Sixty Hours a Week, and Twenty of Them Are Typing

Every article about AI and restaurants starts with AI. That's why none helped. I'm starting with your Tuesday.

You answer the phone between 11 and 1 with a tray in the other hand. Sunday night is the schedule, Monday morning the walk-in count. Forty reviews a month come in; you answer three, badly, at 11 pm. Sixty hours is a quiet week.

The National Restaurant Association's 2026 State of the Industry report says 42 percent of operators weren't profitable last year, and more than 9 in 10 named food, labor, insurance, energy and swipe fees as significant challenges. Four in ten of you didn't turn a profit working those hours.

Price the hours. The Bureau of Labor Statistics puts the median food service manager at $33.36 an hour in 2025, and $65,060 a year inside restaurants, which is $31.28 an hour. That's the floor on your hour.

At $31.28 an hour, a 60-hour week is $1,877 of labor that never shows up on the P&L. Forty of those hours only you can do. The other twenty are typing, counting, replying, reminding and picking up a phone. That twenty is the whole guide.

Three leaks eat most of it: the phone nobody can reach in a rush, the booked tables that stay empty, the dishes that quietly cost more than last spring. A chatbot plugs some, not all; whoever told you otherwise was selling the chatbot.

People hand AI the wrong jobs. They type "write marketing for my restaurant," get soup, and call it a toy.

The only rule you need: hand AI any job you'd hand a smart new hire with no hands, no memory past today, and no way to look anything up. Keep everything else, or buy it.

Give that hire facts and rules and they'll draft all day. But they can't pick up the phone, see the walk-in, or remember Monday on Tuesday. They can't text a guest or charge a deposit; they aren't connected to anything.

Three questions first.

• Does it need hands? "Answer the phone," "count the shelf," "send the reminder at 3 pm." A chatbot can only write the script.
• Does it need yesterday? A chatbot starts from zero every morning unless you paste the past in. Fine weekly, painful daily.
• Does it have to be right every time? A menu description can be wrong once. A deposit charge can't.

Three no's: Chapter 3. Needs yesterday: Chapter 4. Needs hands and has to be right: Chapter 5, where doing it yourself breaks and I'll say what Cenvi is for. You don't need it for most of this.

If it needs a brain for ten minutes, hand it over. If it needs hands, yesterday, or has to be right every time, keep it or buy it.

Three jobs, each wrong at most once. Do them on a laptop; copying reviews out of the Google app on a phone is punishment. Replace anything in [brackets] with your own facts. The examples use Saffron Lane, a fictional 72-seat restaurant I built to test every prompt.

Replying doesn't move your rating; fixing what people complain about does. Replying is how you find out what that is, and it tells a stranger somebody's home.

Tonight: copy your last ten Google reviews into a chat, paste this.

You write Google review replies for [restaurant], a [seats]-seat [cuisine] restaurant in [city]. The owner is [name]. Rules: under 50 words each. Name one specific thing from the review. No "thank you for your kind words", no "we're sorry to hear". Never offer a discount or free food. Sign each reply "[name]". For 1 and 2 stars, say the one thing we will change and invite them to reply by email at [email]. Reviews:
[paste 3 to 10 reviews]

I ran this on a fictional 1-star: "called four times to order pickup, nobody answered." The reply: "Four calls with no answer is exactly the thing I'm fixing. The host phone goes unanswered during the Friday rush. We've moved pickup orders to [link] so you never need the phone, and I'm working on the phone itself." Read it first; it sometimes promises a fix you haven't made.

Ask for "appetizing" and you get "succulent, slow-cooked lamb in a rich, aromatic gravy that melts in your mouth," twenty-five times. People order from ingredients and method.

Rewrite each menu description below. Rules: 8 to 14 words. Start with the main ingredient or the cooking method. Say what it tastes like or how it is cooked, never how good it is (no "delicious", "mouthwatering", "perfect"). Mention heat level if it matters. Keep dish names exactly as written. Do not invent facts about popularity, sourcing or awards. Return a two-column list: dish, new description.
[paste menu]

On my test menu, "Lamb Nihari: slow cooked lamb" became "Lamb shank simmered overnight in bone broth with ginger and fried onion." Now it sounds like $24. Without the "do not invent facts" line, it called butter chicken "our most ordered dish," which it couldn't know. Check it against the recipe: it wrote "bone broth" because nihari usually is, not because it knows yours.

Facts in, fluff banned.

Write 5 Instagram captions for [restaurant] for this week, one per day [days]. Each under 40 words, no hashtags, no emojis, no questions to the reader. Each one is about one real thing from this list, and uses only the facts I give you: [this week's special, an event, a staff member, a dish from the menu, a behind-the-scenes moment]. Plain voice, like the owner typed it. End each with one line telling people what to do (book, order pickup, come by).
Menu: [paste]
This week: [three or four true facts]

A Friday caption from my test: "Friday is already booked past 7. There are still 5:30 and 8:45 tables. Book online, the phone will be busy."

These three need yesterday's numbers. Keep them in one document, paste them in, and a chatbot does the thinking. That works weekly. It breaks the day the job turns daily.

Here are my dishes with menu price and what the plate costs me in ingredients. For each: food cost percentage (cost divided by price). Then list every dish above 32% from worst to best. For each flagged dish, give me the price that would bring it to 30%, rounded to .49 or .99, and one alternative to raising the price (smaller portion, cheaper garnish, a side swap). Show the math. Do not change dish names.
[paste dishes with price and plate cost]

On my 25-dish test menu, eight dishes came in over 32 percent and four at 40 or worse: shrimp masala, lamb nihari, goat biryani, lamb shawarma. Nihari at $24 with a $9.80 plate cost is 41 percent. Getting it to 30 means $32.99, a $9 jump nobody pays. The model's alternative was better: one naan instead of two, or a weekend special at $27.99.

Walk the shelf with a voice memo running, say item and count, paste the transcript.

I'm going to dictate my Sunday night count. Turn it into a table with: item, unit, counted, par, order quantity (par minus counted, rounded up to the case size). Pars and case sizes are in the second list. Flag anything counted above 1.5 times par as "overstocked" and anything under half of par as "urgent". Then write the order as a plain message I can text to my supplier.
Count: [dictate]
Pars and case sizes: [paste once, reuse every week]

The pars list is the trick; without it the model guesses. My test count flagged chicken thighs urgent at 1.5 cases against a par of 4, basmati overstocked at 6 bags, and wrote the supplier text: "Order for Tuesday delivery: 3 cases halal chicken thighs, 2 cases tomatoes, 1 block paneer."

Four weeks of covers and sales by dish beats feel.

Here are covers by day for the last 4 weeks, and what sold by dish on the last 4 [day]s. This [day] there is [event] and [weather] is forecast. Estimate [day] covers as a range, then give me a prep list for the 8 best selling dishes with portions to prep, assuming the kitchen preps 10% over the estimate. Show which number each estimate came from.
Covers: [paste]
[day] dish sales: [paste]

My test asked for a Friday with a high school game at 7 and rain. It averaged the last four Fridays at 210, said a home game pulls dinner early and rain cuts the patio, called 185 to 215, and planned 56 portions of butter chicken, showing every step. Your cook can argue with it.

These four need hands and memory and have to be right every time; a miss is a guest who doesn't come back. A chatbot writes the script for each; it can't run any. The fix is the only pitch in this guide.

The Harris Poll, commissioned by Hostie, which also sells a phone agent, asked 2,065 US adults: 69 percent would likely give up on a restaurant altogether if nobody answers. The NRA, which sells nothing, says 39 percent of off-premises customers ordered by phone in the last six months.

A chatbot can write everything except the answer.

Write three things for [restaurant]'s phone. 1) A one-page answer sheet the host can read from for the 10 most common calls (hours, location and parking, reservations, large parties, catering, menu questions, allergies, takeout, wait times, gift cards). 2) A voicemail greeting under 25 words that tells callers to text the same number or book at [link]. 3) A text message we can send back to a missed call, under 160 characters, with the booking link and the pickup link.
Facts: [hours, address, parking, links, catering minimums]

The voicemail it wrote: "You've reached Saffron Lane. We're on the floor. Text this number or book at [link] and we'll get right back to you." Record it tonight. One catch: on a landline, "text this number" is a lie. Use your booking link instead.

Write a reservation policy for a [seats]-seat restaurant that we can paste on our website and read on the phone. Include: how far ahead we take bookings, how long we hold a table, parties of 8 or more need a card on file, cancellations by 24 hours before, and a note that walk-ins are welcome. Under 120 words, friendly, no legal tone. Then write the confirmation message under 100 words and the questions our web form should ask.

My test wrote "We hold your table for 15 minutes past your time. Parties of 8 or more: we ask for a card to hold the table, and we only charge it if nobody shows." The form asks the question that matters: "may we text you reminders."

Two numbers from the same data say what to do. A delivered text reminder meant a 16 percent lower no-show rate. A required card or deposit meant 0.80 percent against 2.47 percent without. OpenTable, which also sells deposits, says deposits cut no-shows by 57 percent on average. And 56 percent of cancellations arrived within 24 hours of the booking or after it started, so the reminder has to go out early enough to resell the table.

Write a three-message reminder sequence for a restaurant reservation, sent by text: 48 hours before, 24 hours before, and 3 hours before. Each under 160 characters. The 24-hour one asks them to confirm or cancel with one word. Then write the two-line explanation of our card-on-file policy for parties of 8 or more, in the tone of a host, not a lawyer.

The 24-hour message it wrote: "Saffron Lane: tomorrow at 7:30 for 4. Reply YES to confirm or NO to cancel so we can plan the room." That one text is the 16 percent.

Here is my seating: [tables by size]. Here are covers by hour for the last 4 Fridays and Saturdays. Build me a booking grid for Friday and Saturday: how many tables of each size to open per 30-minute slot, how many to hold back for walk-ins, and when to stop taking parties of 6 or more. Assume 90 minutes per table of 2 to 4 and 2 hours for 5 or more. Show your reasoning in 5 lines, then the grid.

My test grid held back three two-tops and two four-tops for walk-ins, stopped parties of 6 or more at 7:30 so the six-tops turn again at 9, and gave the twelve-top one booking a night. Change the 90 minutes if your turns run longer; a model doesn't know your room.

A guest calls Friday at 7:10 while every hand is busy. The agent answers, finds a 7:45 two-top, texts a booking link, takes the card for the hold, puts it on the calendar, sends the reminder, and the table fills. Nobody at the host stand touched it. That's the autonomous restaurant: not the kitchen, the ten minutes a night that were leaking.

The title is a claim, so let's check it. Nothing here takes all twenty of those hours off your plate on its own. About seven of them you can take back yourself with the prompts here, three of them tonight. The other thirteen need something that answers, sends and charges while you're on the line.

Easy wins: reviews, menu copy, social. Three hours a week back.

Medium jobs, on a weekly routine: food cost, counts, prep. Four hours back, if you keep it.

Hard jobs, needing hands and memory: the phone, the book, the reminders, the grid. Thirteen hours, and a chat window can't give them back, because someone has to answer, send, charge and update. That's what Cenvi is for: it answers the phone, runs reservations and deposits, sends the reminders, and keeps the calendar live. It doesn't write your reviews or count your walk-in. You don't need it to.

Twenty hours back isn't less restaurant. It's less typing.

What's left when those twenty hours are gone is the restaurant. The food, the room, the people. The part only you can do.

cenvi.events/content/autonomous-restaurant/