Summary
Sherrod, an engineering manager at Stripe, gives a detailed walkthrough of Kai — Stripe’s internal company brain and agent platform — in this How I AI interview. Built in early 2026 when Stripe faced an avalanche of AI tools but no coherent deployment strategy, Kai was architected around governance from the start rather than raw capability. It is context-aware by design: it knows who you are, your role, your team’s current projects, and the org chart, giving it far more signal to do the right thing than a generic chat interface.
The technical architecture is notable for several concrete mechanisms. Kai implements a three-layer triage for analytics requests: first surface existing reports, then use a pre-built analytics layer with structured queries, then fall back to writing custom SQL against the data warehouse only if necessary (protecting warehouse infrastructure from agent-driven query storms). Tool policies are scoped by role: an HR team member’s agent is explicitly prevented from writing sensitive data to public Google documents, without blanket-restricting all tool access. A skill-sharing platform lets individuals publish their automations company-wide, compounding organizational value over time. Kai now has 86%+ adoption across all of Stripe globally.
The interview is candid about what went wrong during development. Agents “dial up” infrastructure failure amplitudes, Sherrod explains — at least one agent nearly took down a core system before being caught. The conversation is one of the most practical accounts available of how a large engineering organization actually deployed, governed, and scaled an internal AI agent platform, making it essential reading for enterprise AI leaders navigating the same transition.
📺 Source: How I AI · Published September 07, 2026
🏷️ Format: Interview
