Boosting Voice AI Performance for Deutsche Telekom

Deutsche Telekom AG operates one of Europe's most advanced conversational AI platforms, built on their open-source LMOS (Large Model Operations System) framework through the Eclipse Foundation. Their AI assistant serves customers across both chat and voice channels, with chat agents handling over 1,000 use cases across 10-12 specialized agents and achieving resolution rates in the low 50% range. However, their voice channel presented unique challenges, with containment rates around 7% compared to chat's strong performance.

While Deutsche Telekom's chat-based AI assistant was performing well, their voice channel struggled to deliver comparable results. Customers were trained to use short, command-like utterances rather than natural conversation, and the system often delivered verbose, information-heavy responses better suited to text than spoken dialogue. The team needed guidance on voice-specific design principles, prompting strategies, and how to optimize their agent architecture for natural spoken conversations.

Conversation Design Institute (CDI) provided ongoing office hours support for Deutsche Telekom's voice AI agent team, covering prompt optimization, architecture design, and voice-specific conversation patterns. CDI addressed critical topics including:

Voice Design Fundamentals: Turn-taking, expectation management, cognitive load management, and error handling specific to generative AI

Prompt Engineering: System prompt optimization, use case structure, and handling universal intents (like "I don't know," help requests, and ambiguous questions)

Architecture Guidance: Recommendations on routing logic, managing conversation breakdowns, and when to implement handovers versus in-flow resolution

Specific Use Cases: Soft-resistance handling, bot-to-human handover protocols, and managing ambiguous yes/no responses to either-or questions

Results

Through regular office hours sessions, CDI helped Deutsche Telekom:

• Establish clear voice design principles for their team to follow across all use cases
• Optimize their prompt structure, including guidance on where to handle universal intents (system prompts vs. dedicated use cases)

-Develop strategies for managing conversation state and ensuring clean transitions between use cases

• Address specific pain points like barge-in handling, TTS emphasis patterns, and disambiguation approaches

The ongoing collaboration equipped Deutsche Telekom's mixed team of engineers, business stakeholders, and designers with the knowledge and best practices needed to continuously improve their voice AI assistant's performance and create more natural, conversational experiences for their customers.