Just Now Possible From Prototype to Production: How Perk Built a Voice AI Agent That Makes 10,000 Calls a Week
15 snips
Dec 4, 2025 Steven Payne, Product Manager at Perk, Gabriel Stock, Senior Engineering Manager, and Philipe Steiff, Senior Software Engineer, dive into the world of voice AI. They discuss how they automated over 10,000 weekly calls to ensure hotels charge virtual cards, overcoming challenges with prompt engineering and text-to-speech. The team shares their surprising findings from hotel interactions, refining conversational stages, and the importance of manual call reviews despite advanced metrics. Their innovative journey uncovered operational issues, paving the way for future AI applications.
AI Snips
Chapters
Transcript
Episode notes
Calls Were Verbose Until Prompts Got Tighter
- Early calls were verbose and sometimes derailed because prompts lacked off‑ramps, leading to 'please charge the card now' loops.
- One hotel employee even asked the agent if it was an AI and said they preferred speaking to an AI.
Explicitly Format Numbers For TTS Clarity
- Handle number and reference formatting explicitly using TTS markup so hotels hear digits clearly.
- Add spacing or phonetic cues to avoid confusion between letters, zeros, and similar tokens.
Narrow Agent Scope To Prevent Hallucinations
- Keep agent goals narrow and deterministic to avoid unwanted actions or hallucinations.
- Avoid broad role prompts like 'experienced agent' that grant the model excessive authority.