Stripe builds AI assistant Kai in one week

Stripe has successfully deployed a company-wide AI assistant called the Knowledge AI Platform, or Kai, which was built by a single engineer in just one week. Developed on the LangChain and LangGraph stack using the open-source Deep Agents harness, Kai serves as a context-aware coworker for Stripe employees. The rapid development was made possible by leveraging Deep Agents to handle foundational LLM interactions, state management, and tool-calling loops, allowing Stripe's team to focus entirely on company-specific workflows.

To achieve this, Stripe implemented three key middleware components: a virtual filesystem backed by S3 for persistent session context, a sandboxed Python environment for secure data analytics, and summarization middleware to manage long-running sessions. Kai currently navigates over 500 internal Model Context Protocol tools and more than 1,000 skills contributed by over 100 teams. However, managing this scale presents challenges; the engineering team noted a 1024-character limit on frontmatter and observed quality degradation in frontier models when dealing with more than 150 skills simultaneously.

Despite these scaling hurdles, Kai's internal rollout was an immediate success. It met its quarterly adoption target within its first week of open preview, growing sixteen-fold from 296 users to over 5,000 in about four weeks. Today, 83 percent of Stripe employees use Kai weekly across more than 60,000 sessions. Adoption is particularly high in non-technical departments, reaching 95 percent in marketing and 87 percent in go-to-market teams, where employees use it to synthesize data and draft documents.

For AI practitioners, Stripe's success proves the immense return on investment of adopting standardized agent harnesses and Python-native stacks, even in organizations historically dominated by languages like Ruby or Java. By utilizing pre-built middleware for sandboxing and context management, developers can bypass months of infrastructure engineering. This shift allows teams to focus on federated skill development, where individual business units maintain their own agent capabilities, ultimately making generative AI highly accessible to non-technical staff.