"Use customer data to improve the service" is often not allowed in the EU. We talked with Versaia about what it takes to deploy voice agents in European enterprise. The playbook is different.
hamming.ai/blog/building-… #VoiceAI
English
Hamming
122 posts
Hamming
@HammingAI
Making AI voice agents reliable (YC S24) Demo: https://t.co/3uyC3hHTwi
San Francisco Katılım Mayıs 2024
36 Takip Edilen303 Takipçiler
IBM, RingCentral, and Google all added native voice AI in the last 2 weeks. When every enterprise platform ships built-in voice agents, reliability and testing become the differentiator. Not "we have a voice agent." #VoiceAI #EnterpriseAI
English
How do you know when a voice agent “doesn’t sound right”?
Listening + prompt tweaks don’t scale once you have real branching.
Brian Jeppesen's framework:
When users ask “live agent,” don’t guess but copy your best humans.
Use their wording/pace/empathy as the reference, then train + test against it.
English
On a recent podcast with Ronald from Versaia, he pointed out something we didn’t know about the Netherlands:
They had “chatbots” 7–8 years ago, and they were so bad that they trained customers to distrust the whole concept.
GenAI changed what’s possible, but now the job is earning trust back: prove the agent works end-to-end on real customer journeys.
English
Your eval tool scores transcripts. Your users experience audio.
That's where voice agents break.
We found that text-based evaluation misses ~40% of voice-specific failures (latency spikes, ASR misrecognition, interruption handling, and background noise).
The transcript looks fine. The score looks fine. The customer hangs up.
Evaluations tell you what happened. Simulations tell you what will happen. For voice, you need both.
English
Enterprise voice isn’t blocked by capability. It’s blocked by trust.
Leaders aren’t worried about the average call. They’re worried about the one call where the agent says something weird, and it becomes a screenshot and a headline.
You can’t guarantee perfection with probabilistic systems. But you can make the risk low and detectable:
- strong guardrails
- repeated testing (same scenario, many runs)
- production monitoring + fast escalation
English
Mia Labs supports 350+ franchise dealerships, powering 1M+ customer conversations and booking 130K+ sales & service appointments.
Before Hamming:
- Each appointment test cycle took 10+ minutes
- One person, one call at a time
After Hamming:
- 100+ test cycles per week
- After-hours testing that doesn’t disrupt dealership staff
Proud to partner with the Mia Labs team.
Full case study.👇
English
Cameras aren’t “objective.”
In our chat with Thymia, we discussed lots of video models were trained mostly on Caucasian facial features. So, for some Southeast Asian drivers, systems can flag “asleep” when they’re fully awake.
That bias is why automotive is shifting toward audio-first (and other modalities), sometimes removing cameras entirely.
Watch the Full Episode 👇
English
Thymia shared something big on our podcast: they’re on track to be the first speech-biomarker company regulated as a medical device, starting in mental health (with diabetes next, prototype now, deploying in 2026).
Most exciting part: speech biomarkers getting embedded into everyday communication, where they can actually help.
English