Building Lorikeet: How AI Humility and a Dual-Agent Architecture Are Redefining Customer Support

The podcast details the development of an AI customer support concierge, designed to replace generic responses with hyper-personalized, proactive solutions that directly resolve user issues. Founded two and a half years ago, the product initially struggled with user engagement, prompting a pivot from tools like coaching and self-reflection aids to addressing complex support taskssuch as cross-referencing internal data and processing refund requeststhrough advanced language models. Key challenges included aligning AI with real-world workflows, overcoming the lack of formal SOPs among users, and refining models to avoid overconfidence in ambiguous scenarios. The team emphasized iterative development, using direct feedback and real data analysis (e.g., reviewing support traces) to build products that prioritize human-AI collaboration over automation.

A central focus was on solving core pain points in customer support, particularly the high volume and complexity of tasks, by integrating AI as a "co-pilot" alongside existing ticketing systems like Zendesk or Intercom. The product evolved from command-line interfaces and spreadsheets to a conversational AI agent ("Laura Keat") capable of handling nuanced workflows, while maintaining guardrails to prevent errors and escalate ambiguous cases to humans. The company highlighted the importance of balancing AI capabilities with human oversight, especially in regulated industries like fintech, and stressed the need for hybrid interfaces that blend chat-based interactions with structured tools. Challenges in automation included adapting to informal user processes and refining AI responses to match domain-specific needs, while emphasizing the value of iterative testing and user-driven feedback to improve accuracy and compliance.

• What if you built a hybrid interface for AI agent training using test cases instead of traditional SOPs?

  • Move: Develop a tool that allows users to define test cases (e.g., "When a customer says 'Wheres my money?', trigger refund check") to train the AI agent, bypassing the need for formal SOPs.
  • Why Now?: Many customers lack structured processes, and rigid workflows hinder AI adoption. Test cases are easier to define and align with actual user intent.
  • Expected Upside: Faster onboarding for clients, improved AI accuracy, and reduced dependency on outdated or inconsistent documentation.

• What if you created a feedback loop where support agents refine AI responses in real-time during chats?

  • Move: Integrate an inline editing feature in the AI agents conversational interface, letting agents tweak or escalate responses as they occur.
  • Why Now?: Early prototypes showed agents struggling with ambiguity, and iterative improvements based on live interactions can close the "soft ceiling" gap.
  • Expected Upside: Higher agent trust in the AI, reduced escalation rates, and more accurate AI outputs that align with human judgment.

• What if you prioritized a "coach" agent for internal configuration over external support workflows?

  • Move: Build a coding-agent interface (like a "Coach" UI) for non-technical users to define workflows, guardrails, and AI behaviors without reliance on developers.
  • Why Now?: Customers need to configure AI behavior (e.g., refund rules) but lack technical expertise. A coach agent streamlines this process.
  • Expected Upside: Reduced dependency on engineering teams, faster customization, and a scalable system for iterative AI development.

• Integrate AI into existing ticketing systems (e.g., Zendesk, Intercom) as a co-pilot agent, avoiding disruption to current workflows by positioning AI as an additional support layer rather than a replacement.
• Prioritize real customer data analysis by reviewing 50+ support traces or workflows to build effective AI solutions, focusing on resolving complex tasks like refund processing or ticket diagnostics rather than just answering FAQs.
• Implement strict guardrails for AI agents to manage ambiguous or sensitive cases (e.g., compliance checks, customer intent disambiguation), ensuring they defer to human oversight for high-risk decisions or tasks requiring domain-specific knowledge.
• Start with minimal viable tools like spreadsheets or command-line interfaces to provide immediate value, iterating on user feedback to refine workflows before scaling to more complex AI-driven features.
• Develop test cases and predefined scenarios (e.g., refund requests, compliance triggers) to train AI agents and validate outputs, ensuring alignment with desired outcomes while reducing errors from overconfident AI responses.