Intuit was one an early pioneer agent in using AI, but its path to success has hardly been a straight line.
Horse VB Transform 2026Intuit VP of AI Nhung Ho described how the company rebuilt its agent architecture twice in about four months, first moving from a fleet of specialist agents to a central orchestration layer, then abandoning that layer for a skills- and tool-based system when the orchestrator itself failed under its own complexity. The second complete rebuild took 60 days, the first working version took less than 20.
The failure mode that forced the second rewrite was special. Agents in an orchestrated system communicate results to each other in natural language, and each handoff requires the next agent to act correctly, losing context.
"If you have 10 agents and they all switch to each other, every time this switch happens, error combinations," Ho said.
Why did the orchestra collapse?
Ho said the initial push towards specialist agents stemmed directly from customer complaint. A fleet of skilled agents is still something the customer must manage by deciding which agent to use for which task. Intuit’s answer was a system that could take a task and route it internally without requiring the customer to select an agent.
This layer of orchestration lasted about three months, which Ho only half-jokingly described as about a year in the compressed timeline of agent development in 2026.
It broke for a structural reason, not capacity. Passing results between agents in natural language meant how each downstream agent arrived at its result, and that result deteriorated with each additional hop. The ten-agent chain did not fail from time to time, increasing errors by design.
This diagnosis is what brought Intuit back to the skills and tools architecture.
What it took to get 60 days of rebuilding and engineering
Rebuilding the production agent system in 60 days required more than an architectural decision. Ho said the more difficult problem was internal, and convinced both management and the engineers who built the original agents that it was the right call to abandon recent work.
The leadership claim was based on evidence rather than argument. Ho’s team built a demo of the new architecture using real customer requests from production, then showed that it outperformed the existing system on the same tasks.
"The best evidence, at least my belief, is what are customers trying to do? And any system you build to solve these problems needs," Ho said.
Winning over engineering required another case. Hundreds of engineers outside of Ho’s core team had built retired specialist agents, and the request was to separate their agents into individual skills and tools.
Ho said the motivating argument is one of scale. An independent agent has solved a narrow problem, while a common skill or tool built into the new architecture can serve every customer that touches that part of the product. This shift also changed the day-to-day responsibilities of the partner teams, shifting their focus from building agents to executive evaluations, as evaluations became the only way to measure whether the new architecture was actually working.
Bringing a person into the loop and providing feedback on a different scale
The most obvious customer-facing result of the redesign is the human-engaging feature of live agent chat — which, while still in early testing, covers about 1% of Intuit’s customer base. "We’ll be expanding it over the next few weeks," he said.
A customer can bring in an Intuit product support person, or their accountant, or one of Intuit’s own accountants mid-conversation, and that person joins the full context of what the agent is already doing, Ho said.
Ho drew a direct contrast to how most AI chat products handle the same situation. A general-purpose assistant who answers a tax question usually ends with a refusal to consult a specialist. Intuit’s system is designed to connect a customer directly to that professional within the same conversation.
This human handover sits alongside an authorization model built specifically for financial data. Every action an agent takes on a customer’s financial data requires express authorization in the first place, though Ho said the requirement could decrease over time as customers build trust in the system. Intuit keeps an audit log of everything the agent does and can be rolled back if needed.
Agency feedback in the age of artificial intelligence
The restructuring also changed the way Intuit collects and uses feedback, which Ho says is qualitatively different than before.
"In the past feedback was very, very sparse and also very bimodal," Ho said. "They either loved it or hated it, and were generally biased toward the negative."
In a chat-based system, every conversation serves as feedback, which Ho says has moved the company from about 0.3% of customers to something close to 100% giving open feedback.
Ho said he went back to writing code himself specifically to create models that systematically analyzed the volume of feedback, looking for system failures at a scale that no manual review process could handle.
This volume comes with a tone that most product teams are not used to hearing directly. Customers let the agent know exactly where they failed.
"They tell you right that you are bad. I hate it. It’s not right”" Ho said. "But they’re also willing to give the systems grace and tweak it, and so it’s up to all of us to gather that new type of feedback and feedback and really improve the system."
