At the VB Transform 2026 conference, Nhung Ho, Vice President of AI at Intuit, revealed that the company had to completely rebuild its AI agent architecture twice in just four months. This radical decision was made after the previous central orchestration layer exposed fatal flaws, directly impacting the accuracy of financial tasks. The second ground-up rebuild took the engineering team 60 days, with the first prototype delivered in under 20 days.
Detailed Developments
According to Nhung Ho, Intuit's AI agent development journey was far from smooth. Initially, to address customer complaints about manually managing too many specialized agents, Intuit built a central orchestration layer to automatically route tasks. This system operated stably for only about three months before suffering a structural collapse. The root cause was that the agents passed results to one another using natural language, leading to a 'telephone game' effect where errors compounded at each processing step. To resolve this, Intuit was forced to pivot back to a 'skills and tools'-based architecture, breaking down previously specialized agents into shared resources.
Technical & Technology Analysis
The core technical bottleneck lay in the communication mechanism between independent tasks. In the old hierarchical system, because agents communicated results using natural language, the succeeding agent had to infer context from its predecessor. With a chain of 10 agents, this error rate compounded exponentially, leading the system to deliver completely inaccurate results. The new architecture, built on shared skills and tools, allows system components to access APIs and functions directly and consistently. Furthermore, Intuit had to establish a strict authorization model for financial data, requiring explicit user approval before an agent takes any action, while maintaining an audit log to roll back states if necessary.
Expert Insights & Perspectives
Nhung Ho admitted that convincing both leadership and hundreds of engineers to abandon their previous work was an immense challenge. To secure buy-in from executives, her team built a prototype using real customer query data to demonstrate the superior performance of the new architecture. For the engineers, the most compelling argument was scalability. Instead of building a single agent that only solved a narrow problem, a shared tool or skill in the new architecture could serve any customer accessing that product segment. This shift also refocused the teams' efforts from building agents to running evaluation frameworks (evals).
Impact & Future Outlook
The most tangible result of this restructuring is a 'human-in-the-loop' feature that allows a real human agent to step directly into the AI agent's conversation with a customer, currently being tested with 1% of users. Instead of simply offering disclaimer-heavy recommendations and advising customers to find an expert—as typical AI assistants do—Intuit's system directly connects the user with an accountant right within the chat. The transition to a conversational interface has also enabled Intuit to collect feedback from nearly 100% of customers, compared to a mere 0.3% previously, ushering in a new era of real-time product refinement based on direct interaction.