Building Finn Voice: Our Journey into AI-Powered Phone Support

Join me as I share our journey in building Finn Voice, a voice AI agent that answers customer questions and escalates to human support when needed, and explore the key decisions and challenges we faced along the way.

1. The speaker will be discussing Finn Voice, a voice agent for phone support, designed as a frontline teammate for Intercom calls.
2. Intercom is a customer service platform that offers a messenger product and has evolved to include robust tooling for various channels like email, WhatsApp, and phone.
3. Finn, an AI agent launched by Intercom two years ago, has shown significant growth in handling customer interactions without human intervention.
4. The speaker highlights voice as the next big frontier and a cost-effective solution for customer service, with over 80% of support teams using phone support and one-third of global customer service
5. Voice AI agents can provide 24/7 support, eliminate wait times, bypass IVR menus, and support multiple languages, while also offering major cost savings and scalability for businesses.
6. The speaker discusses how Finn Voice was built in approximately 100 days, focusing on seven main areas that had the most significant impact on its development.
7. The use case for Finn Voice started with a broad, flexible, knowledge-based agent capable of answering various customer questions instead of a narrow problem space like scheduling appointments.
8. The initial wedge use case for voice agents was out-of-office hours support, allowing companies to test the technology gradually and build confidence before deploying it during regular office hours
9. The MVP (Minimum Viable Product) for Finn Voice focused on quickly shipping a meaningful product by concentrating on testing, deploying, and monitoring agent behavior.
10. Technical foundations for Finn Voice included leveraging Intercom's existing chat agent setup and native phone support product when possible.
11. The speaker highlights key differences in designing conversations for voice compared to chat, such as addressing latency, answer length, and user mindset.
12. Integrating Finn Voice into support workflows was crucial, with the majority of feedback focusing on how it works with team workflows rather than model performance or latency.
13. Smooth escalation paths and context handoff were essential integration points for Finn Voice, providing video summaries to human agents after AI agent calls.
14. The speaker discusses various evaluation methods, such as manual and automated evaluations, internal tooling for troubleshooting, resolution rate, and using language models as judges for call anal
15. Cost for voice AI agents typically ranges between three and 20 cents per minute, with usage-based pricing and outcome-based pricing being the dominant pricing models on the market.
16. The speaker believes that the market will converge toward outcome-based pricing, as it aligns incentives better between providers and customers.
17. Key takeaways from building Finn Voice include addressing performance issues as both a model and product problem, designing for the realities of phone conversations, building internal and external

Source: AI Engineer via YouTube

❓ What do you think? What are the most significant differences between designing conversations for voice versus text-based interfaces, and how do these differences impact the overall user experience? Feel free to share your thoughts in the comments!