Intuit, an early pioneer in agentic AI, revealed at VB Transform 2026 that it rebuilt its entire AI agent architecture twice in just four months before finding a sustainable approach. The journey offers a cautionary tale for enterprises rushing agent systems into production.
The First Rebuild: Specialist Agents to Orchestration
The company initially deployed a fleet of specialist agents but faced a common customer complaint—users still had to decide which agent to use for each task. Intuit’s solution was an orchestration layer that could route tasks internally without customer intervention.
That architecture held for approximately three months before failing. The problem wasn’t capacity but error compounding through natural language handoffs. When agents passed results to each other in natural language, each downstream agent had to infer how the previous one reached its conclusions—and that inference degraded with each additional hop.
“If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,” explained Nhung Ho, Intuit’s VP of AI.
The Second Rebuild: Skills and Tools
Intuit abandoned the orchestration approach entirely, rebuilding in 60 days (with a first working version in under 20) using a skills and tools architecture. The key insight: rather than passing context between agents, decompose capabilities into reusable skills and tools that any agent can call directly.
The rebuild required convincing both leadership and the engineers who had built the original specialist agents. Ho’s team built a demo using real production customer queries, demonstrating better performance than the existing system—a more persuasive argument than any architectural discussion.
Human-in-the-Loop and Feedback at Scale
The new architecture enables a feature currently in early testing with 1% of customers: bringing a human into the conversation mid-stream. Customers can pull in an Intuit support person, their own accountant, or a bookkeeper with full context of what the agent has already accomplished.
Perhaps most significantly, the shift transformed feedback collection. From roughly 0.3% of customers providing explicit feedback in traditional chat systems, Intuit now collects feedback from nearly 100% of conversations. The tone is direct: “They straight up tell you, ‘You suck. I hate this. This is not right.’” But customers also correct the system and give it grace—creating an unprecedented volume of actionable improvement data.
The lesson for enterprise AI: the fastest path forward sometimes requires the courage to start over.