Can a Takeaway Take Orders on WhatsApp With AI?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Can a Takeaway Take Orders on WhatsApp With AI?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Can a Takeaway Take Orders on WhatsApp With AI?

Yes. A takeaway can take WhatsApp orders with AI by connecting an ordering bot to the WhatsApp Business Platform through a provider, or by using Meta's own agent in the Business app where it's offered. The AI reads the customer's message, matches it to your menu, confirms items, time and total, sends a payment link, and passes the ticket to the kitchen.

The chat is the easy part. What decides whether it works is what happens after it: getting the order into your kitchen printer or till without someone retyping it, handling "no onions, extra chilli" without mistakes, and never letting the bot answer an allergen question on its own. Costs also changed this autumn, because Meta starts charging for replies inside the 24-hour customer window on 1 October 2026.

Follow me on Instagram@sagnikteaches

Three ways takeaways take WhatsApp orders today

Set-upHow the order arrivesAI involved?Suits
WhatsApp Business app with catalogue and cartCustomer picks items from your catalogue and sends a cart; staff confirm by handNo, apart from quick replies and greeting or away messages you writeA handful of WhatsApp orders a night
Meta Business Agent in the Business appMeta's AI answers from your business information and can hand over to youYes, Meta's ownOwners who want AI without a separate provider, where the feature is available
AI ordering bot on the Business Platform via a providerBot takes the order in free text, confirms it, sends a payment link, pushes it to the kitchen systemYes, configured by you or the providerRegular WhatsApp volume and a till or kitchen system the provider can connect to

If you're unsure which side of the app-versus-platform line you're on, WhatsApp Business app vs API for AI replies explains the difference in plain terms.

Connect on LinkedInSagnik Bhattacharya

What the AI does with one real order message

Here is a typical Friday message, and what a well-configured bot does with it.

Subscribe on YouTube@codingliquids

hiya can I get 2 large margherita one with no basil, garlic bread, 2 cans coke. collection about half 7

  1. Parses the items against your menu: two large margheritas, one with the "no basil" modifier; one garlic bread; two cans of cola (your menu might list the brand, a flavour choice, or a size).
  2. Checks what it can't assume. "Garlic bread" might come plain or with cheese on your menu. A good bot asks rather than guesses.
  3. Checks the slot. If 7.30 is already at your kitchen's capacity, it offers 7.45 instead of accepting blindly.
  4. Confirms in a fixed format so the customer can spot errors at a glance.
  5. Sends the payment link once the customer replies "yes", or records cash on collection if you allow it.
  6. Pushes the ticket to your kitchen screen, printer or till, and sends a short "order received" message.

The confirmation is the step that prevents most complaints. Keep it rigid:

Your order, please check:
- 2 x Large Margherita (1 with NO basil)
- 1 x Garlic Bread (cheese)
- 2 x Can of cola
Collection: 7.45pm (7.30 was full)
Total: $[total]
Reply YES to confirm and I'll send a payment link,
or tell me what to change.

Whether it's worth it depends on four things

How many orders already arrive by WhatsApp

If customers message you a few times a night, the free Business app with a catalogue and quick replies is enough, and a bot adds cost for little gain. The case for AI starts when WhatsApp orders arrive faster than someone on the pass can answer them, or when unanswered messages on a busy night become lost orders.

How complicated your menu is

A fish-and-chip or pizza menu with a few sizes and toppings is easy for a bot. Build-your-own curries with spice levels, half-and-half pizzas, meal deals with swaps and multi-buy offers are where bots misprice orders. Test your ten most-changed dishes before going live.

Whether orders can reach the kitchen without retyping

If a member of staff has to copy each WhatsApp order into the till, you have moved the bottleneck, not removed it. Ask any provider which till and kitchen systems they connect to, and see a live order land in yours during the demo.

How you take payment

Unpaid collection orders that never turn up are the takeaway's oldest problem. A bot that sends a card payment link before the ticket prints removes most of that risk. If you allow cash on collection, set a limit for first-time numbers.

What Meta charges from October 2026, and what a provider adds

Meta bills WhatsApp Business Platform messages per message. Until now, every reply a business sent inside the 24-hour window after a customer messaged was free, which suited ordering well. Meta's developer documentation says that from 1 October 2026 those service replies become chargeable per message, at the same rate as utility messages in the customer's market. Each business phone number gets its first 1,000 service messages a month free, and you'll need a payment method on your WhatsApp Business Account. Rates vary widely by the customer's market, so look up yours on Meta's rate card rather than trusting a figure from someone else's.

One detail catches ordering set-ups out. Some providers send the receipt and "your order is ready" messages as utility templates (pre-approved message formats) rather than as ordinary replies, because the kitchen system triggers them rather than a customer message. From 1 October 2026, utility templates sent inside the 24-hour window are charged from the first one, with no free allowance. Ask your provider which of the bot's five messages go out as templates; moving two of them into ordinary replies can take them inside the free 1,000.

Meta's own agent is billed differently: since 1 August 2026, Meta Business Agent has been charged per token, $2 per million tokens, which Meta estimates at roughly 4 to 5 cents a message, with no free window. At ordering volumes that adds up faster than people expect.

Put illustrative numbers on it. A takeaway receiving 400 WhatsApp orders a month, with the bot sending five messages per order (clarifying question, confirmation, payment link, receipt, ready for collection), sends about 2,000 service messages. After the 1,000 free, 1,000 are charged. At 2 cents each that is $20 a month to Meta; at a quarter of a cent it is $2.50. The provider's platform fee and the AI model usage usually cost more than Meta's share, so ask for both as monthly figures at your volume. What a WhatsApp chatbot costs breaks down the provider side.

Run the same 400 orders through Meta Business Agent instead and the sum changes shape. Five AI messages per order is 2,000 messages at 4 to 5 cents, so $80 to $100 a month, every month, with no allowance to absorb the first thousand. That can still be cheaper than a provider's platform fee, but it buys less: Meta's agent answers from your business information and hands over to you, whereas a provider's bot is usually the one that pushes the ticket to your kitchen screen. Compare on what reaches the kitchen, not just the monthly figure.

Where WhatsApp ordering bots go wrong on a Friday night

  • Voice notes. Some customers send audio. Check whether the provider transcribes them, and if not, have the bot reply asking for a typed order.
  • "Same as last time." Only works if the bot can see order history tied to the number. If it can't, it must say so rather than invent an order.
  • Sold-out items. If stock isn't synced, the bot takes an order for the last of the lamb. Give staff one button or message to mark items unavailable, and make sure the bot reads it.
  • Delivery addresses. Customers type half an address or a landmark. The bot should confirm the full address and check it against your delivery radius before quoting a time. "The flats behind the school, top floor" is a real kind of answer; the right reply asks for the building number or name and a contact number for the driver, and doesn't quote a delivery time until it has a full address it can check against your streets list.
  • Offers and codes. The bot applies a discount that expired last month because nobody removed it from its instructions. Keep offers in one list with end dates.
  • Allergen questions. This is the one that matters most. The bot must not say a dish is safe for an allergy. It should hand the chat to a person or tell the customer to call. The reasoning is in whether an AI chatbot should answer allergen questions.
  • Complaints mid-rush. "Where's my food, it's been an hour" needs a human quickly. Set the bot to flag any message with delay or complaint words and alert the shop phone.

One mistake deserves its own paragraph because it is so common. Picture a kebab and burger shop that runs a "meal deal: any burger, fries and a can for $11.50". The bot was given the menu with individual prices and a one-line note about the deal. For two weeks it charged some customers the deal price and others the full item prices ($14.20), depending on how they worded the order. Nobody noticed until a regular complained that his usual order cost more on WhatsApp than at the counter. The fix was to list the deal as its own menu item with its own price, and to add a rule: "If the order contains a burger, fries and a can, offer the meal deal." Test every offer you run with three differently worded orders before going live.

A tricky chat, and what the bot should have said

Here is an illustrative exchange with a bot on its default settings:

Customer: same as last time pls, to the usual address
Bot:      Great choice! I've added 1 x Chicken Tikka Masala, 1 x Pilau
          Rice and 1 x Garlic Naan. Delivery to 14 Mill Street.
          Total $24.40. Shall I send the payment link?

If the bot really can see this customer's order history, fine. If it can't, it has just invented an order and an address, and the customer may say "yes" without reading. The reply you want is:

Bot:      I can't see previous orders on WhatsApp, sorry. Could you
          tell me what you'd like and the delivery address? If it
          helps, our menu is here: [link]

Ask the provider directly: "Can the bot see past orders for this number, and what does it say if it can't?" Then test it yourself with a number that has never ordered.

Pizza shop or curry house: where the bot works hardest

In a pizza shop, most WhatsApp orders are variations on a dozen items, so the bot's job is mostly sizes, toppings and half-and-half pizzas. The risk sits in modifiers ("no cheese" on one of three pizzas) and in capacity on Friday between 6 and 8. A good set-up shows the kitchen screen each pizza on its own line with its modifier in capitals. Take the message "3 large, 2 pepperoni and one half veggie half hawaiian, no cheese on one of the pepperonis for my son". A ticket that prints it the way the customer typed it gets made wrong at speed. The ticket that works on a busy pass looks like this:

#0417  COLLECTION 19:15  PAID
1  LARGE PEPPERONI
1  LARGE PEPPERONI   ** NO CHEESE **
1  LARGE HALF/HALF   L: VEGETARIAN  R: HAWAIIAN

Test this before launch by sending the bot five orders with modifiers buried mid-sentence and checking that each one lands on the right pizza, not the first or last one mentioned.

In a curry house, orders are longer and more personal: spice levels, "the one with spinach", rice swaps and a side of chutney for each person. The bot needs your menu with every common nickname mapped to a dish ("chicken saag" and "spinach chicken" both point to the same item), and it should read the whole order back line by line. Allergen questions are also more frequent, especially nuts in sauces, so the handover to staff has to be quick.

The instructions to give your ordering bot

Whether you configure it yourself or a provider does, the bot's instructions should cover the points below. Keep them narrow: Meta's platform terms prohibit general-purpose AI assistants, and a bot that only handles your menu and orders stays well within the business-specific use Meta allows.

You take food orders for [takeaway name] on WhatsApp.
Only discuss our menu, opening hours, collection and delivery.
Politely decline anything else.

Menu and prices: [attach current menu with modifiers]
Opening hours: [hours]. Last orders: [time].
Delivery: only within [radius or list of streets]; minimum order [amount].
Kitchen capacity: max [number] orders per 15-minute slot.

Rules:
- Always confirm the full order, time and total in the fixed format.
- Ask when an item has options; never guess.
- Never say a dish is safe for any allergy. Reply: "I'll pass you to
  the team for allergy questions" and hand over.
- Hand over to staff for complaints, refunds, late orders, or if the
  customer asks for a person.
- Send the payment link only after the customer replies YES.
- Tell the customer at the start that you are the shop's automated
  ordering assistant.

Should your takeaway switch it on this month?

Work through these in order and stop at the first "no":

  1. Do you get enough WhatsApp orders that messages go unanswered on busy nights? If not, set up the catalogue and quick replies in the free app and revisit in three months.
  2. Can the provider put orders straight into your till or kitchen system? If not, the bot will add a retyping job on your busiest shift.
  3. Will you send payment links before cooking? If not, the no-show risk rises with every order the bot accepts.
  4. Have you tested your ten trickiest menu items and three allergen questions in a trial account? If the bot fails any of them, fix the instructions before customers see it.

If all four are yes, run it on quieter weeknights first, read every conversation for two weeks, and only then leave it on for Friday and Saturday. The failures you find on a Tuesday are much cheaper than the ones you find at 8pm on a Saturday.

Measuring whether the bot is earning its keep

Four numbers, tracked weekly from the conversation log and your till, tell you whether the bot is helping or just moving work around:

  • Orders completed by the bot without staff help. Divide by all WhatsApp orders. If fewer than about two in three finish without a human stepping in, the menu or instructions need work before the bot saves anyone time.
  • Corrections after confirmation. Count orders where the kitchen or the customer had to fix something after the customer said YES. Each one is a wrong item or a wrong price the confirmation should have caught. Look for patterns: the same modifier, the same meal deal, the same late-night slot.
  • Abandoned chats. Customers who start an order and never confirm. A few are normal. A rising number usually means the bot is asking too many questions or the payment link is failing on some phones.
  • Handover response time. When the bot passes a chat to staff, how long until someone replies? A bot that hands over properly but then leaves the customer waiting 20 minutes on a Saturday is worse than no bot.

Put the four numbers on one line each week. A filled-in line for an illustrative third week reads: "96 WhatsApp orders; 71 finished by the bot alone (74%); 5 corrected after YES, 3 of them half-and-half pizzas; 9 abandoned, 6 of those at the payment link; median handover reply 4 minutes, worst 22 minutes on Saturday at 7.40pm." That line tells you three things to do: rewrite how the bot confirms half-and-half pizzas, test the payment link on an older phone, and decide who answers handovers during the Saturday peak. After a month you will know whether to keep it, fix the menu set-up, or go back to the free app with a catalogue. Whichever you choose, keep reading a sample of conversations every week: bots drift when menus and offers change, and the log is the only place you'll see it before customers tell you. When a chatbot should hand over to a human covers the handover rules in more depth.

Takeaway WhatsApp ordering: follow-up questions

Can I keep using WhatsApp on my phone if I connect an AI ordering bot?

Usually yes. Meta supports a set-up it calls coexistence, where the same number runs on the WhatsApp Business app and the Business Platform at once, with messages mirrored between them. You need a recent version of the app, and a few features such as broadcast lists and disappearing messages are switched off once you connect. Ask your provider whether they support it before you sign.

Does Meta allow AI chatbots on WhatsApp at all?

Meta's platform terms ban general-purpose AI assistants, meaning bots that chat about anything. A bot that only takes orders, answers menu questions and hands over to staff is a business-specific use, which is the kind Meta still permits. Keep your bot's instructions narrow and have it decline unrelated requests politely.

Can customers pay inside WhatsApp?

In-chat payment depends on the country and the provider, so most takeaways send a secure payment link from their card processor instead. The bot sends the link after the customer confirms the order, and the ticket only goes to the kitchen once payment clears or the customer chooses cash on collection.

Should the bot tell customers it is automated?

Yes. Customers forgive a bot that says what it is and hands over quickly far more readily than one that pretends to be a person. If you sell to customers in the EU, the AI Act's transparency rules also expect people to be told they are dealing with an AI system. A one-line greeting covers it.

Further reads

Sources: Meta for Developers WhatsApp pricing pages, including the upcoming pricing updates page; Meta Terms for WhatsApp Business Platform section 4.7; Meta's coexistence onboarding documentation; WhatsApp help pages on catalogues and carts; 360dialog's October 2026 service message notice.

Want WhatsApp orders going straight to your kitchen?

On a 1:1 call we'll look at how orders reach you now, check whether your till or ordering system can accept WhatsApp orders without retyping, and decide whether a bot or a simpler catalogue set-up fits your volume.

Book a 1:1 call with me