Knowledge agents for non-engineers: Stripe's Kai platform

Summary

An account from Stripe's product lead on the design and early results of Kai, an AI agent platform built for non-engineers across the company. The piece outlines why knowledge work (research, data analysis, compliance reviews) requires a different agent architecture than coding agents, and describes Stripe's three-layer approach: surface-agnostic APIs, domain-specific agent builders, and shared execution environments with security and access controls. Within two weeks of launch, 83% of Stripe became weekly active users. Account executives using Kai generated 26% more revenue opportunities and closed 39% more deals; new hires used it 2.7x more than baseline cohorts; across the company, Kai shifted 25,000 hours per year from admin work to revenue work.

Key Claims

• Knowledge agents differ from coding agents: knowledge work is domain-specific, multi-turn, iterative, and demands task-specific tools and outputs; coding work has a uniform shape, so a single agent architecture suits it.
• An agent platform scaled to non-technical users requires three layers: surface-agnostic APIs (the agent as a service, not an app), domain-owned agent builders (AgentStudio, where business experts configure agents without centralizing expertise), and shared execution environments with security and access controls built in.
• Kai's agent harness uses RAG and LLM routing to select the right skill from 1,000+ tools and data sources. Without pre-existing folder structure (as coding agents have), a hybrid approach handles skill selection in knowledge agents.
• Multi-turn state management is essential: Kai holds state across sessions reaching 932 turns, allowing iterative reasoning that builds on prior context without degrading or hitting context limits.
• Early adoption data shows: GTM users (Account Executives) produce 2x sales activity, 26% more revenue opportunities, and close 39% more deals in weeks they use Kai; new hires use it 2.7x more than peers; 5,000+ sessions per day focus on data analysis.
• Knowledge agents must enforce implicit access rules not present in coding (e.g. preventing data from unrelated customer contexts appearing in the same session), which a monolithic agent cannot encode; the platform distributes this governance to domain teams.

Quotes

• "Coding agents operate in an environment with decades of fast, verifiable guardrails: compilers reject invalid syntax, tests catch regressions, and git makes every mistake reversible. Knowledge work has very little support for these constructs."
• "The platform has to meet the user where they are."
• "Kai has helped bring that same sense of leverage to knowledge workers across Stripe."