Source: Stripe Dot Dev Blog — 2026-07-30
Summary
Stripe built an internal AI agent platform called Kai, aimed not at engineers but at non-coding knowledge workers — sales, finance, ops, and PMs. It's built on top of LangChain/LangGraph's "Deep Agents" harness and wires into more than 1,000 internal tools contributed by over 100 teams. LangChain published its own companion case study on the same build, but Stripe's post is the primary source here.
Key Takeaways
- Kai is layered on four levels: a foundational agent layer from Deep Agents, a Stripe-specific layer, a configuration layer, and a UI layer — each one narrowing the general-purpose harness toward Stripe's actual workflows.
- It connects to over 1,000 internal skills/tools contributed by 100+ teams, including BI dashboards, Zoom, Google Workspace, and Slack.
- It's designed to span both quick one-off queries and multi-day projects, not just simple Q&A.
- Adoption scaled from 296 to over 5,000 users within weeks of launch, reaching an 83% weekly-active-user rate across 60,000+ sessions.
- Go-to-market teams — marketing, sales, CSMs, TAMs — showed especially strong uptake compared to other functions.
Reel Script
Hook (~15-20s, 35-45 words) Stripe didn't build another coding assistant. They built an AI platform for sales, finance, and ops people — and in a few weeks it went from 296 users to over five thousand, with 83 percent of them coming back weekly. Here's how it's actually built.
Core Concept (~45-90s, 105-200 words) The interesting part isn't the adoption number, it's the architecture underneath it, called Kai. Stripe didn't build an agent from scratch. They started with an open-source harness called Deep Agents from the LangChain team — think of it as the engine that handles the core agent loop: planning, tool calling, keeping track of state. On top of that, Stripe stacked three more layers. A Stripe-specific layer that bakes in their internal context and conventions. A configuration layer that lets different teams tune behavior without touching code. And a UI layer on top for the actual employees using it. That's the pattern worth remembering: instead of reinventing the agent loop, you inherit it, and you specialize by layering, not rewriting. That's also how Kai plugs into over a thousand internal tools — dashboards, Slack, Google Workspace — without every team having to reimplement agent plumbing themselves.
Hands-On (~45-150s, 105-350 words) Picture the request path. A salesperson asks Kai a question in the UI layer — that's layer four, just the chat interface. That request drops into the configuration layer, which applies whatever settings that team or use case needs — maybe which tools are allowed, maybe response style. Below that sits the Stripe-specific layer, which injects company context: internal terminology, data sources, permissions, the stuff that makes generic Deep Agents actually useful at Stripe instead of a blank slate. And underneath all of that is the Deep Agents foundation itself — the actual reasoning and tool-calling engine doing the planning and execution. When the agent needs to act, it reaches into that pool of over a thousand tools built by more than a hundred different teams — pulling a BI dashboard, checking a Zoom recording, searching Google Workspace, posting to Slack. None of those hundred teams had to build their own agent loop; they just registered a tool into a shared foundation. That's the leverage: one harness, one thousand tools, thousands of employees, without duplicating the hard part a thousand times.
Takeaway (~20-30s, 45-70 words) The lesson isn't "Stripe built a chatbot." It's that they treated the agent harness as infrastructure to inherit, not a product to reinvent — and that's what let a thousand-plus tools plug in cleanly. If you're building internal AI tools, ask whether you're building a harness or just borrowing one. Go read how they layered it.