AI phone assistant for restaurants
The phone rings hardest during service, when nobody can leave the floor. Tel-Agent takes the cover, the time and the allergy without stopping anyone.
- Covers and time
- Allergies
- Which sitting
Why a restaurant loses bookings during service
A restaurant phone rings hardest at exactly the hour every person in the building has both hands full. Most of those calls are bookings, which means a call that rings out is a table that goes to the restaurant next door - and nobody on the floor finds out it happened.
How the AI receptionist handles a restaurant booking
It answers during service without anybody leaving the floor.
It takes the name, the covers, the date and the time.
It asks about allergies and dietary needs on every call.
An allergy, a large group or an access need is flagged so a person confirms it before service.
The booking is written down and the caller gets a confirmation instead of a busy tone.
What the agent is told to ask
Written by you, in your own words, as the assistant’s instructions.
What a restaurant phone assistant should and should not do
What it does
- Answers through service and after closing
- Takes bookings, changes and cancellations
- Asks about allergies every time
- Flags groups and access needs for a person
- Keeps the book on the restaurant's own machine
What it never does
- Confirm a dish is free from an allergen
- Overbook or promise a table that is not free
- Negotiate a deposit or a cancellation charge
- Take card details over the phone
- Send guest data to a third-party platform
Questions
Can it take an allergy safely?
It records one; it never confirms that a dish is free from anything. An allergy is set to reach a person before service, because the answer belongs to the kitchen and getting it wrong is not a booking problem.
What about cancellations?
Those are worth more than the booking they replace. A table given up in the afternoon is still re-sellable for the evening; the same table discovered empty at eight is not.
Does it handle large groups?
It takes them and flags them. A group changes the layout, the kitchen and often the deposit, so it is confirmed by a person rather than accepted automatically.
Will it overbook the restaurant?
It proposes from what it has been given and confirms; it does not write into a live book on its own. One phantom table on a Saturday costs more than the calls it saved.
Does it work for a restaurant with sittings?
Which sitting is one of the questions where there is more than one, because a booking at the wrong sitting is a table that has to be moved on the night.
Does the EU AI Act allow an AI to answer a table booking?
Yes, under Article 50 of Regulation 2024/1689, which asks that the caller be told they are speaking to a machine - that line belongs in the greeting you configure. Nothing about this work falls under Annex III, the list of high-risk uses. The distinction the Act draws is between assisting and deciding, and this agent assists: it never decides who gets a table or a room. Austria's KI-Servicestelle at RTR publishes the same reading of the transparency duty.
Who is the data controller for what a caller to a restaurant says on the phone?
You are. A table booking is ordinary personal data, plus whatever a guest volunteers about an allergy, and Tel-Agent is self-hosted with your own model keys, so the recording, the transcript and the record it writes stay on your server and never reach Dpro GmbH. There is no processor agreement to sign with us because we never receive the data, and the source is AGPL-3.0, so you can check that rather than take our word for it.
Can a restaurant run this in the United States?
There is no single federal AI statute to comply with, but a phone line carries older rules. The TCPA governs automated outbound calls, and consent to record is set state by state. Consent to record is set state by state, and California and Illinois among others want every party to agree. Running the software yourself makes you the caller of record, so the disclosure and the recording switch are yours to set.
Not usable yet. The day it answers a real call will be a release, not a newsletter.
Watch the repository
