How Block is becoming the most AI-native enterprise in the world | Dhanji R. Prasanna

What It Covers

Dhanji Prasanna, CTO of Block, explains how his company became AI-native through organizational restructuring, building the open-source agent Goose, and achieving 20-25% manual hours saved across 3,500 employees while prioritizing technology-first culture.

Key Questions Answered

• •Organizational structure drives outcomes: Block shifted from GM-led business units to functional organization where all engineers report to one leader, enabling singular technical focus and AI adoption. This structural change proved more impactful than any individual tool, allowing teams to share platforms, move between projects, and align on technical strategy company-wide.
• •Goose agent delivers measurable productivity: Engineers using Block's open-source AI agent Goose report 8-10 hours saved weekly, with company-wide manual hours savings trending toward 20-25%. The agent uses Model Context Protocol to orchestrate across systems like Snowflake, Tableau, and Git, building features autonomously overnight and opening pull requests without human intervention.
• •Non-technical teams gain most from AI tools: Enterprise risk management, legal, and support teams building their own software tools show the highest productivity gains, compressing weeks of work into hours. This eliminates waiting for internal development teams and enables self-service automation, representing a fundamental shift in who can build software within organizations.
• •Code quality disconnects from product success: YouTube succeeded with videos stored as MySQL blobs and slow Python stack, while Google Video failed despite superior architecture. Engineers should focus on solving user problems rather than refactoring code, as technical excellence doesn't correlate with product-market fit or business outcomes in practice.
• •Future work involves continuous AI operation: LLMs should work overnight and weekends building multiple experimental approaches simultaneously, not sit idle. Engineers will describe several solutions in detail, let AI build them all asynchronously, then evaluate and discard most versions—fundamentally changing from choosing one path to exploring many paths in parallel.