Intuit was an early adopter of agentic AI, but its route to a reliable production system was not linear. At VB Transform 2026, Nhung Ho, Vice President of AI at Intuit, explained that the company rebuilt its agent architecture twice over roughly four months.
From specialists to an orchestrator, then back to skills and tools
Initially Intuit moved from a fleet of specialist agents to a central orchestration layer designed to route tasks internally so customers wouldn’t need to choose agents themselves. That orchestration solution worked for about three months — a period Ho half-jokingly described as roughly a year in the accelerated timeline of 2026 agent development.
The orchestrator failed for a structural reason rather than a capacity one: agents passed results to each other in natural language, and each handoff required downstream agents to infer how upstream agents reached their conclusions. According to Ho, with multiple handoffs the inference errors compound — a chain of ten agents didn’t fail occasionally, it degraded by design.
That diagnosis prompted Intuit to revert to an architecture built around shared skills and tools.
The 60-day rebuild and getting internal buy-in
The full second rebuild took 60 days, with a first working version available in under 20 days. Ho emphasized that beyond the architectural choice, the harder problem was convincing leadership and the many engineers who had built the retired specialist agents that scrapping recent work was the right call.
To persuade leadership, Ho’s team relied on evidence: they built a demo of the new architecture using real customer queries from production and showed it outperformed the existing system on the same tasks. To win over engineering, they argued scale: a standalone agent solved a narrow problem, but a shared skill or tool embedded in the new architecture could serve every customer touching that product area. The partner teams’ day-to-day responsibilities shifted from building agents to running evals, which became the primary way to measure whether the new architecture actually worked.
Human-in-the-loop and permissions for financial data
A clear customer-facing result of the rebuild is a feature that allows a live agent conversation to pull in a human — currently in early testing and live to about 1% of Intuit’s customer base, with plans to scale in the coming weeks, Ho said.
A customer can bring in an Intuit product support person, their own accountant, or one of Intuit’s bookkeepers; the person joining the conversation receives the full context of what the agent has already done. Ho contrasted this with most general AI chat products, which typically end with a disclaimer to consult a professional. Intuit’s approach connects the customer directly to that professional inside the same conversation.
This handoff is coupled with a permissions model tailored to financial data: every action an agent takes on a customer’s financial data requires explicit permission first. Ho said that permission requirements can ease over time as customers build trust, and Intuit keeps an audit log of everything an agent does that can be reversed if needed.
Feedback at a different scale
The rebuild also changed how Intuit gathers and uses feedback. Historically feedback was very sparse and bimodal — users either loved the product or hated it, with negativity more common. In the chat-based system, every conversation functions as feedback, shifting Intuit from roughly 0.3% of customers ever giving explicit feedback to something close to 100%.
That volume requires automated analysis; Ho has returned to writing code to build models that systematically analyze the feedback at scale and surface where the system is failing in ways manual review cannot keep up with. The directness of the feedback also means teams now hear blunt user reactions — users tell the agent plainly where it failed — but they are often willing to give the system a chance to correct itself.
Conclusion
Intuit’s experience shows how architectural choices around inter-agent communication can create systemic failure modes: natural language handoffs caused context loss and compounding errors in long chains. The company’s response combined a return to a skills-and-tools architecture, integrated human handoffs, stricter permissions and auditability for financial data, and large-scale automated feedback analysis, justified by measurable improvements on real customer queries.
