How Block Gives 4,000 Employees AI-Powered Data Access via Claude and Databricks

Block operates one of the more complex data environments in fintech, supporting Square’s merchant platform, Cash App’s consumer payments network, Afterpay’s buy-now-pay-later product, and several other financial services under one engineering organization. With roughly 10,000 employees and vast internal datasets, the challenge was not a shortage of data—it was accessibility. Most of that data could only be reached by employees who knew SQL and understood the architecture of Block’s internal systems, which effectively gatekept a major organizational resource.

The company had worked with Databricks for years as its primary data engineering platform and had built a strong foundation around Spark for large-scale data processing. The idea of combining that infrastructure with large language models emerged organically when a small team began experimenting with an internal AI coding assistant, which they called codename goose. The original purpose was straightforward: help engineers write better code faster. What they discovered was that the tool’s potential extended far beyond software development.

Block chose Claude as the default model for goose after rigorous benchmarking across models. The Claude family consistently outperformed alternatives on the tasks Block cared most about measuring. The deployment runs through a Databricks endpoint, allowing Claude to connect back to Block’s internal databases and datasets via an OAuth-secured integration. This architecture meant Block could extend LLM access to every employee without distributing or managing API keys—a meaningful security and operational simplification for a company handling financial data at scale.

The practical impact of goose on how Block employees work is substantial. Engineers can now write SQL against internal databases without needing to know Block’s specific schema or backend systems. Designers who previously lacked the technical skills to build functional prototypes can now describe a concept in natural language and receive a working version. A machine learning engineer created a system that teaches Claude to write code for Block’s internal Beacon service through an MCP server, removing the need for deep backend knowledge. Operations teams use goose to close support tickets and query cross-functional data in workflows that previously required handoffs to data teams.

Adoption tells the most direct story. Around 4,000 of Block’s 10,000 employees now use goose actively, spanning 15 different job profiles. Adoption doubled in one month and user engagement grew 40 to 50 percent week over week as employees found new applications. With the platform’s security infrastructure in place, Block can make new models available to the entire organization by changing a single configuration—enabling a pace of iteration that would be difficult to achieve with a more distributed deployment model.