Stripe Says 83% of Staff Use Its Internal Kai AI Agent Weekly
- Stripe shipped Kai, its knowledge AI platform, in April 2026 and reports 83% weekly active use company wide today, including nearly all of GTM (marketing, sales, CSMs, TAMs).
- Kai replaces two earlier options: a NoCode Agent Builder that spawned over 4,000 agents with duplicated prompts and no easy way to monitor them, and coding agents that raised security risks and dumped support work on code quality teams.
- The platform connects employees to over 1,000 internal tools and skills for non-coding work, such as querying data warehouses, researching accounts before sales calls, triaging incidents, and modeling revenue scenarios.
- Stripe identified three design requirements: scaling expertise without centralizing it across dozens of domains, embedding the agent into whatever surface the user works in (such as a Finance budget modeling app), and building guardrails that code has natively, for example an invariant barring data from two unrelated customer contexts in one analysis.
- Most Kai sessions run many turns, with users doing deep research and producing artifacts they refine before sharing, so the work differs from coding where the tool set stays roughly constant.
Hacker News opinions
TIL people still use LangChain. It's been bad garbage since day one and oddly it never improved.
I thought it was crap the moment I saw they needed LangChain to develop a harness.
There's like 1000 LangChain alternatives and none are good, they're all vibe-coded unstable projects. I run Apache Burr and it's still not great. I just refuse to use software that has a pricing page if it isn't SaaS.
I built an internal set of agents to run our company (finances, an LLMWiki, plus clients, contracts, billing and time tracking all via MCPs) and it's also called Kai, since our company is Kaizen.
I read the buzzword "Knowledge AI Platform" but saw no real knowledge management feature like verification or transparency. It reads like any generic agent builder, maybe built internally to justify the team.
None of these internal platform posts impress me. They stop at "I orchestrated agents this way", which is about as impressive as "look at my CloudFormation template".
Not these again. Spotify built something similar and made ridiculous claims like saving 90 percent of tokens. Anything goes these days apparently.
The post never says why they built their own instead of using something open source like bionic-gpt.
I'd guess size and scale. An out of the box tool won't have the capabilities they need, and Stripe wants to be opinionated about how the company works with these tools.
Their first paragraph does answer that: no existing tool could handle the data security requirements and specific workflows Stripe needed.
So is this like CloudflareOS? Seems every company with spare engineering resources builds one of these internally.
This is where a lot of companies are heading: on-prem platforms giving non-engineers agents as capable as coding agents but much more managed and governed. I'm building Lightspeed in that same direction.
Very close to the direction we take with windmill.dev, except we call it "operator builders" rather than leaning on "knowledge graph", which the article barely explains. What I read is a developer platform and runtime with a secure code runtime, ACLs and permissions, and app building like CloudflareOS gadgets.