Meet Stripe's Knowledge AI Platform

Stripe's Kai, an internal AI platform for non-engineering staff, treats the current task as a data-access boundary. An employee may legitimately access two customers' records separately, yet an analysis must not combine those unrelated customer contexts. Stripe's architecture account states this as an invariant the platform must enforce beyond checking who the employee is, without detailing the enforcement mechanism. It separately describes Kai's execution layer, which includes a shared access-control framework, per-session sandboxes, and a virtual filesystem where the agent builds artifacts. The design principle is that an assistant needs limits on which information belongs together in one task, alongside permissions governing which information a person may access.