How to Set Up an AI Phone Line for Takeaway Orders

To set up an AI phone line for takeaway orders, choose a system that connects to your till, such as Square's AI voice ordering or a restaurant specialist like Loman. Load your full menu with every option, decide how callers pay, forward your number, then test with 30 or more scripted calls before switching it on for a quiet evening.

Expect the whole job to take one to three weeks, and expect most of that time to go on menu data and testing rather than the AI itself. Below: which route fits your takeaway, what to prepare, how to test it, and what to watch in the first month. One rule from the start: the AI should never be the one answering allergen questions.

Follow me on Instagram@sagnikteaches

Three ways to get an AI answering takeaway calls

There are three broad routes. Which one fits depends mostly on your till, because an AI that takes an order but can't put it into your kitchen system just moves the typing from the phone to a screen.

Connect on LinkedInSagnik Bhattacharya

The details for Square come from its support article on taking orders with AI-powered voice ordering; Square says most businesses finish setup in about five business days. If you're weighing the cost against a missed-call problem, what AI phone answering costs and pays back does that sum in detail. If you're deciding whether phone orders should exist at all versus delivery apps, see AI ordering versus delivery apps.

Subscribe on YouTube@codingliquids

Step 1: count and sort a week of calls (one week, 10 minutes a day)

Before choosing anything, find out what the phone is actually doing. Keep a tally sheet by the phone for one normal week, and download your phone provider's missed-call log.

A completed sheet from an illustrative kebab and pizza takeaway reads: new orders 48 Monday to Thursday and 96 on Friday and Saturday; "are you open / how long?" 22 and 15; changes to an order 4 and 9; allergen questions 3 and 5; complaints 2 and 4; missed calls 11 and 38. Three conclusions come straight off it. The problem is the weekend, where 38 missed calls sit beside 96 orders answered, so a "forward when unanswered" set-up for Friday and Saturday would catch most of the loss. The 37 "are you open?" calls are a sign the Google listing's hours or collection times aren't clear, which is worth fixing whatever you buy. And eight allergen calls a week means the transfer to staff has to work every single time.

This tells you three things: whether there are enough missed orders to justify the cost, what your AI must handle beyond orders, and when peak demand hits. If most calls are "are you open?", note that Square's voice ordering is built for orders, not questions, which pushes you towards a specialist.

Step 2: make the menu machine-readable (two to six hours)

This is where most setups go wrong. Your printed menu is written for humans who can ask "is that with rice?". The AI needs every choice spelled out. Work through each item and list:

• Sizes and prices for each size, including any "large" upgrades in meal deals.
• Required choices (rice or chips, spice level, sauce) and optional extras with prices.
• Removals customers commonly ask for ("no onions") and whether they're allowed.
• The names customers really use. If people ask for "the chicken thing with the sauce", write down which item that is. A list of five or six common nicknames saves many confused calls.
• Combination rules: half-and-half pizzas, meal deals that include "any small side", minimum order for delivery.
• What runs out, and how staff will mark it unavailable during service.

If your till already holds the menu with modifiers, the AI system will usually read it from there. Check it anyway. A till menu built years ago often has "Special 2" where the AI needs real names, and hidden modifiers that staff know about but the menu data doesn't show.

Before and after, for one illustrative item. The till entry: "Special 2 – $11.50". What the AI needs:

Mixed kebab: lamb doner and chicken shish. Sizes: regular $11.50, large $14.50. Required: bread (pitta or naan); sauce (garlic, chilli, both or none). Optional: cheese +$1, extra meat +$3. Removals allowed: onion, salad, chilli. Customers call it: "the mix", "mixed meat", "number two", "special two".

Every line in the second version answers a question the AI would otherwise guess at, or ask the caller in a way that slows the order. The nicknames line matters most: regulars ordering "a large number two with garlic" is exactly the call that fails when the menu data only says "Special 2".

Step 3: decide how callers pay

You have four options, and your route may decide for you:

1. Payment link by text. The caller gets a link and pays online. Safe and simple; some callers won't bother, especially older regulars. Square's system works this way and won't process unpaid orders.
2. Secure card payment on the call. Some specialists take card details using a process designed for the card industry's PCI DSS security standard. Ask exactly how it works before you sign.
3. Pay on collection. The easiest for callers, but you carry the risk of orders nobody collects. Many takeaways allow it only for known numbers or small orders.
4. Cash or card on delivery. Workable if your drivers carry a card reader; same no-show risk.

Whatever you choose, never let card numbers be read to the AI and typed into anything by a person, or saved in call transcripts.

Step 4: write the rules the AI must follow (one to two hours)

Every system has somewhere to set instructions, whether it's a form in a dashboard or a written prompt on a voice platform. Write them down in plain language first, then enter them. A template to adapt:

GREETING: "Thanks for calling [name]. I'm the automated ordering line.
I can take your order, or say 'staff' at any time for a person."
COLLECTION TIMES: quote [20] minutes Mon-Thu, [35] minutes Fri-Sat 6-9pm.
  If the kitchen marks "busy" in the till, add [15] minutes.
DELIVERY: only to [list of areas / within X miles]. Minimum order [$].
  Outside the area: offer collection instead.
ALWAYS TRANSFER TO STAFF:
  - any allergen, ingredient or dietary-safety question
  - complaints, refunds, missing items
  - orders over [$150] or for more than [10] people
  - the caller asks for a person, or isn't understood after two tries
IF NOBODY ANSWERS THE TRANSFER: take name and number, promise a callback
  within [10] minutes, and send it to [staff phone / till message].
CONFIRM EVERY ORDER: read back items, options, total, collection or
  delivery time, and name, before ending the call.

The allergen rule is non-negotiable. An AI can mishear "no nuts", misremember a recipe change or answer from outdated menu data, and the consequence can be someone's health. Tell callers clearly that allergen questions go to staff.

Step 5: connect your number without losing calls

You rarely need a new number. Most setups use call forwarding from your existing line, and the Square route requires it, because its AI answers only Square-provided business numbers. Your options, set with your phone provider:

• Forward all calls to the AI. Simplest, but every regular meets the AI.
• Forward when busy. Staff answer when free; the AI takes the overflow. Good for takeaways where regulars like a person.
• Forward when unanswered after, say, four rings. Catches the calls you currently miss.

Many takeaways start with "forward when unanswered" during the week and "forward all" at Friday and Saturday peaks. Test that caller ID passes through, because the AI's callback and "same as last time" features depend on it.

Step 6: test with scripted calls (a day, spread over two sessions)

Don't launch after three friendly test calls. Write 30 to 40 test calls and have friends or staff ring from different phones, including outside and in a noisy room. Include:

• Simple orders, orders with four or five modifiers, and half-and-half pizzas or meal deals.
• A caller who changes their mind twice, and one who adds an item after the read-back.
• Menu nicknames, an item you don't sell, and an item marked out of stock.
• A delivery address outside your area, and one with a flat number that's easy to mishear.
• An allergen question, a complaint, and "can I speak to someone?" (all three must transfer).
• Strong accents, a child's voice, background television, a caller who mumbles.

One failure worth expecting, from an illustrative test call. The caller says: "Can I get a large mix with naan and garlic, no onions, and a twelve-inch half pepperoni, half veggie." The AI reads back: "One large mixed kebab with naan and garlic sauce, no onions. One 12-inch pepperoni pizza and one 12-inch vegetable pizza. That's $38.50." It heard "half and half" as two pizzas and nearly doubled the pizza price. Nothing in the menu data told it that half-and-half is a single pizza with two toppings, charged at the higher of the two prices (or whatever your rule is). Add half-and-half as a named option with its pricing rule, then rerun every test that involves one.

Score each call on a sheet: order correct, total correct, time quoted correctly, transfer worked, caller would have been happy. I'd want at least 19 out of 20 orders exactly right before going live. Common failures at this stage are fixable: missing nicknames, a modifier set up as optional when it should be required, a collection time that ignores the busy setting. The checks in testing a customer chatbot before it goes live carry over to voice.

A few rows from a filled-in score sheet (illustrative):

Test 31 shows the read-back doing its job; tests 14 and 36 are menu and rules problems to fix before launch, not reasons to give up.

Step 7: launch on a quiet night, then read the first 100 calls

Go live on a Monday or Tuesday, not a Friday. Put a sign on the counter and a line on printed menus telling regulars what to expect. For the first two weeks, someone spends 15 minutes a day reading call transcripts or listening to recordings, looking for wrong items, poor transfers and callers who hung up. Reviewing AI call transcripts for quality sets out a routine for this.

The numbers to track weekly: phone orders taken, orders with a mistake, calls transferred, calls abandoned in the first 20 seconds, and missed calls in your phone log. Compare with the tally from Step 1. If mistakes run above about one in 25 orders after the second week, pause the Friday "forward all" setting and fix the menu data before carrying on.

Two illustrative weeks side by side show how to read them. Week 1: 212 phone orders, 11 with a mistake (about 1 in 19), 34 transfers, 19 calls abandoned in the first 20 seconds, 9 missed calls. Week 2, after adding six nicknames and fixing the half-and-half rule: 231 orders, 6 mistakes (about 1 in 38), 28 transfers, 11 early hang-ups, 7 missed calls. Week 1 was above the one-in-25 line, which is normal for a first week; week 2 is comfortably below it. The early hang-ups halving suggests regulars are getting used to the line, and missed calls down from the 38 a week in the original tally is the number that pays for it.

A worked example: a Friday-night takeaway

Say a pizza and kebab takeaway takes 70 phone orders on a Friday and the missed-call log shows 25 unanswered calls between 6pm and 8pm. The average order is $28. The owner can't know how many missed calls were orders, so they assume half: about 12 orders, or $336 of lost sales each Friday.

The takeaway runs a till that isn't Square, so it looks at a specialist and at a build-your-own option. For the build-your-own route, around 400 AI-handled calls a month at an average of three minutes each is 1,200 minutes. At a mid-range $0.15 a minute that's about $180 a month in usage, plus a developer's time to connect the till. The specialist quote, which is monthly and doesn't charge per minute, turns out simpler because the till connection already exists. The test is simple arithmetic: three extra orders every Friday at $28 is about $360 a month in sales. Set that, after food and packaging costs, against the monthly price. If the recovered orders don't appear in the weekly numbers within a month, the line isn't earning its keep.

Problems that show up in the first month

• Misheard addresses. Make the AI read the address back, spelling the street name for unusual ones, and ask for a phone number on every delivery order.
• Collection times that ignore a slammed kitchen. Link the quote to the till's busy setting or raise the peak-time default. A 20-minute promise that becomes 45 creates complaints the AI then can't handle.
• "Same as last time". Only works if the system recognises the caller's number and can see past orders. If it can't, it must say so and take the order fresh.
• Special requests. "Extra crispy", "cut into squares", "don't ring the bell, the baby's asleep". Decide which the AI can pass to the kitchen as a note and which it should decline. A filled-in list might read: pass as a kitchen or driver note: extra crispy, cut into squares, sauce on the side, don't ring the bell, leave at the side door; decline politely: anything off the menu, "make it like last time but different"; transfer to staff: any question about how food is prepared for a religious or dietary reason, and any request involving a discount.
• Transfers that ring out. If staff can't pick up during a rush, the callback promise must be kept. Put the callback list somewhere the counter team will see it. For when to hand over and how, see when an AI should hand over to a human.

An AI phone line works best as a second pair of hands for the busiest two hours, with clear limits, rather than a replacement for the person at the counter. Set it up that way and the first month will tell you, in plain numbers, whether it's paying for itself.

Questions takeaway owners ask about AI phone ordering

Will regular customers hang up on an AI voice?

Some will, especially at first. Tell regulars before launch with a counter sign and a line on your menu, keep the greeting short, and give callers an easy way to reach a person, such as saying "staff". Then watch the share of calls that end within the first 20 seconds. If it stays high after the first fortnight, shorten the greeting or change the voice.

Can the AI take card payments over the phone?

Some systems can, using a secure payment process, but many send the caller a text message with a payment link instead. Square's voice ordering, for example, requires every order to be paid through a texted link before it's processed. Avoid any setup where a person or an unsecured system writes down card numbers, and ask vendors how they meet the card industry's PCI DSS security standard.

What if the AI can't understand a caller?

A good setup asks once more, then transfers the call to a person or takes a callback number. Set that rule explicitly during setup rather than relying on defaults. If nobody can pick up during a rush, the fallback should be a clear message and a quick callback, not a loop of the AI asking the same question three times.

Do I still need staff answering the phone?

Yes, for the calls the AI hands over: allergen questions, complaints, large or unusual orders, and anyone who asks for a person. The aim is that staff answer fewer calls, at better moments, not none. In most takeaways that means one person keeps a handset nearby during peak hours.

Further reads

• Can AI Answer the Phone and Take Restaurant Reservations? — If you also take table bookings by phone.
• Can AI Answer My Business Phone? What It Can and Can't Handle — What AI voice agents handle well and badly, in any business.
• What to Do When an AI Receptionist Gets a Booking Wrong — What to do when the AI gets an order or booking wrong.
• Should an AI Chatbot Answer Allergen Questions for Your Restaurant? — Why allergen questions should go to a person.
• Your First 30 Days of AI in a Restaurant, Week by Week — A week-by-week plan for a food business's first AI month.
• AI Call Summaries: Log Every Phone Enquiry in Your CRM — Log every call automatically so nothing gets lost.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Square support article on AI-powered voice ordering; Loman pricing page; Retell AI pricing page, all checked September 2026. Call volumes, order values and costs in the worked example are illustrative.