Guides August 25, 2026 14 min read

Voice AI for restaurants: the 90 seconds that decide the order

Everyone counts the calls they miss. The bigger leak is in the calls they answer, and it never shows up in a report.

There are two leaks in every restaurant phone line.

The first one you already know about. The phone rings during the dinner rush, nobody is free, and it rings out. That call is gone. Most operators have seen the number on a call log at some point and winced.

The second leak leaves no trace. The call was answered, on ring nine, by a server carrying three plates, who put the caller on hold, misheard a modifier, never offered a drink, and hung up without repeating the order back. That order still went into the system. It went in smaller, slightly wrong, and attached to a customer who will think twice before calling again.

There is no report for that. No missed call alert fires. No complaint gets logged. The revenue simply arrives lower than it should have, every service, and it looks like a normal night.

This guide walks the ninety seconds where both leaks happen, then gets specific about what to fix, what to automate, and what to leave alone.

The call that never gets answered

This is the well documented leak, so here it is briefly, using published figures rather than any of our own.

In a survey run by The Harris Poll on behalf of Hostie between 29 May and 2 June 2025, among 2,065 US adults, the phone came out as the way Americans most prefer to contact a restaurant. It was cited by 63 percent, ahead of every digital channel. In the same survey, 58 percent said their calls to restaurants have gone unanswered, and roughly one in five said their calls are always or often ignored. Two thirds said they would stop going to a restaurant that does not answer. Worth noting that the survey was commissioned by a voice AI vendor, so it is best read as directional rather than neutral.

The revenue models are equally easy to find, and equally worth reading as models rather than findings. HungerRush published a widely cited worked example: a QSR taking 50 to 75 calls a day, losing 10 percent of them, of which 60 percent would have ordered at a 25 dollar average check, arriving at roughly 27,000 dollars a year. The arithmetic is sound. The inputs are assumptions, and they are almost certainly not your assumptions.

Which is the point. This leak is real, it is well covered, and it is at least visible. It shows up on a call log. You can count it. If you want your own number instead of an industry average, the arithmetic needs four inputs and you already have three of them.

Run your own numbers. The missed call calculator asks for your average check size, how many calls you miss a day, and how often a regular orders in a month. It returns the immediate revenue lost and the full annual figure once repeat orders are counted. It takes about thirty seconds. See what missed calls are costing you

The call that gets answered badly

A missed call is a clean loss. It is binary, countable, and it produces the kind of feeling in the operator's stomach that makes it easy to act on.

An answered badly call may cost more in aggregate, for a reason that is close to arithmetic. It happens to more calls.

If you miss 20 percent of your calls, you answer 80 percent of them. Every one of those is exposed to hold time, to a distracted order taker, to a menu question nobody on shift can answer, to a modifier typed wrong, and, during a rush, to an upsell that never gets attempted. Not all of them go wrong. But the missed calls are a slice, and the exposed calls are most of your phone revenue.

Each one can leak in more than one direction at once.

Ticket size No drink, no side, no dessert, no upgrade. The order is whatever the customer thought of unprompted, which is usually less than what they would have said yes to.
Order accuracy A misheard modifier can cost you the food, the remake, the driver's second trip, and the call where someone has to apologize.
Repeat rate The customer who spent four minutes on hold and then got the wrong sauce tends not to complain. They open a delivery app next Friday, and a marketplace now sits between you and that customer, taking a cut.
Floor service The staff member who took the call badly usually took it badly because they were doing three other things, and those three things got less attention for those ninety seconds too.

Little of this shows up in the reports most operators run. That is not a fault in your reporting. It is the nature of this kind of loss.

Why both leaks have one cause

These are not two problems. They are one problem with two symptoms, and the problem is compression.

Call volume peaks at exactly the hours staff availability bottoms out. Lunch and dinner rush are simultaneously when the most people want to order and when the fewest people are free to talk to them. Every call arriving between six and seven thirty on a Friday arrives into a building where every person is already committed to something else.

The most useful evidence for this does not come from the phone. It comes from the drive-thru, where somebody actually measures performance under controlled conditions. Intouch Insight's 2025 Drive-Thru Study, produced with QSR magazine and based on 165 mystery shops per brand across 13 US chains with fieldwork in June and July 2025, found that order accuracy fell two percentage points year over year, and that dinner was the hardest daypart for accuracy. Quiet afternoon visits scored best, at 90 percent.

Same brands. Same training. Same scripts. Same staff. Accuracy degrades when the pressure goes up.

That is a different channel from the phone, so treat it as directional. The mechanism transfers exactly, and the phone has it worse, because a drive-thru order taker is at least dedicated to taking orders. Your server answering line one is not.

This matters because of what it rules out.

It is not a staffing failure Your team is not underperforming. They are being asked to be in two places at once, and they are choosing the guest standing in front of them, which is the correct choice.
It is not a training failure You can train a perfect phone script. It will not survive a Friday at 6:45, because the constraint is not knowledge. It is attention.
It is not a phone line failure A second line does not help when nobody is free to pick up the first one. Adding lines adds capacity to ring, not capacity to answer.
You cannot hire your way out of it A phone only hire works two two hour windows a day and idles between them, while carrying a full wage plus payroll tax. In a year the National Restaurant Association's 2026 State of the Restaurant Industry report describes as one of elevated costs, softer traffic, and hiring difficulty, with nearly three in four operators planning to hire but expecting trouble filling roles, that math rarely closes.

It is a scheduling impossibility. Which is why the fix has to be structural rather than behavioral.

Neither leak is a people problem. Both are a capacity problem that people are being asked to absorb.

The six moments in a phone order

The call below is a composite, assembled to illustrate the failure points described above. It is an illustrative example, not a transcript of a real call and not a recording from any restaurant. Every line is a pattern operators will recognize. None of it is a research claim.

Ninety seconds. Six moments where money moves.

Ring to pickup

Rings one, two, three, four, five, six, seven, eight.

Staff Thanks for holding. Sorry, thanks for calling, one second.

Eight rings is roughly 25 to 30 seconds of a stranger deciding whether you are worth the wait. Most callers do not have a rule about how long they will hold. They have a threshold, and it is short. Everything after this moment only exists if you clear it.

What it costs The whole order, and often the customer. The tell Rings to answer, if your phone system reports it. If it does not report it, that is its own finding.

The hold

Staff Can you hold?

Caller Sure.

Two minutes forty seconds.

The caller was asked and said yes, which is why this never registers internally as a problem. It registered externally. The hold is where a warm order turns cold, and where a caller starts idly considering the delivery app already open on their phone.

What it costs A share of orders outright, and the goodwill on the ones that survive. The tell Average hold time, and how many callers hang up during it.

The question nobody can answer

Caller Is the green curry made with fish sauce? My daughter is vegetarian.

Staff I think it is fine. Let me, actually, hang on, I will ask the kitchen.

Forty five seconds.

Staff Yeah, should be fine.

This is the most expensive line in the call, and it fails in two directions. The commercial failure is another forty five seconds of dead air. The real failure is "should be fine," which is an allergen answer given under time pressure by someone guessing.

Menu, allergen, hours, delivery radius, and parking questions are the highest volume and lowest judgment calls a restaurant takes. They are also the ones most likely to be answered by whoever happened to be nearest the phone.

What it costs Time, trust, and a genuine safety exposure. The tell Count the calls that require someone to walk to the kitchen and come back.

Taking the order

Caller Two pad thai, one with no peanuts, and a green curry, chicken, medium spice.

Staff Two pad thai, one green curry chicken. What was the spice?

Caller Medium. And no peanuts on one of the pad thais.

Staff Got it.

Watch what happened. "No peanuts" was said twice and confirmed back zero times. It is now travelling to the kitchen on the memory of someone who is also watching table six try to flag them down.

When this goes wrong, the cost is never just the dish. It is the food, the remake, the labor, the driver's second trip on a delivery, the comp, and the review. The same drive-thru research found that inaccurate orders also take materially longer to fulfill than accurate ones, so the error costs you throughput at the exact moment you have none to spare.

What it costs Food cost, labor, delivery, comps, and a customer. The tell Remake and comp rate on phone orders specifically, compared against your other channels.

The upsell that never happens

Staff Okay, that is twenty eight ninety. Fifteen minutes.

No drink. No spring rolls. No dessert. No mention of the mango sticky rice on special. Not because the staff member does not know to ask, since they have been trained to ask and they do ask on quiet Tuesdays, but because there is a queue at the counter and asking takes six seconds they do not feel they have.

This is not a small leak. It is the largest one in the call, and it gets its own section below.

What it costs Every add on the customer would have said yes to. The tell Average phone ticket compared against average ticket on every other channel. If phone is your lowest, this is why.

Confirmation, or the lack of it

Staff See you soon.

Call ends.

No order read back. No name confirmed. No pickup or delivery restated. No phone number captured. No text confirmation.

The read back is the cheapest quality control in the restaurant. Fifteen seconds that catch the peanut error from moment four before it costs a dish, a driver, and a review. It is also the first thing dropped when the room is loud, because it feels redundant to the person saying it. It is not redundant. It is the only checkpoint in the whole ninety seconds.

And because no number was captured, that customer is now anonymous. You cannot text them an offer, add them to loyalty, or win them back. You served them and learned nothing.

What it costs Error rate, and the entire customer record. The tell What share of your phone orders have a usable customer record attached.

What the upsell is worth

Everyone writing about restaurant phones writes about missed calls, because loss makes a good hook. The more optimistic number, and probably the larger one, is on the other side. It is the orders you are already getting, at the size they should be.

Published figures here deserve skepticism, because they do not agree with each other. Vendor reported lifts for AI upselling run from around four percent at the conservative end to forty percent at the promotional end, depending entirely on who is publishing. That spread is too wide to be a finding. Treat all of it as marketing until someone shows their methodology.

Two things, though, are independently observable.

Even the best drilled operators upsell inconsistently. In the 2025 Drive-Thru Study, suggestive selling varied enormously between brands. Within one segment, a leading brand offered an add on in nearly 80 percent of visits while another came in below 40 percent. These are national chains with corporate training programs, mystery shopping regimes, and scripts. If their execution swings that widely, an independent restaurant's phone line will not hold a consistent standard through a rush.

When the AI took the order, upselling happened more often. The same study visited AI enabled drive-thru lanes at three brands and found upselling was more frequent than at staff run lanes, alongside faster service times and overall satisfaction six percentage points higher. Again, drive-thru rather than phone, and directional rather than definitive. The mechanism is not mysterious. The AI asks every time because asking costs it nothing. It does not get tired at seven, does not read the caller's tone as impatient, and does not skip the question because four people are waiting at the counter.

Rather than borrow anyone's percentage, do the arithmetic with numbers you already own.

Take your phone orders per week and multiply by your average phone ticket. That is your current weekly phone revenue. Now run it again with three dollars added to the ticket.

Three dollars is one soft drink, or half a side. At 120 phone orders a week, an extra three dollars on each is 360 dollars a week, or a little over 18,700 dollars a year, from asking a question you already train your staff to ask.

Then check the second effect. Compare your average phone ticket against your average ticket on kiosk, online, and in person. Digital channels prompt on every order without fail, because prompting is what software does. If your phone ticket sits meaningfully below your online ticket for the same menu, the gap is not a customer preference. It is the six seconds nobody had.

What voice AI changes

Voice AI for restaurants is software that answers the phone in natural conversation, takes the order, and writes it into your systems without a person on the line. The useful way to evaluate it is not whether it can talk, since they all can now. It is what it does to each of the two leaks.

It answers the ones you would have lost

The structural change is concurrency. A person answers one call at a time. That is the hard limit, and it is why a second phone line never solved anything. Voice AI answers call one, call two, and call five simultaneously, on ring one, at 6:45 on a Friday and at 11:40 when you are closed.

That turns your worst hour into a normal one. It also captures the after hours calls, the catering inquiry, the next day reservation, the customer checking whether you open at eleven, that currently reach nobody at all.

This is the leak the whole category advertises. It is real. It is also the easier half.

It holds a standard on the ones you would have answered

This half matters more, and it is about consistency rather than intelligence.

Menu and allergen questions get the same answer every time, drawn from a menu you configured once rather than a guess from whoever picked up
The upsell is attempted on every order, at 6:45 on a Friday exactly as on a quiet Tuesday
Modifiers are captured as structured data rather than held in working memory next to seven other things
The order is read back every time, because there is no social awkwardness in doing it
A phone number is attached to every order, so the customer exists in your system afterwards
The order arrives in the POS already entered, not on a ticket somebody has to re key while the queue builds

The honest counterweight comes from the same independent study. At AI enabled lanes, friendliness scored lower than with staff, and when orders went wrong, a staff member stepping in improved accuracy by fourteen percentage points. Voice AI is consistent, not warm, and it needs a clean human escalation path for the calls it cannot finish. Any vendor claiming it handles every call with zero intervention is selling something that will disappoint you in month two.

Get your baseline first. The comparison only means something against a real number. The calculator covers the missed call half of the picture in about thirty seconds. Calculate your missed call revenue

What to automate first

The operators who get the most out of AI tools for restaurants are specific about scope. The failure mode is not automating too little. It is automating the wrong things and then concluding the technology does not work.

A clean rule: automate what is high volume and low judgment, keep what is low volume and high judgment.

Hand to AI

Repeat pickup and delivery orders
Menu, hours, allergen, and location questions
After hours and overflow calls
Standard reservations and simple changes
Order status calls
The upsell prompt on every order

Keep with people

Complaints and service recovery
Large catering and private events
Regulars who expect to reach the owner
Severe allergy conversations that need someone who can speak for the kitchen
Press, suppliers, and anything unusual

Start narrow. After hours and overflow only is a good first configuration, because it cannot cannibalize anything. Those calls were being lost anyway. Once accuracy holds for a few weeks, widen to peak hour coverage, then to primary answering. Operators who go straight to answering everything on day one tend to discover their edge cases in public.

The point is not to remove people from the phone. It is to remove the calls that never needed a person.

Straunt AI

Answer every call, and hold a standard on the rest

Straunt AI is built around both leaks, not just the calls you miss but the ones you would have answered badly.

AI order taking: calls answered instantly, concurrent calls handled, during the rush and after close
Missed call recovery: calls that would have gone nowhere turned into confirmed orders
Menu and question handling: your menu, hours, policies, and delivery rules, answered the same way every time
Upsells built in: prompted on every order, contextually, at seven on a Friday as reliably as three on a Tuesday
Orders into your POS: clean tickets and structured modifiers, with card payment captured over the phone
Clean human handoff: transferred to a live line, or a message captured with full context, when a call needs a person

You can also have Straunt call you and take a sample order off a real menu. The demo sits on the AI order taking page, next to the calculator.

Schedule a 15 minute demo

Frequently asked questions

How many calls does a restaurant actually miss?

It varies by format, volume, and daypart, and most published figures come from vendors with an interest in the number being large. The Harris Poll survey commissioned by Hostie in mid 2025 found 58 percent of US adults reported having calls to restaurants go unanswered. The more useful number is your own. Pull rings to answer and unanswered call counts from your phone system for one Friday and one Saturday, and you will have a better figure than any industry average.

Does AI phone answering actually upsell?

Yes, and the mechanism is simply that it asks every time. Independent mystery shopping by Intouch Insight in 2025 found AI enabled drive-thru lanes upsold more frequently than staff run lanes. That is a different channel from the phone, so read it as directional. Be skeptical of specific percentage claims, since vendor reported lifts range from around four percent to over forty, which is too wide a spread to be meaningful. Model it with your own average ticket instead.

What happens to calls that come in while another call is in progress?

With a person answering, the second caller gets a busy signal, a ring out, or voicemail. Adding phone lines does not fix it, because the constraint is that one person can only talk to one caller at a time. Voice AI handles concurrent calls, so the fifth simultaneous caller is answered on ring one exactly like the first. For most restaurants this is the single largest source of recovered revenue, because it targets the peak hour where the losses concentrate.

Will customers know they are talking to AI, and will they mind?

Most will work it out fairly quickly, and disclosure norms are moving toward telling them. On whether they mind, the 2025 Drive-Thru Study found overall satisfaction at AI enabled lanes ran six percentage points higher than at staff run lanes, even though friendliness scored lower. The reading is that speed, availability, and accuracy matter more on a transactional order than warmth does. Which is also why complaints, regulars, and catering should still reach a person.

Is voice AI cheaper than hiring someone for the phone?

Usually, because a phone hire is an awkward shape of job. Two two hour peaks a day with idle time between them, at a full wage plus payroll tax. Voice AI is a fixed monthly software cost with no idle hours. Compare it against the right baseline though, which is not what a host costs but what you are losing now: the missed calls, plus the ticket size gap on the calls you already answer.

Can it handle complicated orders and substitutions?

The good systems handle a lot, and none handle everything. Ask for completion rate, meaning the share of orders finished without a person stepping in, rather than accuracy. Test it with a deliberately messy order off your own menu before signing. The 2025 Drive-Thru Study found that when staff intervened on problem AI orders, accuracy improved by fourteen percentage points, which is a useful reminder that the escalation path is a feature rather than an admission of failure.

How long does it take to set up?

Setup is generally measured in days rather than weeks: menu import, hours, policies, delivery rules, greeting, and escalation rules. The longer part is tuning. Expect a few weeks of listening to real calls and fixing edge cases specific to your menu and your customers. Start with after hours and overflow, where there is nothing to lose, and widen once accuracy holds.

Data Sources

Intouch Insight, 2025 Drive-Thru Study, produced with QSR magazine. 165 mystery shops per brand across 13 US chains, plus 120 voice AI orders at three brands, with fieldwork in June and July 2025.
The Harris Poll on behalf of Hostie, 2,065 US adults, 29 May to 2 June 2025. Commissioned by a voice AI vendor.
National Restaurant Association, 2026 State of the Restaurant Industry.
HungerRush missed call revenue model, which uses illustrative assumptions rather than measured results.

The call in this guide is an illustrative composite built to demonstrate common failure points. It is not a recording or transcript of any real call.

Guide 02 · Straunt All guides