Intuit Overhauled Its AI Agent System Twice in Four Months

Intuit, the company behind TurboTax and QuickBooks, has been an early adopter of AI agents. However, its journey to a working system was not smooth. At a recent tech conference, Intuit's AI VP Nhung Ho revealed that the company rebuilt its AI agent architecture twice in about four months. Initially, Intuit used a fleet of specialist agents. But customers had to decide which agent to use, so Intuit added a central orchestration layer to route tasks automatically. That worked for about three months, then broke down. The problem was that agents passed results to each other in natural language. Each handoff lost context, causing errors to build up. A chain of ten agents would compound mistakes every time. Intuit scrapped the orchestration layer and rebuilt using a skills-and-tools approach. The second rebuild took 60 days, with a working version in under 20. To get approval, Ho's team built a demo using real customer queries that outperformed the old system. The new system also allows a live agent conversation to bring in a human, like an Intuit support person or the customer's own accountant. This feature is in early testing with about 1% of customers. Feedback has changed too. In the past, only 0.3% of customers gave feedback. Now, every chat conversation provides feedback, giving Intuit a much richer data set to improve the system. Customers are honest, often saying things like, "You suck." But they also help correct mistakes. Ho now writes code herself to analyze the massive feedback volume, hoping to catch problems that manual review cannot.