How to compare AI restaurant voice agents.
Most demos get judged on the voice. A restaurant needs to know something else: whether the booking the guest heard is the booking in your book. These are the tests we'd run on any AI host, ours included.
A natural voice is the easy part.
Any vendor can play you a smooth call. Booking agents fail in quieter ways, and you only see them in the book.
It says “confirmed” and saves nothing
The guest hangs up happy. There's no booking in the book, and you find out when they arrive.
It offers a table that's taken
The agent answers from an old copy of your book, or doesn't check at all. Two parties turn up for one table.
It takes a message and calls it a booking
“We'll see you Friday” turns out to mean “someone will call you back”. Nobody does.
Ten tests to run on any AI host.
Run them on a test line set up with your own tables and times, not a rehearsed demo. Check the book after every call. Do it once at 7pm on a weekday and again late on a Friday.
Book a table, then open the book
Make an ordinary booking. Then compare what the agent said with the saved record: date, time, party size and name. They should match exactly.
Ask for a time that's full
Fill 8pm first, then ask for it. A good agent offers times that are actually free. It should never confirm 8pm, and never offer a time it hasn't checked.
Change your mind mid-sentence
Say “four people, actually six”. The saved booking should say six. Read it in the book rather than trusting the read-back.
Hang up halfway
Give a date and party size, then end the call. Nothing should be held for you, and nothing half-finished should sit in the book.
Ask how two callers are handled
Ask what stops two guests getting the last table at the same moment. The answer should describe a check and save that happen together, not a sync every few seconds.
Move a booking you made earlier
Call back and change the time to one that's full. The original booking should still be there afterwards, untouched.
Ask for a person
Find out what really happens: a live transfer, a message or a request in a queue. Then look at what your staff see, and listen for promises the agent can't keep, like a callback within ten minutes.
Ask something it hasn't been told
Ask about corkage or a dish that isn't on the menu. It should say it doesn't know and pass it on, not make up an answer.
Switch languages
In Dubai this one matters. Start in English and switch to Arabic halfway through. The booking should still come out right.
Talk over it
Interrupt while the agent is speaking. It should stop, listen and answer what you actually said.
Questions to put to every vendor.
Where does availability come from?
Does the agent book into its own book, read your current system, or write back into it? If it only reads, staff can miss the bookings it makes.
What happens when something is down?
If the booking system or the agent's own service fails mid-call, what does the guest hear? An agent that can't confirm a booking should say so.
How is quality checked after launch?
Ask whether calls are tested on a schedule, and who grades them. An agent that grades its own calls proves very little.
Who can hear your guests' calls?
Ask who can open recordings and transcripts, where they are kept and how deletion works under UAE data protection law. Ours is on the Data page.
What will it cost on a busy month?
Per-call and per-minute fees add up on a full Friday. Ask for the whole monthly cost at your call volume, any setup fee and how long the contract runs.
How Onla does on the same tests.
We'd rather you test us than take our word for it. This is what you should find. You can build an AI host for your own restaurant and run the tests yourself.
| Test | What Onla does |
|---|---|
| Book, then open the book | Amira confirms only after the booking is saved, then reads back the saved date, time, party size and name. |
| A time that's full | She offers the closest of the times the availability check found free, never one it didn't. |
| A change mid-sentence | The booking is saved with the corrected details, taken from the guest's own words. |
| Hanging up halfway | Nothing is held. A table is taken only when the booking is saved. |
| Two callers, one table | Checking and saving happen in one locked step, so the second guest is offered other times. |
| Moving a booking | Done on the call. If the new time is full, the original booking stays. |
| Asking for a person | A request in your team's queue with a summary and the guest's number. There's no live transfer, and Amira says so instead of promising one. |
| A question it can't answer | She answers from the facts set up for your venue. Anything else becomes a request for your team. |
| Switching languages | Amira switches between English and Arabic on the same call. |
| Talking over it | She stops when you speak over her and answers what you said. |
| Quality after launch | Onla calls itself four times a day with scripted test guests, and the booking system grades each call, not the agent. |
| Cost | One monthly plan per venue. No setup fee, cancel any month, and a first-month guarantee. |
Asked when comparing AI hosts.
What is the most important test for an AI host?
Book a table, then compare the saved booking with what the agent said. If those two don't match, nothing else on the list matters.
Is a demo call enough to judge an AI host?
No. Demos are rehearsed. Ask for a test line set up with your own tables and times, and check the book after every call.
Should the AI host book into its own system or ours?
Either can work. What matters is one book everyone trusts, and bookings made by the AI reaching your team. A read-only link to your system means staff can miss them.
Does Onla answer in Arabic?
Yes. Amira answers in English or Arabic and can switch between them on the same call.
How can I test Onla myself?
Call our AI host on the live demo, or build one for your own restaurant from its website on the Your AI host page. Then book a walkthrough to run the tests on your own floor.