Inside Kai: Stripe's In-House AI Agent Platform for Knowledge Work

Stripe has rolled out Kai, an internal AI agent platform aimed at non-engineering staff—sales reps, finance analysts, and technical account managers—who were sidelined by earlier tooling. Its previous attempts fell short: a no-code builder spawned thousands of inconsistent one-off agents, while coding-oriented agents introduced security risks in the hands of non-technical users. Kai consolidates that sprawl into a single agent service that connects employees to more than 1,000 internal tools and skills, handling everything from one-off questions to multi-day projects.

Architecturally, Kai is split into three layers. A surface-agnostic API exposes the same underlying agent through a web app, Slack, and Chrome-extension embeds inside third-party tools. AgentStudio serves as the control plane where domain experts build, tune, test, and monitor custom agents against their own data. The execution layer is built on LangChain’s deepagents running on Kubernetes, giving each session its own sandbox and a multi-tenant virtual filesystem; it uses a hybrid RAG/LLM approach to pick the right skill from the 1,000-plus available and sustains very long sessions—one reportedly hitting 932 turns—without context degradation. Guardrails are baked in, including blocks on mixing data across unrelated customer contexts.

Stripe reports aggressive adoption and business impact: 83% weekly active users company-wide within months, with account executives on Kai showing roughly double the sales activity, 26% more revenue opportunities, and 39% more closed deals. The company estimates the platform shifts about 25,000 hours a year from administrative busywork to revenue-generating work. The takeaway for other enterprises is the model itself—centralized, sandboxed, embedded in existing workflows rather than bolted on—which matters as much as the headline metrics.