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

The podcast discusses the development and implementation of AI agents within organizations, focusing on challenges, governance, and practical applications. AI agents can amplify system vulnerabilities by brute-forcing queries or bypassing human oversight, necessitating strong governance. Key strategies include tool policies, a three-layer triage system for data access (existing reports, analytics layer, data catalog), and human-in-the-loop workflows. Stripe's custom AI agent, Kai, exemplifies these principles by incorporating context awareness, user-controlled data access, and secure, cloud-hosted infrastructure.

AI agents are used to streamline data workflows, particularly in creating dashboards and handling "last-mile" data tasks, enabling non-technical users to build lightweight applications. The system supports skill creation and reuse through a project-based model, where projects act as governance units that define access controls, model usage, and team-specific configurations. Skills - packaged sets of tools - are automatically documented, monitored for quality, and subject to deprecation if unused. A secure sandbox environment allows agents to execute code safely, while resilient data infrastructure ensures stability under high query loads. The approach emphasizes balancing automation with security, scalability, and usability across diverse enterprise needs.

• What if you built a project-governed AI skill that automates your most frequent data tasks with built-in cost controls?

  • Move: Identify your top 3 recurring data workflows (e.g., report generation, dashboard updates), then create a reusable AI skill for each inside a project-scoped environment with enforced model routing (e.g., use cheaper models by default).
  • Why Now?: AI agents can drift into expensive or insecure queries without governance - by codifying workflows now, you prevent cost creep and errors as usage scales.
  • Expected Upside: Reduce execution time by 70%+ per task and cut AI spend by constraining model usage, while creating auditable, shareable workflows you can reuse or monetize.

• What if you implemented a three-layer triage system for your personal data queries to protect your infrastructure?

  • Move: Structure your data access so AI first checks existing reports, then uses an analytics layer (e.g., pre-built views), and only queries raw data as a last resort - enforce this via a simple routing script or agent prompt template.
  • Why Now?: Unchecked AI agents can brute-force queries and overload personal or small-scale databases - triage prevents burnout on your backend before it becomes critical.
  • Expected Upside: Achieve 10x query efficiency, reduce load on your data store, and ensure faster, more reliable responses by prioritizing optimized paths.

• What if you sandboxed your AI agent in a secure, user-controlled environment where you dictate data access per session?

  • Move: Wrap your AI agent in a lightweight sandbox (e.g., Docker or isolated runtime) that requires explicit, per-session opt-in to access sensitive tools or data (e.g., Google Drive, CRM).
  • Why Now?: Rogue agents can leak data or trigger unintended actions - giving you session-level control now prevents breaches as your automation grows.
  • Expected Upside: Gain trust in your AI workflows, enable safe experimentation, and maintain compliance with personal or customer data boundaries - critical if you plan to productize your agent.

• Implement a tiered data access system (existing reports analytics layer data catalog) to reduce load on your data warehouse and prevent brute-force queries from AI agents.
• Build AI agent workflows within a secure sandbox environment that isolates execution, allowing safe code and query testing without risking production systems.
• Design project-based governance for your AI tools, where each project defines model usage, tool access, and permissions, enabling scalable and secure configuration across different user groups.
• Create reusable, documented skills for common tasks (e.g., generating dashboards or running SQL queries) and implement automated quality suggestions to maintain reliability and performance over time.
• Prioritize investment in your data and developer infrastructure (e.g., resilient querying layers like Trino) before scaling AI features, ensuring systems can handle high-volume agent-driven workloads.