A caller dials your restaurant, the AI picks up on the first ring, chats naturally, and a confirmed order lands in your kitchen. It feels like magic — but under the hood it's a few well-understood steps happening in about a second. Here's what actually happens, in plain English, and why the details matter when you're choosing a provider.
The four steps of a voice-AI call
Almost every restaurant voice AI does the same core job, in this order:
- 1. Listen (speech-to-text).The caller's voice is transcribed into text in real time, word by word, as they speak.
- 2. Understand & decide (the language model). An AI model reads that text along with your menu, hours, and rules, figures out what the caller wants, and decides what to say or do next — add an item, ask a clarifying question, quote a price.
- 3. Act (tools). When the model needs to dosomething — look up a dish, add it to the order, check if you're open, place the order — it calls a tool connected to your menu and systems, not just chit-chat.
- 4. Speak (text-to-speech). The reply is turned back into a natural-sounding voice and played to the caller. Then the loop repeats for the next thing they say.
All of that happens in the pause between the caller finishing a sentence and the AI answering — usually around a second, smoothed over with natural conversational timing so it doesn't feel robotic.
The two main approaches
Providers build that loop in one of two ways, and it's the biggest technical fork in the category:
Cascade (text-in-the-middle)
The steps above run as separate stages: speech-to-text, then a text-based language model, then text-to-speech. Because the middle is plain text, everything the AI “hears” and “says” is captured as a clean transcript, tool calls are precise, and the same logic can power a chat simulator you can test without calling. It's reliable, auditable, and cost-effective — which is why it's the common default.
Speech-to-speech (realtime)
A single model listens to audio and speaks back directly, without a text stage in the middle. This can feel a touch more fluid and natural — quicker to interrupt, more expressive — at the cost of being newer, pricier, and harder to audit. Some platforms (Depla included) offer it as an option for the calls that benefit most.
Why the plumbing matters to you
- Accuracy comes from your menu, not just the model.A voice AI is only as good as the menu, hours, and rules it's given. Ask how a provider ingests and keeps your menu current.
- Tools are what make it take a real order. Without live connections to your menu and POS, an AI can only talk about ordering, not actually do it. This is the difference between a receptionist and an order-taker.
- Transcripts are a feature, not a given. A text seam means every call is logged and reviewable — useful for training, disputes, and spotting missed revenue.
- Latency is engineered, not free.Good providers mask the think-time with natural pacing so callers never feel they're talking to a machine.
See it for yourself
The fastest way to understand voice AI is to talk to one. Try Depla's live demo on your own menu, read the how it works overview, or see how AI order taking turns a call into a confirmed order.