AI news ·
Stripe builds company-wide AI agent Kai on Deep Agents, hitting quarterly adoption target in one week
Stripe built its internal AI assistant Kai in one week and now 83% of employees use it weekly across 60,000+ sessions. Marketing and go-to-market teams adopted at 95% and 87%, outpacing engineering.
Stripe built Kai, a company-wide AI assistant, in one week using LangChain's open-source Deep Agents harness. The platform hit its quarterly adoption target in a week and now sees 83% of Stripe employees using it weekly across more than 60,000 sessions, with adoption strongest among marketing and go-to-market teams.
Kai is a session-based agent that connects to Stripe's internal data warehouse, Slack, and Google Suite. Employees use it to synthesize data, draft documents, analyze trends, and produce reports. Unlike generic AI assistants, Kai arrives preloaded with Stripe context - it understands internal systems, data sources, and company norms without users needing to explain their job function every time.
What made the one-week build possible
Sharadh Krishnamurthy, Engineering Manager for the Agent Foundation team, said the Deep Agents layer "solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems." Staff Software Engineer Anupam Upadhyay built the initial version in a single week by using three middleware components from Deep Agents: a virtual filesystem backed by S3 for persistent file context, a sandbox for secure code execution, and summarization middleware to manage context across long-running sessions.
Stripe spent over a decade building internal tooling around Ruby and Java. Adopting a Python-native stack meant rebuilding service scaffolding around a new language. Kai proved the return on that investment immediately. Chrissie, Head of AI Platform, said, "Kai completely solidified people's belief that Deep Agents is the way to go."
How skills route Kai through hundreds of tools
Kai navigates over 500 internal MCP tools and more than 1,000 skills contributed by over 100 teams. The system uses a federated approach: individual teams own and maintain their own skills, which are structured modules that encapsulate how to accomplish specific tasks and which tools to load. A sales ops user gets a different skill set than a finance user, and individuals can add more skills beyond their function defaults.
Skills use allowedTools lists to drive dynamic tool loading. Rather than loading all tools upfront, the LLM selects which skills to load, and those skills gate which tools enter context. Anupam Upadhyay explained, "The LLM is spending effort figuring out which skills to load, so that is what we then rely on to load the relevant tools as well." Certain foundational skills remain pinned regardless of model decisions to ensure consistent policy behavior.
Adoption spread fastest outside engineering
Kai went from 296 users to over 5,000 in roughly four weeks. Sharadh Krishnamurthy said, "It just caught fire. People used it, and it just clicked for them." The bulk of new users were sellers, finance analysts, and business operations employees - people who had been told to use AI but never found a tool that fit how they worked.
One user told the team, "My Stripe career is divided into before and after Kai." Another said, "Kai is absolutely unbelievable for salespeople. Tens of thousands of hours will be saved." Marketing teams adopted at 95%, GTM teams at 87%, both higher than engineering. Finance teams use it for data analysis and dashboard generation. New hires use it to accelerate onboarding.
Krishnamurthy described the feedback from reluctant users: "There are a lot of people who said, 'I found the previous thing unapproachable. I had to tune a bunch of knobs. But I was being asked to use AI, and I really tried.' And then: 'this is great, I don't have to do any of that stuff.'"
What the team is building next
With over 1,000 skills now available, no LLM can perform reliable selection across the full catalog without assistance. The team is building a hybrid selection system that will use RAG or a classifier layer to prefilter before the LLM makes final decisions. They are also investing in security and compliance guardrails as adoption spreads, and working on ML-driven personalization that adapts agent behavior to user context automatically - some teams want Kai to plan before acting, while others want fast, direct answers. Collaborative sessions where multiple employees iterate on the same AI outputs together are on the roadmap.
For professionals building production AI agents, Deep Agents is LangChain's open-source agent harness. The team's experience suggests that starting with a solid foundation layer for tool-calling loops, middleware composition, and state management can dramatically reduce time-to-deployment. Professionals looking to build similar internal tools can explore AI Agent Courses that cover agent architecture and deployment patterns. Software engineers interested in the underlying stack may also find relevant material in AI Engineering Courses.
Why this matters for creatives, support teams, and business professionals
Kai's adoption pattern is instructive: the tool gained traction fastest among people who don't write code. Salespeople, finance analysts, and marketing teams adopted it at higher rates than engineers. The lesson for any team evaluating AI tooling is that adoption depends less on technical capability and more on whether the tool removes friction from existing workflows. Kai succeeded because it didn't ask users to learn a terminal, manage context windows, or configure model parameters. It met them in a chat interface with their company's context already loaded. For customer support, HR, sales, and marketing professionals, the bar for useful AI is not raw model performance - it's whether the tool reduces the number of steps between a question and a usable answer.