Framework Business Data StrategyMCPAnalyticsAI Agents

Building the Data Foundation for Automated Analytics

How data mesh, a governed semantic layer, and purpose-built MCP servers can make AI-powered analytics both trustworthy and flexible.

Originally published on Block engineering blog here.

I coauthored this piece with Alyssa Ransbury, Andrew Kuttig, and Phil Azar about the data foundation behind automated analytics at Block.

The short version: AI-powered analytics only works when the underlying data system is trustworthy. The post walks through how we approached that problem by combining data mesh principles, a governed semantic layer, and two purpose-built MCP servers with different jobs:

  • Block Data MCP for deterministic, governed metric answers where accuracy matters most.
  • Query Expert MCP for exploratory analysis, data discovery, and SQL generation with richer context.

The piece also covers why the project started with a concrete set of recurring business questions, why we optimized for 100% metric accuracy on governed metrics, and how governance, ownership, and metadata became part of the product instead of a separate process layered on top.

Adoption grew quickly: by Q1 2026, Block Data MCP reached 1,812 quarterly unique users and Query Expert MCP reached 1,899, while controlled tests showed an 87% reduction in analyst time.

Read the full article on the Block Engineering Blog.

Glowing data foundation illustration showing messy inputs becoming structured analytics outputs