Amadeus shares lessons from AI voice agent development

Amadeus IT Group has shared insights from its journey in developing a production-grade voice AI agent. The company highlighted that transitioning from a demo to real-world application introduced significant technical and user experience challenges.

In a blog post, Amadeus emphasized that eliminating unexpected delays and employing asynchronous orchestration are critical for ensuring a smooth voice conversation. The system must respond without lag to prevent conversation breaks. To achieve this, Amadeus utilized a high-performance voice-to-voice model, bypassing the traditional speech-to-text and text-to-speech pipeline. Latency from backend systems and API calls was managed by configuring tool executions to run asynchronously.

Regarding user experience, Amadeus stressed the need for conciseness and careful curation in voice interface design. Unlike a website, a voice agent must present a limited and clear set of options, enabling users to follow the conversation easily without visual aids.

The company also pointed to the importance of reliability and data governance. Production systems require robust guardrails to ensure the agent functions as intended and avoids errors. Regulations concerning personal data processing, such as GDPR, also shape how the system is built and improved.

Amadeus partnered with Microsoft to develop an AI agent capable of autonomously handling traveler requests, including modifying bookings. If the agent cannot complete a task or if wait times become excessive, calls can be automatically transferred to a human agent.