Why AI voice agents for small restaurants need honest scoping

Choosing AI voice agents for small restaurants looks like a vendor comparison. It is really a question about your phone line. Before you book demos, you need to know who calls, what they ask, and what a missed call costs you. This article walks through that decision the way an advisor would with an owner.

What a voice agent does on a restaurant call

An AI voice agent answers the phone, understands spoken requests, and completes a task such as booking a table or logging a takeaway order. Yelp's voice agents, announced in April 2025, show the typical scope: handling calls, answering basic questions, adding a guest to a waitlist, and passing complex requests to a person. After each call, the business receives a summary, a transcript and a recording. That handoff and paper trail matter more than the voice itself, because they decide how much risk you carry.

Vendors differ mainly in what sits behind the voice. OpenTable lists around fifteen partners, including Loman, Certus, Hostie AI and Slang AI, so the category is crowded. Yelp's chief product officer told TechCrunch he expects voice technology to become commoditized, with data and the handling of customer queries as the real differentiators. Treat that as one executive's view, but it points the right way: judge the setup around the voice, not the voice.

Audit your own phone line before you compare vendors

Most buying guides start with a vendor shortlist. Start with a one-week call log instead. It costs nothing, and it gives you numbers that belong to your restaurant rather than to a sales deck.

Pull every inbound call from your phone provider's call history, answered or missed, and add a short note on the reason for each. Then sort the reasons into categories. The table below shows a sensible first sort, and it also shows where automation fits and where it does not.

Call type	What to look for in the log	Usual fit for automation
Table reservations	Volume at peak hours, party size, special requests	Strong fit when your booking system connects
Takeaway and delivery orders	Order length, modifiers, payment step	Good fit if the order reaches your POS cleanly
Opening hours and menu questions	Repeat questions at predictable times	Strong fit
Allergen and dietary questions	Frequency and risk of a wrong answer	Route to a person
Complaints and refunds	Tone and frequency	Route to a person
Large groups and events	Value per call	Human follow-up

The log answers three questions. How many calls do you miss, and at which hours? Which call types dominate? What is a typical call worth? Multiply the missed calls that would have become bookings by your average spend, and you get a figure to set against a subscription quote. Use your own conversion estimate, and keep it conservative.

The same discipline applies to any automation project, which is why our checks to run before automating a business process start with measurement rather than tools.

How to read the numbers vendors publish

Vendors publish large figures, and you should treat them as marketing claims until your own log confirms them. Certus AI's Y Combinator profile states that restaurants lose up to $100,000 per year per location from missed calls, that the average restaurant gets 75 calls a day, and that 25% go unanswered. Those are the company's own statements. A quiet neighborhood bistro and a high-volume takeaway sit far apart on that average.

Put three questions to every vendor and ask for the answers in writing:

1. Which systems does the agent connect to today, and which are only planned?
2. What happens when the agent cannot answer, and who gets the call?
3. How is pricing structured, and what changes as call volume grows?

Certus lists Toast, Square, Clover, UberEats and DoorDash among its integrations. That is the kind of specific answer you want, matched against your own stack rather than a generic list.

Integration and handoff decide the outcome

A voice agent that books a table nobody sees in your reservation book is worse than a missed call. Check that confirmations land in the system your staff already use, and ask how double bookings are handled. Yelp says its agents can connect to a restaurant's management software to send reservation details after a call. Whether that works with your software is a question for the provider.

Handoff deserves equal attention. Decide in advance which calls go to a person, during which hours, and how the transfer happens. Allergen questions and complaints are sensible candidates. Staff also need a short routine for reading summaries and transcripts, otherwise the paper trail goes unread.

McKinsey's analysis of AI agents suggests they could eventually act as skilled virtual coworkers. That is a useful frame, because a coworker needs onboarding, supervision and clear limits. Callers also meet your brand through the agent, a point Harvard Business Review makes about AI concierges, including OpenTable's AI Concierge.

Recordings, transcripts and privacy

A recording is personal data the moment a caller gives you a name or mentions an allergy.

Call recordings and transcripts hold names, phone numbers and sometimes health details. Before launch, ask the provider where recordings are stored, how long they are kept, and who can open them. Callers should also hear that an AI is answering and that the call may be recorded. Rules differ by country, so a legal adviser should confirm what applies to your restaurant.

Off-the-shelf product or custom build

For a single location, a packaged product connected to your existing booking and POS tools is the sensible start. Custom work earns its cost when you run several sites, have unusual ordering rules, or need call data to feed your own systems. A studio can build those cases, for example through the Xerx AI development and workflow automation service, although a pilot with a packaged product often shows first whether you need that.

If you are weighing this against other projects, where business process automation AI really pays off gives a wider view of which processes justify the effort. Voice is rarely the first one for a kitchen with a quiet phone.

Run a four-week pilot before you commit

A pilot limits your exposure because you can stop it. Start with overflow calls or off-peak hours, and keep a person on the line for anything the agent hands back. Compare each week with your baseline log: share of calls answered, bookings completed, handoff rate and complaints about the call experience.

Read a sample of transcripts every week, since averages hide bad calls. At the end of week four, decide to extend, adjust the scope, or stop. If the numbers do not beat your baseline, you have lost a month and a small fee, not a contract. Start on Monday by logging calls.

Action steps

1. Log one week of calls: Record every inbound call with its time, reason and whether anyone answered.
2. Sort calls by type: Mark which reasons suit automation and which should always reach a person.
3. Estimate missed-call value: Multiply missed calls likely to convert by your average spend, using conservative assumptions.
4. Ask vendors for written answers: Request current integrations, handoff behavior, pricing structure and data retention terms.
5. Define handoff and review routines: Decide who takes transferred calls and who reads summaries and transcripts each week.
6. Pilot for four weeks: Start with overflow or off-peak calls, compare against your baseline, then extend, adjust or stop.

Frequently asked questions

What can an AI voice agent do for a restaurant?

It answers calls, handles routine requests such as bookings or basic questions, and passes complex ones to staff. Yelp's announced agents, for example, can add guests to a waitlist and send the business a summary, transcript and recording after each call.

How much do AI voice agents for restaurants cost?

Pricing varies by provider and structure, so ask each vendor for a written quote that shows fixed fees, usage charges and integration costs. Compare it with the value of missed calls from your own call log.

Will an AI voice agent work with my reservation or POS system?

It depends on the provider and your software. Certus, for instance, lists Toast, Square and Clover among its integrations, but you should confirm your exact system and version with any vendor before you sign.

Is it legal to record calls handled by an AI?

Rules on recording and on telling callers that an AI is answering differ by country. Ask the provider how it handles consent and storage, and have a legal adviser confirm what applies to your restaurant.

Should small restaurants build a custom voice agent?

Usually not at first. A packaged product tested in a pilot shows whether the idea works for you. Custom work makes sense for multiple locations, unusual ordering rules or deep integration with your own systems.