Stripe's Kai: Enterprise AI Governance Framework
Stripe's engineering leader reveals how they built Kai, an internal AI agent serving 10,000+ employees with a team of fewer than 10. This analysis covers the critical role of project-based governance, tool policies, and infrastructure hardening in scaling enterprise AI adoption safely and effectively.
The Strategic Shift to Governed AI
Stripe’s internal AI agent, Kai, demonstrates that the primary barrier to enterprise AI adoption is not model capability, but governance. Built by a team of fewer than 10 engineers, Kai serves over 10,000 employees by solving the complex problem of replicating corporate structure in an AI context. The core insight is that AI deployment must be treated as an infrastructure and governance challenge, not merely a software feature.
Governance Through Project Boundaries
A key differentiator of Kai is the use of "projects" as a governance layer. Unlike generic chatbots, Kai uses projects to define specific tool policies, model routing, and permission sets for distinct workflows. For example, HR teams can operate with restricted tool access to prevent data leakage, while engineering teams can utilize broader permissions. This context-aware approach minimizes friction for users while maintaining strict security boundaries, allowing a small core team to manage risk across a massive organization.
Infrastructure Resilience and Data Triage
The transcript highlights that agents "dial up all your failure modes," often brute-forcing data queries and stressing infrastructure. Stripe’s success relies on pre-existing investments in data platform resilience and a three-layer triage system: checking existing reports, then the analytics layer, and finally the raw data catalog. This hierarchical query strategy ensures accuracy and protects the data warehouse from being overwhelmed by agentic behavior. The lesson for leaders is clear: harden your data infrastructure before deploying agents, as they will exploit any weaknesses in query handling or load management.
Democratization via Skill Platforms
Kai transforms one-off AI interactions into repeatable workflows through a skill-building platform. Employees can package successful sessions into skills, which are then shared and governed across the company. This creates a feedback loop where the organization’s collective knowledge is encoded into the AI system, improving efficiency and standardizing best practices. The platform includes automated suggestions for skill improvement, ensuring that the library of automations remains high-quality and relevant.
Conclusion
The success of Kai underscores that enterprise AI value is derived from the integration of AI with existing organizational structures and infrastructure. Companies must invest in governance frameworks, data resilience, and skill management platforms to scale AI safely. The focus should shift from simply adopting AI tools to building the rails that allow AI to operate effectively within the complex realities of large-scale business operations.
Key insights
-
Enterprise AI requires project-based governance to manage permissions and tool access dynamically. This allows specific workflows to have tailored security policies without imposing universal restrictions on all users.
Impact: Reduces security risks and operational friction, enabling faster adoption of AI tools across diverse departments.
-
Agents significantly amplify infrastructure failure modes, often brute-forcing data queries and stressing systems. Data warehouses must be hardened to handle high-volume, unpredictable agentic traffic.
Impact: Prevents system outages and data corruption, ensuring reliable AI performance in production environments.
-
A three-layer data triage system (reports, analytics, catalog) improves agent accuracy and efficiency. This hierarchical approach prevents agents from generating incorrect queries against raw data.
Impact: Enhances data quality and reduces compute costs by optimizing query paths for AI agents.
-
Skill-sharing platforms enable non-technical employees to create and share reusable automations. This democratizes AI capabilities and builds a collective knowledge base for the organization.
Impact: Increases overall organizational efficiency and reduces the dependency on specialized engineering resources for routine tasks.
-
Pre-AI investments in developer and data platforms provide extreme leverage for AI adoption. Strong foundational infrastructure allows agents to operate more effectively and safely.
Impact: Maximizes the return on AI investment by leveraging existing high-quality internal tools and data structures.
Action items
-
Implement project-based governance structures for AI agents, defining specific tool policies and permission sets for different workflows. This ensures that sensitive data is protected while allowing flexibility for other tasks.
Impact: Enhances security and compliance while maintaining user productivity and reducing the need for manual oversight.
-
Audit and harden data infrastructure to withstand high-volume, unpredictable queries from AI agents. Focus on load shedding, resilience, and query optimization to prevent system failures.
Impact: Ensures stable AI performance and protects critical data assets from agentic stress and potential outages.
-
Develop a hierarchical data query strategy for AI agents, prioritizing existing reports and analytics layers before accessing raw data catalogs. This improves accuracy and reduces the risk of incorrect data retrieval.
Impact: Increases the reliability of AI-generated insights and reduces the computational burden on data warehouses.
-
Build a skill-sharing platform that allows employees to package and share successful AI workflows. Include automated quality checks and suggestions to maintain the integrity of the skill library.
Impact: Accelerates the adoption of best practices and creates a scalable repository of organizational knowledge and automations.
-
Invest in and optimize internal developer and data platforms to support AI integration. Ensure that existing tools and data structures are agent-ready to maximize the effectiveness of AI deployments.
Impact: Leverages prior investments to enhance AI capabilities, leading to faster and more secure AI adoption across the organization.
Quotes
“It turns out that agents just like dial up all your failure modes. It just multiplies the amplitude of problems you can get.”
“The harder problems are in trying to replicate the way a company works at scale.”
“Agents are incredibly good at figuring this out. However, the key part, and you asked about the Ask Data skill, the key part is we have some really smart data scientists as well who sort of said, hey, this is probably the right way that most data queries should be handled.”