The problem: missed calls cost small businesses real money

Dental practices, hair salons, and restaurants live and die by the phone. A missed call during a busy shift is a missed booking, and most small businesses can't staff a receptionist for every hour they're open. That's the problem we set out to solve when we built an AI phone receptionist (in German: a KI-Telefonassistent) for the German SMB market.

This isn't a product pitch, it's a rundown of the technical decisions and tradeoffs that mattered most while building a production voice AI system that has to work in German, on real phone lines, under GDPR.

Core architecture

A phone-based AI receptionist breaks down into four stages that all have to happen inside roughly 1-2 seconds of perceived latency for a call to feel natural:

1. Telephony ingestion - a SIP trunk / PSTN bridge that hands the call audio stream to your application in near real-time.
2. Speech-to-text (STT) - streaming transcription, not batch. German has long compound words and regional accents (Bavarian vs. Berlin German sound very different), so generic English-tuned STT models underperform here.
3. LLM reasoning + business logic - intent detection (booking, rescheduling, FAQ, emergency triage), slot-filling against a calendar API, and guardrails so the model doesn't hallucinate appointment slots that don't exist.
4. Text-to-speech (TTS) - needs to sound natural in German specifically, not translated English prosody mapped onto German words.

What actually broke in practice

A few lessons that weren't obvious until we shipped this to real practices:

• Latency budget is the real constraint, not model quality. A smarter LLM that takes three extra seconds to respond loses to a slightly dumber one that responds instantly. Callers hang up on silence far more readily than they forgive an imperfect answer.
• Barge-in / interruption handling matters more in voice than in chat. Callers talk over the bot constantly ("no wait, actually..."). If your pipeline can't cancel a TTS stream mid-sentence and re-listen, the experience falls apart immediately.
• GDPR/DSGVO data residency isn't optional in this market. German SMB owners ask where call data is stored far more often than English-speaking customers do. EU-region hosting and a clear data-processing agreement are a sales requirement here, not a nice-to-have.
• German-language search demand for this category dwarfs English. While validating the market, we found the German term "KI-Telefonassistent" gets roughly 14x the search volume of the closest English equivalent ("AI receptionist") within Germany. A reminder that translating your product copy isn't enough — the whole content and SEO strategy needs to be built around the language your customers actually search in.

Where this landed

We turned this into Hallodesk, an AI phone receptionist (KI-Telefonassistent) built specifically for dental practices, salons, and restaurants in Germany — answering calls, booking appointments, and handling FAQs in German, with EU-hosted data.

If you're building anything in the voice AI / conversational AI space, happy to swap notes in the comments, especially on STT accuracy for non-English languages, which still feels like the most underrated part of this stack.