Intuit Redesigned AI Agent Architecture Twice in Four Months

Intuit, renowned for its financial software solutions, has aggressively adopted agentic AI technologies, but achieving a stable architecture proved challenging. Over four months, the team twice overhauled its approach: first shifting from specialized agents to a centralized orchestrator, then abandoning it for a modular system based on skills and tools.

The orchestrator’s flaw lay in agents exchanging data in natural language, losing context at each stage. Errors compounded, making the system unreliable. As Nhung Ho, Intuit’s AI vice president, noted, a chain of ten agents didn’t just occasionally fail—it was inherently error-prone due to the architecture itself.

The transition to the new system took just 60 days, with a working version live in 20. Success hinged not only on architectural changes but also on internal communication. To win over leadership and engineers, the team created a demo using real customer requests, proving the new approach’s effectiveness. For developers, the ability to scale skills and tools across all users became a compelling argument.

The new architecture also transformed feedback mechanisms. Previously, customer feedback was sparse and polarized, but now every AI agent interaction generates data. This accelerates the identification of weaknesses and system improvements. Additionally, users can now include live specialists—Intuit employees, accountants, or personal advisors—in conversations while maintaining full dialogue context.