layer, which reduces ambiguity and translates stakeholder terms like “weekly active users” into concrete, controlled entities.

The analysis architecture is described as four layers:

Data foundation – governed models, metrics, and metadata.

Knowledge layer – semantic definitions, lineage, and business context.

Skill layer – repeatable analysis workflows encoded as skills.

Verification system – checks for output correctness and consistency.

Anthropic summarizes three principles for successful AI analysis:

Maintain a single source of metric data.

Ensure users can easily locate the correct data.

Continuously monitor and update stale definitions.

Increasingly, real‑world deployments show that AI performance is often limited more by context definition than by model capability.