Why voice AI agents feel weird in production

Why do voice AI agents still feel "off" to users, even when the tech works?

Aidan Hornsby and Jack Bridger break down the real production challenges that voice AI builders face: from subtle issues that break user trust to deep technical problems that don't surface until you're serving real calls.

This episode covers what we've learned talking to dozens of teams building voice AI systems in production.

---

What We Cover

Naturalness & Trust

• Why mispronouncing one word in 100 makes your agent feel broken—especially names and numbers
• The difference between voice and text agents: problems that don't exist in chat
• Designing the right "vibe"—why Domino's intentionally made their voice AI "dead inside"
• ChatGPT Voice quirks and why overly enthusiastic agents backfire

The Technical Reality

• Stacked non-determinism: how STT → LLM → TTS errors compound across turns
• Why voice calls are "one-shot"—users hang up, they don't reroll like with text
• Background noise and "presence": why silence makes users think the call dropped
• Browser audio gotchas (mobile Safari mic latency)
• Testing voice systems: why traditional QA breaks down

Production & B2B Deployment

• 60+ day onboarding cycles and why self-serve doesn't work yet
• CRM integration nightmares in vertical markets
• Why business owners are scared to deploy voice AI during business hours
• Multi-agent handoffs, tool latency, and keeping conversations flowing

Latency & UX Patterns

• Turn-taking tuning vs. perceived naturalness
• Filler sounds, clicks, and other latency-masking techniques
• Fast/slow agent architectures for background processing
• How to give feedback in voice when there's no visual UI

---

Timestamps

• 0:00 — Intro: Voice AI Happy Hour
• 2:02 — Naturalness: mispronunciations, names & voice vs text
• 4:27 — Real-world calls: unexpected user behavior & production QA
• 7:16 — Testing voice systems: STT → LLM → TTS non-determinism
• 10:03 — One-shot UX: why users hang up and LLMs are "slot machines"
• 11:33 — Background noise, silence & presence cues
• 14:08 — Designing the right vibe: Domino's "dead inside" + ChatGPT voice
• 16:09 — B2B adoption: onboarding, integrations & CRM challenges
• 21:16 — Building complex agents: handoffs, workflows & tool latency
• 23:37 — Latency patterns: turn-taking + UX feedback techniques
• 28:33 — Browser audio gotchas: mobile Safari mic
• 30:25 — Demos vs reality (Sesame, Her) + wrap-up

---

Resources

• Layercode
• Layercode Docs
• Follow Aidan on X
• Follow Jack on X

Mentioned in This Episode

• Domino's AI voice ordering
• "Dead inside" Domino's discussion on Reddit
• ChatGPT Voice Mode
• Sesame voice demo
• "Her" (film)

---

Have a topic you'd like us to cover? Reach out on X @uselayercode or email us at podcast@layercode.com