Chapter Summary: The AI-First Mindset¶

Key Takeaways¶

Build capabilities, not features: Harvey, Dust, and Glean won by building underlying capabilities that compound—legal reasoning, multi-agent infrastructure, organizational understanding. The paradox: you discover capabilities by building specific applications first. Features accumulate complexity; capabilities multiply value.
Think Buy, Boost, or Build: 65% of AI costs occur after deployment. Klarna saved $40M buying. Morgan Stanley boosted with proprietary data. Bloomberg spent $8M building. Three questions determine your path: Is this capability your moat? Does the data require isolation? Do you have resources to execute?
Design your data flywheel: Usage generates signal, signal improves the model, better models drive more usage. Three requirements make flywheels spin: structured data for learning, fast feedback loops, visible improvement to users. If you can't articulate the loop, you don't have a flywheel.
Iteration speed is the competitive variable: 720 learning cycles per year (daily) versus 52 (weekly). Same talent, same resources, 14x more learning. Your slowest loop sets your pace. Vercel iterates daily. Stripe deploys model updates in hours.
Collaboration requires design, not assumption: MIT found human-AI teams don't automatically outperform the best single performer. Partner under direction, not teammate. Design for complementarity—information and capability asymmetry—or let the best performer work alone.