Everything we've learned building voice agents
in production.
Winning the ElevenLabs Hackathon was the easy part. Putting Kallina in front of real customers answering real phones taught us what we didn't learn at the hackathon. These articles are the lessons — free, public, with receipts.
From the shop floor.
Voice agents: what service looks like when nobody picks up.
How Kallina processed 1,745 calls for GalTrans in a single quarter without a human operator — and the three places where it still hands off.
The 2.3 ms latency budget — how we broke it down.
End-to-end: STT, LLM, TTS, network. The budget we shipped with, where the 2.3 ms actually comes from, and the three places where we're still cheating.
When NOT to use a voice agent — four cases we've turned down.
Emotional calls, legal calls, high-objection sales, complex technical triage. We've said no to all four — and here's why, with receipts.
Prompt packs for voice agents — what's different.
Voice prompts aren't chatbot prompts. Timing, interruption handling, turn-taking, tool-calling latency — our production patterns.
The open questions.
Per-minute is the wrong unit.
Clients care about resolved tickets, not call duration. The industry prices on minutes; value accrues on outcomes. Someone will crack outcome pricing and take the market.
Romanian voice TTS is almost solved.
We're 12 months from native-quality Romanian voice out of the box. The last gap: dialect handling and code-switching (RO ↔ RU mid-sentence). We're working on it.
The CRM layer is the moat.
Voice quality commoditises. What stays proprietary: deep CRM integration, tool-call orchestration, and observability. That's why Kallina is wired end-to-end into Mega-CRM.
AI disclosure is coming.
The EU AI Act will require "you are speaking to an AI" disclosure on most voice bots in 2027. We've had it on by default since day one — you can turn it off only with explicit client consent.