Intuit's AI VP: Rebuilding Agent Architecture Twice Was the Fast Path
Intuit's Vice President of AI, Nhung Ho, revealed at VB Transform 2026 that the company had to completely rebuild its agent architecture twice within a four-month period. Initially, Intuit moved from a system of specialized agents to a central orchestration layer designed to manage tasks without customer intervention. However, this orchestrator proved too complex, leading to context loss as agents communicated in natural language, compounding errors with each handoff. The failure mode was so significant that a full second rebuild was initiated, taking 60 days to complete, with a functional version ready in under 20 days.
The company then transitioned to a skills and tools-based architecture. This decision was supported by demonstrating to leadership that the new system performed better on real customer queries. Convincing the engineering team involved reframing their work from building standalone agents to developing reusable skills and tools, emphasizing the scalability benefits. This shift also redefined team responsibilities towards evaluation metrics. A key outcome of the rebuild is an early-stage feature allowing customers to bring human support agents into AI conversations, providing seamless context transfer. This contrasts with typical AI assistants that end with a recommendation to consult a professional. Intuit's system also incorporates a robust permissions model for financial data and maintains an audit log of all agent actions, with mechanisms for reversal. The company also experienced a dramatic increase in feedback, with nearly every conversation now serving as data, moving from a sparse, bimodal feedback rate to near 100% engagement. Ho has personally returned to coding to build models that systematically analyze this vast amount of direct customer feedback to identify and address system shortcomings.
Intuit's experience highlights the inherent challenges in developing complex AI agent systems, particularly concerning emergent complexity and inter-agent communication. The initial reliance on natural language for agent-to-agent context transfer proved to be a fragile design, susceptible to compounding errors. The rapid iteration and architectural pivots, while disruptive, demonstrate a pragmatic approach to problem-solving driven by performance and customer needs. This iterative process underscores the dynamic nature of AI development, where initial architectural assumptions may quickly become obsolete. The shift towards a skills and tools-based system, coupled with a focus on explicit permissions and human-in-the-loop capabilities, suggests a maturing understanding of how to build more robust, trustworthy, and user-centric AI applications, especially within sensitive domains like finance. The massive increase in actionable feedback data represents a significant opportunity for continuous improvement, provided the company can effectively scale its analytical capabilities to harness this input.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.