Navigating AI Evolution: Intuit's Rebuild of Agent Architecture

In the fast-paced world of artificial intelligence, adaptability is key. Intuit, the financial technology giant known for its innovation in accounting and tax software, recently faced a significant challenge: the failure of its AI agent architecture. Within a brief span of just four months, Intuit underwent two major overhauls of its AI systems, reinventing its approach to customer service and automating processes. This evolution, described by Nhung Ho, Intuit's VP of AI, at the VB Transform 2026 conference, highlights the intricate balance between developing specialized AI agents and maintaining seamless communication between them.

Intuit's journey serves as a critical case study in the broader context of AI development, illustrating the rapid pace of technological change and the necessity for companies to pivot quickly in response to shortcomings. As they navigated the complexities of AI orchestration, Intuit discovered crucial lessons that not only affected their internal processes but also reshaped their customer interactions.

The Rise and Fall of Agentic AI

Intuit began its foray into AI with a focus on agentic systems—intelligent agents that autonomously carry out specific tasks. The initial strategy involved deploying a fleet of specialized agents, each designed to handle distinct customer queries. However, this approach quickly revealed its limitations. Customers were often overwhelmed by the need to choose the appropriate agent for their needs, leading to frustration rather than satisfaction.

Transitioning to Orchestration

To streamline customer interactions, Intuit introduced a central orchestration layer intended to manage the flow of tasks among these agents without requiring customer input. This layer was designed to route requests automatically, thereby enhancing efficiency. However, the orchestration layer struggled under its own complexity. Agents communicated results in natural language, which meant each transfer of information lost context, leading to compounded errors. As Nhung Ho noted, “If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds.”

Lessons from the First Rebuild

The breakdown of the orchestration layer was not merely a matter of capacity; it was a structural flaw in how information was shared. The first iteration of the system lasted approximately three months before it became evident that a major overhaul was necessary. Intuit’s response was to pivot to a skills and tools-based architecture, where individual agents would be dismantled and their capabilities repurposed into shared resources.

The 60-Day Race

This pivot was not just a technical challenge but also a cultural one. Convincing leadership and engineering teams to abandon months of work required a compelling case. Ho and her team built a demonstration using real customer queries to showcase how the new architecture could outperform the existing system. This hands-on approach highlighted the importance of addressing customer needs as the driving force behind technological decisions.

Integrating Human Support

One of the standout features emerging from this rebuild is the ability to seamlessly integrate human support during AI interactions. Currently in early testing with about 1% of Intuit's customer base, this feature allows conversations with live agents, such as Intuit product support personnel or user accountants, to occur mid-dialogue. This contrasts sharply with traditional AI chat systems, which often end with generic disclaimers advising users to consult a professional.

Building Trust through Transparency

This integration is complemented by a robust permissions model tailored for financial data, ensuring that any actions taken by the AI on a customer's behalf are explicitly authorized. Intuit also maintains an audit trail of all agent interactions, allowing for transparency and trust-building over time as customers engage more with the system.

Revolutionizing Feedback Mechanisms

Another significant change that emerged from the rebuild is the way Intuit collects and utilizes feedback. Previously, feedback was sparse and often polarized, with users either loving or hating the experience. The new chat-based system, however, allows every interaction to serve as feedback. This transition has led to a dramatic increase in customer feedback, from approximately 0.3% of users providing explicit feedback to nearly 100% engaging in the process.

Responding to Customer Sentiment

Ho emphasizes the value of direct feedback, stating that customers often express their frustrations candidly. This influx of data allows Intuit to quickly identify pain points and areas for improvement. By returning to coding herself, Ho is working on models that can analyze this feedback at scale, ensuring a continuous cycle of enhancement in the AI's performance. The goal is to transform negative feedback into actionable insights while maintaining a balance of grace and understanding from customers.

Key Takeaways

• Intuit underwent two major AI architecture rebuilds in four months to address customer service challenges.
• The transition from an orchestration layer to a skills and tools architecture improved efficiency and reduced error propagation.
• Integrating human support into AI interactions enhances customer experience and builds trust.
• New feedback mechanisms have led to a significant increase in customer engagement and actionable insights.

Frequently Asked Questions

What prompted Intuit to rebuild its AI architecture?

Intuit's decision to rebuild its AI architecture was primarily driven by customer feedback indicating frustration with the complexity of choosing among a fleet of specialized agents. The first orchestration layer attempted to streamline this process but ultimately failed due to compounded errors arising from the natural language communication between agents, necessitating a more effective solution.

How does the new AI architecture benefit customers?

The new skills and tools-based architecture allows for more efficient task handling without requiring customers to choose specific agents. Additionally, the integration of human support during AI interactions ensures that customers can receive personalized assistance when needed, enhancing the overall customer experience.

What changes were made to feedback collection and analysis?

Intuit's approach to feedback has transformed dramatically with the introduction of chat-based interactions. Every conversation serves as a source of feedback, leading to a notable increase in customer engagement. This data is now analyzed systematically to identify areas for improvement, allowing for a more responsive and adaptive AI system.

What are the implications of these changes for the future of AI development?

The lessons learned from Intuit's experience underline the importance of flexibility and responsiveness in AI development. As companies increasingly rely on AI to enhance customer service, adopting a human-centered approach and maintaining open channels for feedback will be crucial for ongoing improvement and innovation in the field.