← ALL EVENTS
Builders of Voice AI · Episode 1
June 18, 2026
/
~1 hour
/
Mai Medhat & Michael Sujith
About this episode
VEC AI has run over 1.3 million conversations across logistics, recruitment, debt collection, and enterprise customer support. Michael brings real scar tissue to this conversation — covering conversation design, model selection, latency hacks, completion rates, memory architecture, and why the hardest work only begins after you go live.
10 things we learned
01 · Silence kills calls more than latency
Add ambient background noise during processing gaps — typing sounds, light office noise. The caller stays engaged. Simple fix, real impact.
02 · 60% completion at launch is completely normal
Well-tuned agents hit 89–92% after 8+ months. Don't panic on day one — the number goes up as you learn.
03 · The real work starts after deployment
Build, test, deploy, fine-tune, stabilize. Budget at least 6 months before you can call an agent truly stable.
04 · A small copy change moved completion from 40% to 60%+
Moving the AI disclosure later in the call, and explaining why you're asking questions, was enough for one car dealership to see a 20-point jump.
05 · Single-prompt agents don't scale
Complex calls need 7–8 sub-agents. Each handles a specific task with its own context — and limits prompt injection risk.
06 · STT is still the weakest link
Regional accents and domain jargon still cause errors. Fine-tune with your customer's specific terminology and names before you go live.
07 · Match your LLM to the complexity of the call
GPT-4.1 as a default. Mini models for simple outbound. Full model for multi-intent inbound. Model choice is a cost lever, not just a quality one.
08 · Voice selection is a brand decision
Let the brand choose. Give each agent in a call a distinct voice so the caller can follow handoffs. Rotate and test every few months.
09 · Memory separates good agents from great ones
Short-term: inject the last 2 call summaries at the start. Long-term: RAG over historical transcripts. The goal is an agent that actually remembers the customer.
10 · Observability isn't optional past 1,000 calls per month
You can't manually review at scale. Set evals that define success and failure, let automation surface patterns, and manually verify the edge cases.
Standout moments
“It's not latency — it's silence.”
The single most actionable insight from the conversation. Ambient noise masks processing gaps in a way that lower latency alone can't.
The “supervisor check” trick
During a negotiation, the agent says “let me check with my supervisor” and pauses 2 seconds, then returns with the same offer. Customers accepted it at higher rates.
Voice vs. chat: the debate
Michael put the overlap at 70%. Mai pushed back hard — voice has background noise, interruptions, accent variation, and zero tolerance for silence.
“If your completion rate is under 70%, don't do voice AI yet.”
Michael's blunt take on readiness. Below that floor the economics don't work — voice AI costs about one-third of a human agent, but only if it's completing calls.
After-hours is the best place to start
Low stakes, high value. VEC AI's first deployment was after-hours support, and the team listened to every call for the first 10 days.
Guests
Mai Medhat — Founder and CEO, Tuner
Building the reliability layer for voice AI: observability, evals, and call simulation.
Michael Sujith — Founder and CEO, VEC AI
Building conversational AI for enterprises across logistics, recruitment, and customer support.
Get notified when we announce new episodes and upcoming live conversations with voice AI leaders.