Half your WhatsApp users don't type — they hold the record button and talk. Especially in Latin markets, especially over 50s, especially any customer who's driving / cooking / carrying groceries. If your "AI concierge" only handles text, you've just excluded a big slice of your audience from the party.
What the AI does with a voice note
Transcribe → understand → answer
The chain is invisible to the customer. They hold and speak; a few seconds later, they get a text answer (or a voice answer, if you've enabled it). Under the hood: Whisper (or equivalent) transcribes, the LLM reads the transcript in-context, the response goes back on the same channel.
Language auto-detect works on voice too
You don't configure the language per user. The transcription pipeline detects Portuguese vs. Spanish vs. Italian from the audio itself. Reply follows the detected language.
Handles noise, accents, half-sentences
Real voice notes are messy — background noise, driving, kids, half-finished sentences. Modern transcription models handle 90-95% of this. When they don't, the AI defaults to "Sorry — could you send that again?" rather than guessing.
The reply flow: text back? voice back?
Both are options; each fits different markets.
- Text back. Default. The customer can re-read, forward, screenshot. Every reply becomes searchable. Faster to generate.
- Voice back. Optional. A synthesised voice replies in the same language + tone as the customer's voice note. Feels warmer for over-50 markets, Latin markets, hospitality. Slower but higher-signal.
- Custom voice (enterprise). Clone a specific human voice (with permission) so replies sound like your actual concierge. Requires ~30 minutes of recorded material to train.
The trust angle: does grandma trust an AI voice reply?
Yes, if you don't try to pretend it's a person. The trust move is transparency: the AI introduces itself as an AI in the first voice reply ("Hi — I'm the concierge assistant; happy to help right away"), and the customer feels informed rather than tricked. In our data, guests who know they're talking to AI + get an accurate answer rate the interaction the same as guests who talked to a human — as long as the answer was right. Faking humanness backfires within 2-3 exchanges.
Setup: on by default, no config
Voice-note support is on for every ChatGenius instance from day one. No toggles, no per-channel setup, no separate KB. You upload your docs, connect WhatsApp, and voice notes work — text or voice reply, per your preference.