Intuit AI Agent Architecture Overhauled Twice in Four Months – VP Calls It the Fast Path

Intuit AI agent architecture faced a rapid evolution in early 2026, as the company scrapped and rebuilt its core system twice within four months. At VB Transform 2026, Nhung Ho, Intuit’s VP of AI, described a 60‑day sprint that delivered a functional prototype in under ↑ 20 days, after the first orchestration layer collapsed under its own complexity.

Intuit AI agent architecture: two rebuilds in 120 days

The initial shift moved from a fleet of specialist agents to a central orchestrator designed to hide agent selection from customers. The orchestrator survived roughly three months before a structural flaw emerged: agents exchanged results in natural language, causing context loss at each handoff. “If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,” Ho said.

“The error compounded by design, not by load,” Ho explained.

To stop the decay, the team abandoned the orchestrator and adopted a skills‑and‑tools model, treating each capability as a reusable component. Convincing leadership hinged on a live demo that processed real production queries and outperformed the legacy system. “The best proof is what customers are trying to do,” Ho asserted. Gaining engineer buy‑in required a scale argument: a shared skill serves every user, whereas a standalone agent solves only one narrow task. The new architecture also re‑engineered human‑in‑the‑loop handoffs; a live chat can now summon an Intuit support specialist, a user’s accountant, or a bookkeeper with full context. This contrasts with generic AI assistants that merely advise users to consult a professional. Permission controls now guard every action on financial data, and an immutable audit log records each step. Early testing shows the feature active for ↑ 1% of customers, with plans to expand. Feedback collection shifted dramatically: from a sparse 0.3% response rate (↓ 0.3%) to near‑universal input as each conversation becomes implicit feedback. Ho is personally coding models to parse this torrent of user comments, which range from blunt criticism to constructive correction. “They tell you, ‘You suck,’ but they also help fix it,” she noted. The overhaul positions Intuit at the forefront of agentic AI, a claim echoed by industry analysts in Reuters and Bloomberg.

Reported by: Julian Reed

Consumer Electronics Expert