Build your own company brain: the enterprise AI playbook from Stripe’s engineering team | Sharadh Krishnamurthy

The episode begins with Sharadh Krishnamurthy introducing himself and explaining Stripe's motivation behind building Kai, their internal AI agent. Stripe aimed to create an AI system that would serve over 10,000 employees weekly, built from scratch rather than bought off-the-shelf, to tailor governance, context-awareness, and security to enterprise needs. The discussion highlights the importance of infrastructure, governance, and embedding AI deeply into organizational workflows, emphasizing that their approach leverages their existing investments in developer experience and data infrastructure.

The conversation delves into the technical architecture of Kai, including its context-awareness—knowing who the user is and their organizational role—and its integration with Stripe’s core systems to ensure safe, scalable queries. Key investments in resilience at the data layer, comprehensive data cataloging, and tiered access are emphasized. These foundational elements allow Kai to perform high-volume queries safely while enabling flexible, user-controlled access and privacy management. The importance of infrastructure hardware, such as load shedding and failure mode mitigation, is also discussed, illustrating the depth of engineering effort behind Kai.

A major focus is on how Stripe uses 'projects' as a governance layer—defining boundaries, permissions, and safety controls—rather than simply organizing workflows. Projects allow specific tool policies, such as restricting access to sensitive data or requiring human approval for certain actions. This setup ensures safe automation, especially when dealing with confidential information. Sharadh emphasizes how this layered, purpose-built management keeps AI safe and aligned with enterprise policies, enabling more bespoke, scalable usage of AI across varied business units.

The episode features live demos showing how Kai is used for creating dashboards, iterating on data visualizations, and packaging workflows into reusable skills. These examples illustrate how non-technical users can collaborate with Kai through multi-turn conversations, turning simple prompts into complex data artifacts like dashboards and insights. The platform's UI and management layers support editing, sharing, and evaluating skills, making it accessible to teams beyond technical experts. This demonstrates Stripe’s goal of democratizing AI-powered productivity tools.

Stripe’s lean team managed to deploy Kai rapidly—within a few weeks with less than 10 people—initially through internal enthusiasm and gradual expansion. The episode discusses strategies for encouraging organic adoption across teams, the importance of infrastructure investments in enterprise AI, and how to manage a large library of skills with telemetry, quality assessments, and deprecation policies to keep the system current and efficient. Sharadh notes that the success of Kai is rooted in their pre-existing data and developer platforms, which provide the leverage for AI integration.

Closing remarks touch on lessons learned from building Kai, including the need for continuous iteration and improvement. Sharadh acknowledges that they are still in early days of enterprise AI development, with many challenges ahead. The episode ends with personal insights from Sharadh about his own AI use, including using Kai as a personal assistant and for entertaining his children’s questions about the universe and dinosaurs. He advocates for gentle prompting and respectful human-AI interaction, emphasizing that the enterprise tools require ongoing evolution.