Intuit, a pioneer in financial software, faced a complex journey in developing AI agents. At VB Transform 2026, Nhung Ho, Intuit's VP of AI, revealed that the company rebuilt its AI agent architecture twice in just four months. The first architecture used a fleet of specialist agents, then moved to a central orchestration layer, which was later abandoned for a skills and tools based system. The second full rebuild took 60 days, with a first working version in under 20.
The orchestration layer failure
The reason for the orchestration failure was structural. Agents in the orchestrated system passed results to each other in natural language, and each handoff lost context needed by the next agent. As Ho explained, "If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds." The orchestration held up for about three months but broke due to design: a ten-agent chain did not fail occasionally, it compounded errors by design.
Sponsored Protocol
The 60-day rebuild and internal buy-in
Rebuilding a production agent system in 60 days required more than an architectural decision. Ho said the harder problem was internal: convincing both leadership and the engineers who had built the original agents to scrap recent work. The pitch to leadership relied on evidence: Ho's team built a demo of the new architecture using real customer queries, showing better performance. Winning over engineers required a different case: the argument was scale. A standalone agent solved one narrow problem, while a shared skill or tool built into the new architecture could serve every customer. This shift changed partner teams' focus from building agents to running evals, since evals became the only way to measure whether the new architecture was working.
Sponsored Protocol
Human-in-the-loop and feedback at scale
The clearest customer-facing result is a feature that lets a live agent conversation pull in a human, currently in early testing with about 1% of Intuit's customer base. Ho said they will scale it in the coming weeks. A customer can bring in an Intuit product support person, their own accountant, or an Intuit bookkeeper mid-conversation, with full context. Unlike generic AI chatbots that end with a disclaimer, Intuit connects the customer directly to a professional. This sits alongside a permissions model for financial data: every agent action requires explicit permission, with an audit log. The rebuild also changed feedback collection. Previously, only 0.3% of customers gave explicit feedback; now every conversation serves as feedback, nearing 100%. Ho returned to writing code to build models that analyze this feedback volume systematically. Customers are blunt: "You suck, I hate this, this is not right," but they are willing to give grace if the system corrects itself.
Sponsored Protocol
For more on AI challenges, read the article on China's AI leap leaving Europe behind. Another tech failure story is the AWS billing glitch showing billion-dollar fees. For more about Intuit, see the Intuit Wikipedia page.
