How Kai’s three-layer data-query workflow works

Stripe’s internal AI agent Kai creates and iterates on data dashboards by retrieving an “Ask Data” skill, querying Stripe’s Hubble data layer backed by Trino, and using a secure cloud sandbox to process data and generate dashboard artifacts.

• Kai’s data-query workflow routes requests through three layers: existing direct artifacts and reports first, Stripe’s “blessed” analytics layer second, and the data catalog and underlying datasets only as a final fallback.
• Stripe built Kai’s V0 with one and a half engineers over two weeks; the core team managing the experience remains under 10 people while more than 10,000 employees use Kai weekly and “86 plus%” of the company uses it.
• Kai has approximately 2,000 skills; users can convert a chat session into an OpenSpec skill, keep it private or publish it to an area, and project-level tool policies can require human approval for actions such as creating a calendar invite.

Anish Acharya describes AI-native companies as cascading loops: agents with tools, memory, and skill files run tasks within functions, while humans provide out-of-distribution ideas when automated hill-climbing reaches a local maximum.

• A coding loop can take a bug report through repro generation, bug-fix creation, review, risk assessment, production deployment, and a customer email; Anish Acharya said this loop can complete in 5 minutes.
• For a growth team, the loop generates and measures every experiment variant, automatically merges a variant after it reaches statistical significance with a sufficiently high P value, maintains a long-term holdout, and starts the next experiment.
• Kavak runs a 6-week “Jedi Academy” that teaches employees, including mechanics, to use new tools and technologies; participants ship an in-production agent at the end of the course, and Kavak’s customer agents call a human when stuck, capture the resulting traces, and use them for later cases.

Brilliant’s AI tutor Cooji guides students through difficult problems without supplying answers or direct explanations, while human learning designers define lesson sequences and AI implements them using reusable interactive primitives.

• In a recorded fractions session, a student needed to shade 5/12 of a shape; Cooji guided the student to recognize that five shaded pieces could have different values rather than explaining the answer directly.
• Brilliant uses a library of modular, composable primitives with APIs that let models create, annotate, highlight, and ask subquestions in interactive lessons; the structure provides deterministic grounding instead of having the model generate an entire interactive experience from scratch.
• Before incremental user rollout, Brilliant runs evaluations for correctness, physics constraints, visual overlap, and tap-target size, then sends synthetic students through roughly 1,000 tutoring sessions; learner success is measured by assessment performance and retention rather than completion rate or time spent.