The Enterprise AI Playbook: Lessons from Stripe’s Kai

Stripe engineering manager Sharadh Krishnamurthy explains how to build AI systems within a large company. His main point is that successful enterprise AI is not a product challenge, but a matter of governance and infrastructure. Deploying AI agents reveals existing architectural weaknesses, so companies that have already invested in developer tools and stable data platforms will have a significant advantage. Engineering leaders and product managers should use this as a guide to move past simple chatbots and build a company-wide agent platform by solving governance issues early.

Key Insights & Analysis

The Infrastructure-First Precondition

Most companies treat AI as a layer to add onto their current systems. Krishnamurthy’s work at Stripe shows this is a mistake. Stripe’s internal agent, Kai, works for over 10,000 employees because of work done years before the current AI boom. Stripe invested in developer platforms, standardized data querying with Trino, and thorough data cataloging.

When you deploy an agent, you are adding an entity that interacts with your entire stack. If your data warehouse is not stable, an agent will overwhelm it until it fails. If your developer experience is messy, the agent will inherit that mess. Teams that skip foundational work often end up fixing agent-induced outages instead of building new features.

"It turns out that agents just dial up all your failure modes, it just multiplies the amplitude of problems you can get. There were agents that went rogue, there were agents that may have almost taken down core systems."

-- Sharadh Krishnamurthy

Governance as a Competitive Moat

A key insight is using Projects as a governance tool rather than just a folder structure. Many companies treat security and compliance as obstacles to AI. Stripe does the opposite: they use Projects to define tool policies and model routing, creating a sandbox where employees work safely without needing to configure every permission manually.

This provides a lasting advantage. Because governance is built into the project structure, a small team of fewer than 10 people can manage a system used by 10,000. The system responds to user intent rather than just a prompt, reducing the cognitive load on employees who can trust that the agent stays within safe boundaries.

The Three-Layer Triage for Data Agents

When building data agents, many assume the AI should have access to raw databases to reason through them. Krishnamurthy suggests a more structured approach: a three-layer triage. The agent should first look for existing reports, then query the vetted analytics layer, and only access the raw data catalog as a last resort or when specifically authorized.

This mimics how a human analyst works and protects the system from inefficient queries. By forcing the agent to prioritize high-level abstractions, the company ensures the AI is not guessing at table structures, which lowers the risk of hallucinations and performance issues.

"I love this idea of this like three layer triage that a data agent can go through and that's really smart. Find existing reports then use the analytics layer to find the right query and then if like you really have to fall down to the data catalog and write your own query, your data warehouse has to be very resilient to high volume queries."

-- Claire Vo

The Necessity of Skill Lifecycle Management

As an enterprise scales to thousands of AI skills, the system faces a quality problem. If an agent is cluttered with 2,000 outdated or redundant skills, its performance and ability to find the right context will suffer. Krishnamurthy notes that Stripe treats skills like production code, using telemetry to find failing tool calls and implementing policies to remove unused workflows. This prevents the context rot that makes many internal AI tools unusable over time.

Key Action Items

• Audit your Data Layer (Immediate): Before scaling AI agents, ensure your data warehouse can handle high-volume, automated queries. If your infrastructure struggles with manual use, it will fail under agentic use.
• Implement Project Governance (Next Quarter): Stop using global AI settings. Group AI access by project or team to define specific tool policies and model budgets at the group level.
• Adopt a Triage Hierarchy (Next Quarter): Structure data agents to search for reports and analytics layers before accessing raw data schemas to reduce hallucination risk and improve performance.
• Establish a Skill Deprecation Policy (12-18 Months): Treat AI skills like code. If a skill is not used for 30 days, move it to a deprecation queue to prevent context bloat.
• Build for Human-in-the-Loop (Immediate): Identify high-risk actions like sending emails or modifying production databases and require an explicit human confirmation step. This adds friction but prevents major errors.
• Invest in Developer Experience (Ongoing): Treat developer experience as the foundation for your AI strategy. The more efficient your human-facing tools are, the more effective your agentic tools will be.