Intuit Rebuilt Its Agentic AI Architecture Twice in Four Months

What happened

Intuit rebuilt its agentic AI architecture twice in the space of roughly four months — and its VP of AI, Nhung Ho, told attendees at VB Transform 2026 that the repeated scrapping was not a failure but the fastest route to a system that actually works.

The company's first pivot moved away from a fleet of specialist agents toward a central orchestration layer designed to coordinate them. That layer collapsed under its own complexity: agents passed results to one another in natural language, and every handoff degraded the context the next agent needed to act correctly. As Ho explained, with ten agents exchanging outputs, compounding error became inevitable. The orchestrator, meant to simplify coordination, instead amplified the problem.

The second rebuild replaced the orchestration model with an architecture built around skills and tools. The full reconstruction took sixty days; a first working version was running in under twenty. Ho framed both rewrites not as setbacks but as deliberate, rapid learning — the kind of iteration that only becomes possible when teams are willing to abandon sunk investment quickly.

Why it matters

For anyone designing AI-assisted customer journeys, Intuit's experience is a live case study in how invisible failure modes accumulate. The breakdown was not a model quality problem or a data problem — it was an interaction design problem. Each agent-to-agent handoff was, in effect, a lossy translation: meaning degraded with every step, much as customer intent degrades when it passes through too many service touchpoints staffed by people who each only see their own slice of the journey. The behavioral parallel is direct: context collapse in multi-agent pipelines mirrors the frustration customers feel when they must repeat themselves across channels, departments or automated flows.

The service-design implication is that orchestration complexity is not neutral. Adding coordination layers to manage complexity can generate more complexity than it resolves — a systems-thinking principle that applies equally to org charts, IVR trees and AI pipelines. The Intuit case suggests that leaner, tool-based architectures with fewer handoffs may produce more reliable outcomes, and that the willingness to measure and respond to compounding error is itself a core capability.

By the numbers

• Twice — the number of times Intuit fully scrapped and rebuilt its agent architecture within the period described.
• ~4 months — the total window across which both architectural pivots occurred.
• 60 days — time required to complete the second, skills-and-tools-based rebuild.
• Under 20 days — time to first working version of the new architecture.
• 10 agents — the scale at which compounding handoff error became critical in the orchestrated system.

The Renascence take

Most coverage of this story will focus on the technical pivot — orchestration bad, skills-and-tools good. That misses the more important signal: Intuit's real advantage was not the architecture it landed on, but the organisational reflex to kill a failing system before it became load-bearing.

The lesson here is not about AI architecture — it is about error tolerance in customer-facing systems. Every handoff, human or machine, is a moment where context can be lost and trust eroded. Customer-obsessed operators should audit their own journeys for exactly this pattern: where does intent get translated, summarised or inferred rather than carried intact? The teams that will win with agentic AI are those that treat compounding handoff error as a CX metric, not just an engineering one — and that build the institutional courage to rebuild fast when the numbers say the current design is failing their customers.