A look at what the receptionist actually does between the moment it hears a caller and the moment it speaks back. With real sample calls and notes on what's happening behind each line.
An AI receptionist isn't one thing. It's a loop of six components, and each one has a specific job. The easiest way to understand the whole is to understand what each part does.
The receptionist turns voice into text as the caller speaks, not after they finish. It picks up the language on its own and copes with accents, background noise, and the little pauses people make when they're thinking.
Before the caller's second word is out, the receptionist has already matched their phone number against your CRM, PMS or EMR. So the call starts with context. It knows who's calling and what they did last time.
A language model reads your business knowledge base (menu, hours, services, prices, policies) and works out what to say, what to ask, and what to do next. This is the part that makes the agent feel like it actually knows your business.
When something needs to be saved (a booking, an order, a quote, a ticket) the receptionist writes it straight into your reservation system, CRM, PMS, EMR or ERP. Nobody on your team has to retype anything afterwards.
The reply comes out in a natural voice, in the caller's language, in the tone you set for your brand. Audio starts streaming while the response is still being written, so the caller doesn't sit through an awkward silence.
The receptionist knows when a call shouldn't be handled by AI. It hands the conversation to a teammate live, with the full context already on screen, the moment a caller asks for a person, confidence drops, or any of your rules fires.
Here's a normal call from start to finish. On the left is what the caller and the agent actually say. On the right is what's happening in the background, line by line.
(call answered on the first ring)
"Bonjour Marie, merci d'appeler Le Bistro. Comment puis-je vous aider aujourd'hui?"
"Yes hi, I'd like to book a table for 4 people tomorrow night around 7 pm."
intent = reservation · party_size = 4 · date = 2026-05-19 · time ≈ 19:00 · checking availability…
"7 pm is fully booked tomorrow, but I have a table for 4 at 7:15 pm. Would that work for you?"
"7:15 works. Can we get the table by the window?"
reservation_id = R-48217 · table_pref = "window" · confirmation SMS dispatched
"You're all set, Marie. 7:15 pm tomorrow, table for 4, with a note for a window seat. I've texted you the confirmation. See you tomorrow!"
Outbound uses the same six components, just in a different order. The agent picks who to call, dials them, listens, and books. Here's a real reactivation for a dental clinic.
queue = recall-q2 · target = James K. · last_cleaning = 2025-11-04 · local_time = 14:32 (within calling window) · dialing…
"Hello?"
"Hi James, this is the front desk at Bright Smile Dental. Just a quick courtesy call. Our records show you're due for your 6-month cleaning. Would you like me to find a time that works?"
"Oh, sure. Do you have anything next week in the afternoon?"
"I have Wednesday at 2:30 pm or Thursday at 4 pm with Dr. Chen. Which works better?"
"Wednesday at 2:30 is perfect."
appointment_id = A-9921 · provider = Dr. Chen · type = cleaning · reminder_sms = T-24h · written to PMS
"Booked. Wednesday at 2:30 pm with Dr. Chen. You'll get a text reminder the day before. Thanks James, see you then!"
Calls feel natural because each step finishes in milliseconds, and the next one starts before the previous one is done. Here's what a typical turn looks like in numbers.
Streamed transcription. The text is ready as the caller's last syllable lands.
Working out what the caller wants, looking it up, deciding the action, drafting the reply.
Time until the first audio chunk plays. The rest of the sentence streams while the caller is already hearing it.
About the same as a fast human receptionist on a good day.
Every Orpanai deployment ships with escalation rules. The agent transfers to a human the moment any of these triggers fire, and the transcript and context are already on the teammate's screen when they pick up.
If the caller says "can I speak to someone?", "get me a manager", or anything close, the agent transfers right away. No friction, no script to escape.
If the model isn't sure it understood the caller, or the request falls outside the knowledge base, the agent hands the call off instead of guessing.
Medical concerns, complaints, threats, or anything you flag as "always human" gets escalated by policy. The agent doesn't have to make the judgment call.
Catering above a threshold, large group bookings, custom quotes. Any time you'd rather a person close the deal, you can route it to a teammate.
If the agent fails to resolve a request after two attempts, it stops trying and brings in a teammate. The caller never gets stuck in a loop.
If you can describe it, you can use it as a trigger. VIP customers, specific keywords, time-of-day constraints, anything that fits how your team already works.