HeadsUpAI

Anthropic automates 95 percent of internal analytics using new Claude agent stack

Anthropic released a technical blueprint for building self-service business analytics agents using Claude. The framework addresses the "data is not software" problem—where analytics requires a single correct answer amidst ambiguous data—by implementing an agentic analytics stack. This architecture moves beyond text-to-SQL generation to focus on data foundations and verification loops.
Internal Automation Rate
95%
Aggregate Accuracy
~95%
Accuracy Gain (Adversarial Review)
6%
Token Cost Increase (Adversarial Review)
32%
Latency Increase (Adversarial Review)
72%

This release builds on Anthropic's evaluation framework, shifting focus to domain-specific reliability. While tools like Claude Code excel at creative software tasks, business data often suffers from staleness. By forcing agents to use a semantic layer (a central source of metric definitions), organizations can reach the 95% accuracy levels Anthropic reports internally.

Teams can implement this stack by adopting the provided Skill File Skeleton to define procedural knowledge. This includes setting up adversarial review sub-agents to challenge query assumptions and using MCP (a universal connector for AI tools) to connect agents to governed warehouses. The framework is available now for organizations deploying Claude in Claude Cowork.

View the full update on claude.com
ClaudeDevs
ClaudeDevs
@ClaudeDevs
X

How do we automate business analytics with Claude? New blog post covering our best practices for skills, data foundations, and evaluations when building agents to perform data analysis: https://t.co/mfEJMAQFBU

14retweets165likes
View on X

Still wondering? A few quick answers below.

It is a four-layer architecture designed to ensure AI agents produce accurate data insights. The stack includes data foundations to eliminate ambiguity, sources of truth like semantic layers, procedural skills for navigating warehouses, and validation loops. This system helps agents map natural language questions to specific, governed data entities.

Skills are structured markdown files that provide an agent with procedural knowledge, such as which tables to join and which filters to apply. Anthropic uses a "Skill File Skeleton" to standardize these instructions, ensuring agents follow the same logic as senior analysts when performing complex data tasks.

Unlike coding, which allows for creative solutions and deterministic testing, business analytics often has only one correct answer hidden within ambiguous data models. Anthropic argues that analytics accuracy is a context and verification problem rather than a code generation issue, requiring agents to be strictly routed to governed sources.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update