Intuit rebuilt its AI agent stack twice as orchestration broke down

Intuit rebuilt its production AI agent architecture twice in roughly four months after a central orchestration system created compounding errors, VP of AI Nhung Ho said at VB Transform 2026. The shift matters for enterprise software teams betting on agentic AI because Intuit found that routing work among specialist agents did not hold up once agents had to pass context through multi-step workflows.

Ho said Intuit first moved from separate specialist agents to a centralized orchestration layer so customers would not have to decide which agent should handle a given task. That system lasted about three months before the company replaced it with an architecture based on reusable skills and tools. The second rebuild took 60 days, with an initial working version ready in less than 20 days, according to Ho.

Why the orchestrator failed

The problem was not described as a compute limit or scale issue. Ho said the orchestrated system used natural language handoffs between agents, which meant each agent had to interpret what the prior agent had done and why. As the workflow lengthened, each handoff increased the chance that necessary context would be lost.

“If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,” Ho said.

That diagnosis pushed Intuit away from agent-to-agent orchestration and toward a model in which capabilities are broken into smaller skills and tools that can be reused across product workflows. Ho said the earlier specialist-agent approach had addressed narrow tasks, but it also created a management burden for customers and a maintenance problem inside the company.

The internal sale was part of the work

Ho said the technical redesign was only one part of the process. The AI team also had to persuade leadership and the engineers who had built the specialist agents that recently shipped work should be dismantled and repackaged.

For executives, Ho said her team used production customer queries to compare the new architecture against the system already in place. “The best proof, at least my belief, is what are customers trying to do? And whatever system you build needs to address those problems,” she said.

The engineering argument was different. Ho said hundreds of engineers outside her core team had contributed to the existing agents. The new approach asked those teams to turn agents into individual skills and tools, with a broader potential surface area across Intuit’s products. It also changed their operating model: partner teams shifted more attention from building agents to running evaluations, because evals became the main way to determine whether the revised system was improving.

Human handoffs and controls

One customer-facing result of the rebuild is a human-in-the-loop feature now in early testing with about 1% of Intuit’s customer base. Ho said Intuit plans to expand it in the coming weeks.

The feature allows a customer to bring an Intuit support representative, the customer’s accountant, or one of Intuit’s bookkeepers into an active agent conversation. Ho said that person joins with the context of the interaction to that point, rather than starting from a blank support handoff.

Ho contrasted that with general AI assistants that answer tax or finance questions and then advise users to consult a professional. Intuit’s design, she said, is meant to bring the professional into the same conversation.

Ho also said Intuit requires explicit customer approval before an agent acts on financial data. The company keeps an audit trail of agent actions, and actions can be reversed if needed. Ho said permission requirements may loosen over time as customers become more comfortable with the system, but did not provide a timeline or thresholds for that change.

Feedback volume changed the product loop

Ho said chat-based AI has also changed Intuit’s feedback system. Previously, she said, about 0.3% of customers provided explicit feedback, and that feedback tended to cluster at positive or negative extremes. With conversational systems, she said, nearly every interaction becomes feedback.

Ho said she has returned to writing code to build models that analyze the resulting volume, looking for recurring failures that manual review could not process at the same scale.

That feedback can be blunt. “They straight up tell you, ‘You suck. I hate this. This is not right,’” Ho said. She added that customers also correct the system, creating a feedback stream Intuit is trying to use to improve agent performance.

This story draws on original reporting from VentureBeat.