Two months ago the only server at my family's restaurant walked out, and we did not post a job ad.
The restaurant is Wok & Karahi in Spring, Texas, north of Houston. We serve Hakka-style Indo-Chinese food next to Indian and Pakistani dishes, all zabihah halal. My family is Indian by origin and Canadian by citizenship, and we run the place as a family. I used to be a software engineer at Google.
That server was the whole front of the house. One person answered the phone, took orders, ran payments and greeted the door, all at once. At peak hours it fell apart. When that one person did not show up and nobody could cover, we closed for the day. A closed day cost us about $2,000 in sales, and we still paid everyone their full wage. When someone was short, a family member came in and covered, whether we could or not.
We had been through that cycle enough times. So when the server left, I spent the next 24 hours building something to do the job.
By the following day, every table had a QR code. An AI host named Haniya was answering the phone and sitting inside the menu. Every order was paid for before it reached the kitchen. Nobody has been hired into that job since.
What a guest sees
They sit down and scan the code on the table. Haniya greets them and asks which language they want: English, Spanish, Urdu, Hindi, Gujarati, Punjabi or Arabic. If they are a regular, she says welcome back. They can tap through the menu, or hold a button and talk to her the way they would talk to a waiter. Say "chicken chow mein" and the dish opens on screen with chicken selected, and she says "Adding chicken chow mein" while the price updates. Ask what goes with the karahi and she suggests a pairing based on what other guests actually order together. She asks only the questions the kitchen needs answered, one at a time. A note like "no onions" or "nut allergy" prints on the kitchen ticket word for word, because the guest typed it. Then they pay from their own phone with Apple Pay or Google Pay.
For older guests, and anyone who finds apps hard, there is nothing to learn. Hold the button and say what you want.
What the engineers want to know
Every order from every channel goes into Clover, our point-of-sale system, so the kitchen sees one queue. Counter, phone, tables, website, DoorDash, Uber Eats and Grubhub all land on the same screen in the back. A phone order ends with a pay link texted to the guest. The kitchen ticket prints only when the payment clears. If a ticket fails to print, I get an alert. I also get a text on every table order. Nothing is cooked until it is paid for, so dine-and-dash is gone.
Haniya is the same host on the phone, the website and the table. She knows each guest's past orders, favourites and rewards across all three. Pickup guests get a text the moment their food is ready, and a thank-you a few hours later asking how it went. The pickup time she quotes reads the live order queue across every channel, so when she says 35 minutes it is 35 minutes.
Every call is recorded. Every call gets reviewed. That review loop is most of the real work.
Some numbers from our own Clover data, over the last couple of months. In mid-August, 17 percent of our direct orders were self-serve. Over a recent 14-day stretch it was 94 percent. 100 out of 101 dine-in orders came in by QR code. 317 out of 317 app orders were paid up front. About 1 in 6 calls Haniya takes come in while we are closed.
That last number is the one I did not expect. Catering callers used to ring after hours, find nobody, and give the order to someone else. Now the phone answers 24 hours a day. Haniya takes the request, answers the question, or books the order for the next day. She also offers an add-on on every call, and about 4 in 10 callers say yes.
What broke
Guests asked for a person. Some people do not trust an AI on the phone, and the first versions handled that badly. I tried a fixed line that pitched her abilities. It sounded scripted, so I pulled it. What works is answering what they asked first, in her first words, before anything about the order or payment. If someone asks for a person a third time, she tells them she has passed their situation to the owner and they will hear back within 24 hours, and I get one text. That count lives in code. The model does not get to decide it.
Paying first was a change for some regulars. We take no in-person payment on any channel now. When a pay link does not reach a guest, she resends it or tries another number or email. The hard news and the help go in the same breath: we are all digital now, and I can stay on the call with you while you make that payment.
She has to say "we don't carry that" when we don't, with confidence. Early on she would say she "couldn't find" a dish, which makes a guest think she is broken. Internal notes must never be spoken aloud. One language per call, no mixing. One question per reply, never asking the same thing twice. Every one of those rules came from listening to a real call where it went wrong. I fix the spots where guests got stuck every week, and the system gets a little less strange to people each time.
Then the back office
Once ordering ran on its own, the same system took on the work behind it. It writes our social posts and guest texts. It runs our Google Ads, and every ad dollar is traced to the order it brought in. An ad that does not bring orders turns itself down. It makes production-quality photos and short video ads of our dishes from pictures of our real food, with no photographer and no shoot day. Traffic by hour and by day tells the chef what to prep and how many hands each shift needs.
The tips changed too. Servers used to keep most of them. Now they go to the cooks who make the food.
By my count, the system saves or wins back about $10,250 a month. That is the server's $3,500 monthly salary, about $2,000 for the day a month we used to close when nobody could cover, family members no longer pulled in to work the floor, marketing and Google Ads we would otherwise pay an agency for, the photo and video shoots, catering that no longer goes to competitors, fewer wrong orders and remakes, and no dine-and-dash. Some of those are measured and some are my estimates, and I mark which is which when I show the breakdown. Running it costs about $10 a day.
Does this take jobs?
At Wok & Karahi, the job that went away was the one my family could not keep filled. The people who stayed are calmer in the rush, because one person is no longer doing four jobs at the door.
America's biggest chains are doing the same thing. Yum Brands says voice AI now takes orders at more than 890 Taco Bell drive-thrus. McDonald's has said it will bring AI to all 43,000 of its restaurants. The National Restaurant Association's 2026 industry report found that 80 percent of US restaurants were short at least one position last year. Those chains have engineering teams and large budgets. An independent restaurant has a family, and usually a family that is already stretched.
Most of what we built is ordinary engineering. Webhooks, a POS integration, a pay link, a print rule, a phone agent and a weekly review of real calls. The hard parts were the edges: the guest who wants a person, the order that does not print, the caller at 11 at night who wants food for 40 people tomorrow. A chain can hire a team to find those edges. A family restaurant finds them one bad call at a time, unless someone builds it with them.
I now build the same kind of system for other independent restaurants through my company, AZ Restaurant Partners. If you run one, or you know someone who does, I would like to hear from you.