Key Takeaways

  • Stripe launched the initial prototype of Kai, its internal AI agent, with 1.5 engineers in two weeks.
  • More than 10,000 Stripe employees use Kai every week, maintained by a core team of fewer than 10 people.
  • Vendor AI tools fail inside complex companies because they lack native access to org charts, OKRs, project trackers, and internal email archives.
  • Stripe treats Kai as a live proving ground to harden agent infrastructure before shipping AI tools to external customers.

Why Off-the-Shelf Tools Fail at Company Context

Most enterprise AI buyers think their biggest challenge is model quality. It is not. The bottleneck is context.

When Stripe evaluated third-party AI assistants, the tools stumbled on basic organizational realities. A generic bot does not know who reports to whom, which team owns a disputed service, or which project relates to an executive OKR. As Sharadh Krishnamurthy explained:

“The harder problems are in trying to replicate the way a company works at scale and Stripe is an incredibly complex business... so we quickly realized that it's not about providing AI, it's about providing the correct governance structures so that everyone can just go use AI and know it'll do the right thing for them.”

Instead of bending an external tool to fit internal security boundaries, Stripe connected its own data graph directly to Kai. The agent knows the user's role through the org chart and can search project management tools, shipped update emails, and active roadmaps without leaking restricted data across teams.

The Two-Week Rule for Internal Tools

Founders often stall AI projects by over-planning architecture, writing lengthy security reviews, and comparing dozen-vendor feature matrices. Stripe moved in the opposite direction.

“It took us like one and a half people over two weeks to get Vzero out the door,” Krishnamurthy noted. “Something we realized was a lot of the answers became apparent once we could show people something.”

By putting a rough V0 in front of real employees immediately, the team stopped debating hypothetical edge cases. They saw exactly where users hit errors, which permissions broke, and what standard queries employees actually asked. That feedback loop kept the engineering footprint small: even with over 10,000 weekly active users across technical and non-technical roles, the core team maintaining Kai remains under 10 people.

Dogfooding as an Infrastructure Strategy

Building an internal agent is not just an operational shortcut; it is product research. If your company plans to sell AI capabilities to customers, running your own internal agent exposes the real failure modes of agent architecture long before your clients see them.

Claire Vo highlighted this feedback loop during the conversation:

“It's like the perfect dogfooding agent experience, which is if you're going to build a great agent for your customers, you should learn to build a great agent for yourself. And then it sounds like you're reusing some of that infrastructure, which is nice because then you can pressure test it against Stripe employees or external customers.”

When you build internally, your employees become your strictest quality assurance team. If an agent hallucinates a database schema or misinterprets an API change, an internal engineer catches it in minutes.

What to Do With This

Stop running three-month vendor evaluations for internal AI tooling. Assign one engineer to spend next week wiring a private LLM endpoint to your team's primary documentation store and internal directory. Put the crude interface in front of five teammates on Friday, log their failed queries, and fix the context mapping before writing another spec.