Automate the repeatable, checkable jobs: answering the phone orders you currently miss, keeping the menu identical on every platform, drafting review replies and turning sales into a prep list. Skip anything where a wrong answer hurts a customer or costs real money: allergy advice, refunds, pricing decisions and the call on when the kitchen is full.
Most of the value sits in the dull jobs. A phone agent that answers every call but mishears one order in ten can cost more in remakes than the calls it recovers, so judge every job by the orders that arrive right, not by how many calls or tasks the software touched. The sorting rule is short, and the ranking comes from your own numbers.
Two questions that sort every takeaway job
Put any job through two questions. How often does it repeat, and how long does it take each time? And if the output is wrong, who finds out first: you, or the customer? Vendors talk at length about the first question and rarely raise the second.
| The job... | Mistake is caught before a customer sees it | Mistake reaches a customer directly |
|---|---|---|
| Repeats every day or week | Automate: menu updates, review drafts, prep lists | Automate only with a safety net: phone orders with a read-back, replies limited to opening hours and address |
| Happens rarely or once | Use a chat assistant when it comes up; no set-up needed | Skip: allergy answers, refund decisions, supplier disputes |
Phone orders, menu changes, reviews and stock all repeat, so they sit in the top row. What separates them is the column. A wrong prep list reaches only your own kitchen. A wrong phone order reaches a customer who is already hungry. That difference, together with setup hours, drives the ranking below.
Four jobs ranked by effort and payback
Every figure from here on comes from one invented business, so the sums can be followed: a single-site burger and wings takeaway, open 26 evenings a month, with about 110 orders an evening (40 by phone, 50 through two delivery apps, 20 walk-in), an average order of $24, food and packaging at about 35% of sales, and your own time valued at $25 an hour. Swap in your numbers; the method is what carries over.
| Job | Setup effort | Weekly upkeep | Net gain a month (illustrative) | How sure the gain is | What a mistake costs |
|---|---|---|---|---|---|
| Menu changes | About 3 hours | About 1 hour | $100 to $200 in time saved, plus fewer price mismatches | High: the time is countable today | A wrong price sits on one channel until someone spots it |
| Review replies | About 1 hour | About 10 minutes | $25 to $50 in time; the rest is reputation, which nobody can price | High | An awkward reply goes public if you skip the read-through |
| Stock and prep list | About 8 hours | About 45 minutes | $0 to about $400, depending on how clean your item names are | Low until you back-test it | Short of a fresh item on a Friday, or over-prepped |
| Phone orders | About 12 hours | About 30 minutes reviewing call logs | From a small loss to about $480, depending on missed calls | Medium, once test calls are done | A wrong order, a remake and an annoyed caller; at worst an allergy slip |
Read the table in two directions. Ordered by what each job is worth when it works, the list runs phone orders, stock, menu changes, review replies. Ordered by effort and risk, it runs the opposite way. The order I would work in is menu changes, review replies, stock, then phone orders, because the cheap jobs give you a clean menu sheet and the habit of checking AI output, and the expensive jobs need both.
Costs differ in kind as well as size. The first three jobs run on an ordinary chat assistant: a paid plan at about $20 a month list price is the usual step up, and it is worth testing whether a free tier covers your volume first. None of them needs customer names or phone numbers pasted into a chat tool, because the menu sheet, the prep counts and public reviews hold none. Only the phone agent adds a bill of its own, and it is the only one that talks to your customers unsupervised.
Phone orders: the missed-call tally decides
Vendors lead with phone agents because the demonstration is impressive and the pain is real. Whether one pays depends on a number you can collect in two weeks: how many callers you lose at your busiest hours.
Run the tally for ten trading evenings. Use the missed-call list on your phone line, or a tick sheet beside the handset, and count the calls nobody answered between 6:30 and 8:30 pm. Then cross out every number that rang again within 15 minutes or turned up later as an order in your system, because those callers were not lost.
| Evening | Missed, 6:30 to 8:30 pm | Called back or ordered later | Lost |
|---|---|---|---|
| Thursday | 4 | 2 | 2 |
| Friday | 9 | 3 | 6 |
| Saturday | 11 | 4 | 7 |
| Sunday | 6 | 2 | 4 |
| Ten evenings in all, including the six not shown | 52 | 17 | 35 |
That is 3.5 lost callers an evening, about 91 a month, in this illustration. The sums follow, with every assumption in view.
- One in three lost callers would have ordered had someone answered: about 30 orders a month. That is a guess. If you cannot test it, use one in four.
- The agent captures 80% of those correctly: about 24 orders, worth roughly $580 at $24 each.
- At a 65% gross margin that is about $377 of gross profit.
- One recovered order in ten needs a remake or refund, costing about $58 a month.
- Net before the tool's price: about $320.
As of October 2026, Maple lists its Voice plan at $150 a month, or $85 a month billed annually, and its Pro plan at $350, or $220 billed annually, with no per-minute fees. Yelp Host is free for up to 30 days, then from $249 a month. Smith.ai's AI Receptionist Pro plan starts at $150 a month but charges for every call beyond the plan's allowance. That pricing model matters more than the headline figure: this takeaway handles roughly a thousand phone calls a month, so a per-call agent should answer only the calls staff miss, while a flat-priced plan can take the lot.
At about $250 a month, close to Yelp Host's starting price, the agent nets about $70 for this takeaway. At Maple's $85 annual rate it nets about $235, and at $350 it loses about $30. Break-even at $250 is roughly 19 recovered orders a month, which means about three lost callers every evening. The setup, around 12 hours of loading the menu with its modifiers, connecting the order system and making test calls, comes on top. For this takeaway the honest answer is "marginal". A takeaway losing eight callers a night, with every other number unchanged and the same $250 plan, nets about $480 a month and has an easy yes.
The failure to test for is the negative modifier. "No onions", "without cheese" and "sauce on the side" are where a voice agent slips, and the caller finds out when the bag is opened. A good read-back repeats the modifiers and the total, and a bad one repeats only the dish names.
When the tally is low, a cheaper move catches some of the same callers: a missed-call text back that sends your online ordering link, covered in setting up missed call text back. When it is high, the sums are the easy part and the quality check is the hard one. Run the twelve test calls in the test checklist for menus and delivery areas, and set the agent to transfer to staff whenever a caller mentions an allergy. The set-up guide for an AI phone line covers the build.
Menu changes: one sheet that every platform copies from
Changing a price or marking a dish sold out means touching the order system, two delivery apps, the Google profile, the website and sometimes a printed board. Pushing a change to several platforms automatically is an order-system or menu-sync feature, not an AI one, so ask your EPOS supplier which platforms it updates. What AI does well here is the writing and the checking.
Keep one master sheet with a row per item and size: price, description, allergen entry copied from your allergen matrix, the channels where it is live, status, and the date last changed. Edit the sheet first, then paste only the changed rows into the assistant with a prompt like this one.
Here are the changed rows from our master menu sheet. For each row write:
1. the one-line website update;
2. the note to type into the delivery apps;
3. a 15-word line staff can say on the phone.
Use only facts that appear in the rows. Do not write allergen wording: copy the
Allergens column exactly. List anything in the rows that looks inconsistent.
Spicy wings (10) | $11.50 | UNAVAILABLE until 7 pm tomorrow | Allergens: [from matrix]
Wings box (20) | $20.00 | live | Allergens: [from matrix]A typical draft (illustrative) comes back with a banner line, an app note and a phone line, and, because the prompt asked, a flag such as "the 20-piece box costs $20.00 but two 10-piece portions cost $23.00: is the saving intended?" Two faults appear often. The assistant writes "back tomorrow" when the row says 7 pm, so check the wording against the status column every time. And it sometimes adds a detail the row never gave, such as "served with fries". Strip it.
Descriptions need the same discipline. Before: "Chicken Burger - chicken burger with toppings and sauce." After a draft built from the sheet: "Crispy chicken thigh, pickles, lettuce and house hot sauce in a toasted brioche bun." If the kitchen uses a seeded bun, "brioche" is the invented part. Models fill gaps with the common choice, which is why the sheet has to carry every ingredient you want mentioned. Allergen text never goes through the assistant for rewriting; it is copied from the matrix, and when the matrix changes someone re-checks every channel.
The checklist is short, and each line ends with a check that takes under a minute.
| Place | Who or what updates it | Check that it worked |
|---|---|---|
| Order system (till) | You, first | Ring up a test order for the changed item |
| Delivery app one | You, or the menu-sync feature if your system has one | Open the app on a phone and search for the item |
| Delivery app two | Same as above | Same check, on the second app |
| Google profile | You | View the menu section on a phone while signed out |
| Website | You or your web provider | Reload the page and read the price |
| Printed board or counter menu | Staff | Compare it with the sheet at opening |
Setup is about three hours: building the sheet, saving the prompt and writing that checklist. In the illustration it turns about 2.5 hours of weekly menu work into about one.
Review replies: AI drafts, a person posts
Twenty-five reviews a month is about 100 minutes of replies at four minutes each, if they all get answered, which they rarely do. Drafting with AI brings that to about 40 minutes. The time saved is small. The real gain is that every review gets an answer in one consistent voice. Give the assistant the facts, the rules and the review, and nothing else.
Draft a public reply to this review. Rules: under 70 words; thank the customer
by first name if shown; do not blame the driver or the delivery app; do not
promise a refund or discount; do not state anything I have not given you;
invite them to message us with their order number; sign off as [your name].
Facts: our kitchen ticket time was 14 minutes (from the order system). We
cannot see the courier's journey.
Review (2 stars, delivery app): "Chips cold and delivery took over an hour.
Not ordering again."An illustrative draft comes back like this.
Hi [first name], thank you for telling us. Cold chips and a long wait are not
what we want for anyone. Our kitchen had the order ready in 14 minutes, so the
delay may have happened on the delivery leg, but we would still like to look
into it. Please message us with your order number and we will check what we can.
[your name], [takeaway name]What to fix: the middle sentence edges towards blaming the courier, which the prompt ruled out. Cut it. Quoting the kitchen time invites an argument, so keep it only if you would say it to the customer's face. Never add order details to a public reply. And if a reviewer mentions illness or an injury, stop drafting and handle it yourself.
Two rules from Google's review policy matter here. You cannot offer a discount or gift in exchange for a review, or for changing or removing one, and you should not invite only your happiest customers to post. Replies can take a few minutes to appear and occasionally much longer, so do not post twice. Letting a tool post replies by itself is a bigger step, and the arguments are in whether AI should auto-post Google review replies. For a takeaway, drafts read by a person are enough.
Stock and prep: only as good as your item names
A takeaway's menu repeats, so the daily question is how much of each prepped item to have ready. A spreadsheet of weekday averages answers most of it. AI helps in two places: turning your till export into that sheet, and explaining the result in plain words.
The export is where it goes wrong. Order systems often list one dish under several names: "Spicy Wings 10", "10 spicy wings", "Wings x10 (promo)". Square's custom item report, for example, lists a renamed item as a separate entry, so a mid-year rename splits one dish into two lines. Merge the names before analysing anything, and mark the one-off weeks.
| Friday | Wings boxes sold | Note |
|---|---|---|
| Week 1 | 118 | |
| Week 2 | 124 | |
| Week 3 | 131 | |
| Week 4 | 120 | |
| Week 5 | 127 | |
| Week 6 | 163 | Big match on TV |
Ask for the working, not just the answer.
Here are six weeks of Friday sales of spicy wings (10-piece boxes): 118, 124,
131, 120, 127, 163. Week 6 was a one-off big-match night. Leave it out of the
normal average and say so. Suggest prep for a normal Friday and for a
match-night Friday, show the arithmetic, and tell me what else would change
your answer.A typical (illustrative) reply averages weeks 1 to 5 at 124 boxes, adds a 10% margin for a prep figure of about 136, and says the match night is a single data point, so 180 is a rough guide only. What to fix: recompute the average in your sheet with =AVERAGE, because models can slip on sums, and make sure the one-off week is flagged. Left in, week 6 lifts the average to 130.5 and normal Fridays are over-prepped by six or seven boxes.
Before trusting it, back-test: run the method on the last four weeks as if you did not know the answer, and compare it with what sold and with "same as last week". If the sheet is not closer, do not use it. If it is, the sums are modest. At about $5,500 a week of purchases, a 2% saving is $110 a week, about $475 a month, or roughly $400 once your 45 minutes a week are paid for. Even 0.4% ($22 a week) covers that time. The saving remains an assumption until the back-test shows it. For one site with a repeating menu I would not pay for a forecasting add-on before this sheet has had a month. The par-level tutorial for restaurants goes further on setting quantities.
The skip list, with reasons
- Allergy and dietary answers. A wrong answer harms a person, and the right answer lives in a written matrix that a named person signs off. Let the assistant format the matrix if you like; let staff answer the customer. Any phone or chat tool should transfer on words such as allergy, intolerance, coeliac or nuts. The wider argument is in should a chatbot answer allergen questions.
- Refunds and complaints involving illness or injury. Drafting is fine. The decision, the wording and the record of what happened are yours.
- Prices and discounts. AI can run margin sums in seconds, and you should check them in a spreadsheet. Deciding what a burger costs, or whether the apps get a higher price, stays with you because it depends on your commission plan and your regulars.
- Kitchen capacity. Pausing app orders or stretching quoted times is a call made at the pass by someone who can see the tickets. No tool sees your kitchen as well as the person standing in it.
- Anything posted unread. Review replies, social posts and food photos made by a generator that do not match the plate. If a customer can compare it with what they were served, someone should have looked at it first.
- Rota fairness and pay conversations. A draft rota is fine. The conversation about hours or pay is not a job for software.
Decoding the vendor pitch
The left-hand lines below are paraphrased from the way automation is commonly sold, not quotes from one supplier. The right-hand column is what to count in a trial.
| What the pitch says | What to ask | What to count |
|---|---|---|
| "Answers every call" | How many orders come out correct end to end, including mid-order changes and noisy lines? | Correct tickets out of twelve test calls |
| "Saves twenty hours a week" | Which tasks, on a menu the size of mine? | Hours you timed before and after |
| "Connects to your order system" | Does the order land as a kitchen ticket, or as an email someone re-keys? | Minutes of re-keying per order |
| "Simple pay-per-call pricing" | What counts as a call or a minute, do spam calls count, and is there a cap? | The bill for your busiest month |
| "Trained on thousands of restaurants" | Can I hear calls with an allergy question and a caller who changes their mind? | Whether allergy words trigger a transfer |
The trial order, with a stopping rule for each job
- Week 1. Run the missed-call tally and build the master menu sheet. Switch nothing else on.
- Week 2. Use the menu-change and review prompts live. Decision: do the drafts need fewer than two edits each? If not, tighten the rules in the prompt before going on.
- Week 3. Back-test the prep sheet on four past weeks. Decision: is it closer than "same as last week"?
- Week 4. Decide on phone orders with the tally and the menu sheet in hand. If you lose about three or more callers an evening and a plan fits the break-even sum, run twelve test calls. Keep the agent only if at least 11 produce a correct ticket and every call that mentions an allergy is transferred to a person. An allergy call handled by the agent alone ends the trial.
That order leaves every job with a number and a stopping rule. If you'd rather map this with someone, that's what my AI implementation consultation covers: finding the jobs worth automating first and handing over something your team can run.
Questions takeaway owners ask before switching anything on
Is AI worth it for a one-site takeaway?
Yes for the cheap jobs. Menu changes and review drafts need one to three hours of set-up and little or no extra cost, and the time saved is countable. Phone agents and stock forecasting only pay if your own numbers clear break-even, and a missed-call tally plus a four-week back-test will show that within a month.
Can AI take allergy questions on the phone?
It should not. Set any phone or chat tool to transfer to a person whenever a caller mentions an allergy, an intolerance or a dietary need, and keep a written allergen matrix that staff check. A wrong 'no nuts' can seriously harm someone, and a vendor demonstration rarely tests it, so make it part of your own test calls.
Should AI post my review replies automatically?
No. Draft with AI, read the draft, then post by hand. A reply that names the wrong order or blames a driver cannot be unsent, and reviews that mention illness or injury need a person from the start. Batch the drafts once or twice a week and the whole routine takes minutes.
Which job should I automate first?
Menu changes. Keep one master sheet that every platform copies from and let AI draft the update wording. It costs the least, pays back inside a month, and gives you the clean menu data that a phone agent or a prep forecast needs later.
Further reads
- Takeaway EPOS Systems Compared: Which AI Features Matter — Check which platforms your order system already updates for you.
- Takeaway Delivery App Menus: Rewrite Descriptions and Photos With AI — Rewrite delivery app descriptions once your menu sheet is clean.
- Can a Takeaway Take Orders on WhatsApp With AI? — A cheaper route to out-of-hours orders than a voice agent.
- Delivery App Commission: Should Takeaway Menu Prices Rise on Apps? — Work out whether app prices should differ from shop prices.
- What Are Diners Really Saying? Using AI to Analyse Your Reviews — Turn a month of reviews into the three things to fix.
- Restaurant AI Voice Agent: 12 Test Calls to Run Before Go-Live — The twelve test calls to run before a voice agent goes live.
- AI-Generated Marketing Photos: A Cafe Test of What Looks Honest — Same menu item three ways: what looks honest in a food photo.
- Kiosks AI for Cafés and Takeaways: Self-Service Ordering Costs — Hardware, software fees, menu set-up and a break-even sum for a self-service kiosk, plus the cases where a person at the counter still wins.
- Pizza Shop Inventory: Plan Dough and Toppings From Weekend Orders — Turn eight weeks of orders into dough balls, cheese and topping quantities for Friday to Sunday, then tune the plan from a simple waste log.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Maple, Smith.ai (AI Receptionist) and Yelp Host pricing pages, checked October 2026; Square's help page on the custom item report; Google's review and contribution policy pages.