Intuit Rebuilt Its AI Agent Architecture Twice in Four Months and Called It Progress

Quick Facts

• Intuit rebuilt its AI agent architecture twice in roughly four months, completing the second overhaul in 60 days with a working version in under 20.
• The company moved from large multi-agent systems to a skills and tools architecture after discovering that a 10-agent chain compounded errors by design.
• When Intuit shipped AI agents to 3 million customers, 85% returned; TurboTax users now complete taxes 12% faster, and QuickBooks customers get paid an average of five days sooner.

Intuit scrapped its AI agent architecture not once but twice in four months. At VB Transform 2026, Nhung Ho, Intuit’s VP of AI, described that cycle of demolition and rebuilding as the fastest path forward, not a sign of failure.

The company behind QuickBooks, TurboTax, and Mailchimp has been building toward an AI-driven platform since 2019. Its GenOS platform, anchored by a component called GenRuntime, functions as the operating system for its agents. GenRuntime receives data, reasons over it, and determines what action to surface to the customer.

Why the Architecture Kept Breaking

The first rebuild moved Intuit from a fleet of large specialist agents to a central orchestration layer. That layer held for about three months before structural problems forced another overhaul. The cause was error compounding.

When agents pass outcomes to one another in natural language, each downstream agent must infer how the upstream agent reached its conclusion. That inference degrades with every additional handoff. A 10-agent chain does not fail occasionally. It fails by design.

‘If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,’ Ho said at the conference.

The second rebuild decomposed those large agents into granular skills and tools. Rather than each agent owning a broad capability, shared skills now run across every part of the product that needs them. The rebuild also decoupled Intuit’s orchestration from specific model providers, letting the company swap in different models without rebuilding the stack around them.

What the Numbers Show

The business results are concrete. According to Ho’s presentation, TurboTax users with AI agent support complete their returns 12% faster, with nearly half finishing in under an hour. QuickBooks Online customers are getting paid an average of five days sooner, and overdue invoices are 10% more likely to be paid in full. The company’s Business Tax Agent helped 58% of small businesses uncover deductions they had previously missed.

When Intuit rolled out AI agents to 3 million customers, 85% came back. The company attributed that retention rate to pairing AI with human expertise rather than replacing one with the other.

Human Experts Built Into the Flow

The architecture change also restructured how human experts enter a conversation. A customer can bring in an Intuit product support person, their own accountant, or one of Intuit’s bookkeepers mid-conversation. That person joins with full context of what the agent has already done, rather than starting from scratch.

Intuit is currently testing a live handoff feature with about 1% of its customer base. Ho said the company plans to scale it in the coming weeks.

The new system also changed how feedback works. In a chat-based product, every conversation generates a signal. Ho said Intuit went from roughly 0.3% of customers ever providing explicit feedback to something close to 100% with the new architecture. That volume of signal is what drives evals, which have become the primary measure of whether the system is performing.

What This Means for Enterprise AI Teams

Intuit’s experience offers a direct warning for software companies building agent systems at scale. Orchestration layers that look clean in design can become brittle under real workloads, particularly when agents communicate in natural language rather than structured outputs. Decomposing agents into shared, reusable skills reduces compounding errors and makes the architecture easier to test.

Every agent action on customer financial data at Intuit requires explicit permission, and the company maintains a full audit log of reversible actions. As agent systems move deeper into sensitive workflows, that governance layer matters as much as the underlying model.

Ho framed the repeated rebuilds not as instability but as the price of moving at the speed the market demands in 2026. ‘We changed the orchestrator, we changed the planner, we changed the brain, and we also changed what everybody had to build across the whole company,’ she said.

Read more: Intuit scrapped its own AI agent architecture twice in four months. At VB Transform 2026, its AI VP called that the fast path