Intuition was a early pioneer in the use of agent AI, but its path to success has not been a straight line.
In VB Transformation 2026Nhung Ho, vice president of AI at Intuit, described how the company rebuilt its agent architecture twice in a span of about four months, first moving from a fleet of specialized agents to a central orchestration layer and then abandoning that layer for a skills- and tools-based system once the orchestrator itself began to fail due to its own complexity. The second full rebuild took 60 days, with a first working version in less than 20.
The failure mode that forced the second rewrite was specific. Agents in the orchestrated system passed results to each other in natural language, and each transfer lost the context that the next agent needed to act correctly.
"If you have 10 agents and they all pass each other, every time that pass occurs, the error gets worse," Ho said.
Why the orchestration layer broke
Ho said the original push toward specialty agents came from a simple customer complaint. A fleet of capable agents is still something that the customer must manage and decide which agent to use for each task. Intuit’s answer was a system that could take a task and route it internally, without asking the customer to choose an agent.
That layer of orchestration lasted for about three months, which Ho only half-jokingly described as about a year in the compressed timeline of the agent’s development in 2026.
It broke for a structural reason rather than a capacity one. Passing results between agents in natural language meant that each downstream agent had to infer how the upstream agent reached its conclusion, and that inference degraded with each additional hop. A chain of ten agents did not fail occasionally, but rather compounded design errors.
That diagnosis is what returned Intuit to an architecture of skills and tools.
The 60-day rebuild and what it took to get engineers’ buy-in
Rebuilding a production agent system in 60 days required more than an architectural decision. Ho said the hardest problem was internal, convincing both the leaders and the engineers who had built the original agents that discarding the recent work was the right decision.
The speech to the leadership was based on evidence rather than arguments. Ho’s team created a demo of the new architecture using real customer queries pulled from production and then demonstrated that it performs better than the existing system on the same tasks.
"The best test, at least in my opinion, is what are customers trying to do? And any system you build must address those problems," Ho said.
Earning engineering required a different case. Hundreds of engineers outside of Ho’s core team had built the specialized agents that were retiring, and the request was to separate his agents into individual skills and tools.
Ho said the motivating argument was scale. An independent agent solved a specific problem, while a shared skill or tool built into the new architecture could serve all customers who touched that part of the product. That shift also changed what partner teams were responsible for on a day-to-day basis, moving their focus from creating agents to running assessments, as assessments became the only way to measure whether the new architecture was actually working.
Bring a human into the loop and receive feedback on a different scale
The clearest result of the reconstruction for the customer is a feature that allows a conversation with a live agent to attract a human being; Although it is currently in early testing, it is live for about 1% of Intuit’s customer base. "We will expand it in the coming weeks," she said.
Ho said a customer can bring in an Intuit product support person mid-conversation, or their own accountant, or one of Intuit’s accountants, and that person joins in with the full context of what the agent has already done.
Ho drew a direct contrast to how most AI chat products handle the same situation. A general-purpose assistant answering a tax question usually ends with a disclaimer to consult a professional. Intuit’s system is designed to connect the customer with that professional directly, within the same conversation.
That human transfer is combined with a permissions model built specifically for financial data. Every action an agent takes on a customer’s financial data first requires explicit permission, although Ho said that requirement may decrease over time as customers build trust in the system. Intuit maintains an audit trail of everything an agent does that can be rolled back if necessary.
Feedback in the age of agent AI
The rebuild also changed how Intuit collects and uses feedback, a change that Ho said is qualitatively different from what came before.
"Feedback in the past used to be very, very sparse and also very bimodal," Ho said. "They either loved him or hated him, and he usually tends toward the negative."
In a chat-based system, every conversation functions as feedback, which Ho says took the company from about 0.3% of customers ever giving explicit feedback to something close to 100%.
Ho said she has rewritten code herself specifically to build models that analyze that volume of feedback systematically, looking for where the system is falling short on a scale that no manual review process could reach.
That volume comes with a tone that most product teams aren’t used to hearing directly. Customers tell the agent exactly where they went wrong, in simple terms.
"They tell you directly: ‘You suck.’ I hate this. This is not right.’" Ho said. "But they’re also willing to give the systems grace and fix them, so the onus is on all of us to collect this new piece of feedback and type of feedback, and really improve the system."
