AI phone ordering costs a pizza shop about $220 to $350 a month on a flat plan that puts orders into the till, or roughly $0.08 to $0.13 a minute on a voice platform. It pays back when saved Friday orders exceed that bill: about 12 to 20 extra orders a month.
That range hides three billing models and several extras. The sections below set them side by side, work out the payback on a made-up Friday, and list the costs that never appear on a pricing page. All prices are USD list prices read from vendor pages in October 2026; check each page before you sign.
The three ways phone AI is billed
Before comparing prices, find out what the unit is. A $0.11 minute and a $220 month cannot be compared until you know how many minutes your shop uses.
Per minute
Voice platforms such as Retell and Vapi charge by the minute and leave the rest to you. Retell's headline is $0.07 to $0.31 a minute, built from a voice engine at $0.055, a voice at about $0.015, a language model from a few thousandths of a dollar up, and telephony at $0.015. Add up a typical build (the engine, a standard voice, telephony and a low-cost to mid-range model) and you land between about $0.09 and $0.15 a minute; the sums below use $0.11 as a midpoint. Phone numbers are listed at $2 a month. Vapi's page estimates $82 to $129 a month for 1,000 minutes, with hosting at $0.05 a minute plus transcription, model and voice costs passed through at cost.
These are building blocks. They do not come with a pizza menu, an understanding of half-and-half toppings, or a connection to your till. Someone has to build and test that, and the bill for that person is not on the page.
Flat monthly plan
Restaurant-specific vendors sell a fixed monthly plan. Maple's page lists Maple Voice at $85 a month on annual billing or $150 monthly, which answers calls, gives hours and FAQs, and transfers calls, but does not take orders. Maple Pro is $220 a month on annual billing or $350 monthly, and adds phone orders into a point-of-sale system (it names Toast, Square, Clover and Shift4 Dine among others), upselling, payment over the phone and order confirmation texts. The page says there are no per-minute fees and states no call allowance or setup fee, so ask both questions before you rely on that.
Per unique customer
A third model bills by how many different phone numbers called. Goodcall's page lists Starter at $79 a month for 100 unique customers, Growth at $129 for 500, and Scale at $299 for 2,000, with overage at 79, 26 and 15 cents per extra customer. A customer counts once a month, whatever the number of calls. The page I read does not mention taking food orders, so treat it as an answering plan for a pizza shop until the vendor shows otherwise.
Some phone and POS packages bundle a basic assistant instead. Toast IQ's base assistant, for example, is included with Toast POS, but how much of a full phone order it handles is not something its pricing makes clear, so ask your POS vendor before assuming a separate bill is needed.
The same 600 calls under each model
Here is an illustrative month for a single shop: 600 calls, about 450 different callers, with 2 minutes per straightforward call or 3.5 minutes per call if customers change toppings, ask about allergens and spell out addresses. The sums are mine, using the list prices above.
| Billing model | Easy calls (2 min each, 1,200 min) | Messy calls (3.5 min each, 2,100 min) |
|---|---|---|
| Retell at an illustrative $0.11 a minute | $132 (plus $2 for a number) | $231 (plus $2) |
| Vapi at its quoted spread, about $0.082 to $0.129 a minute | about $98 to $155 | about $172 to $271 |
| Maple Pro, annual billing | $220 | $220 |
| Maple Pro, monthly billing | $350 | $350 |
| Goodcall Growth (450 customers, within 500) | $129, answering only | $129, answering only |
Read it as a pattern, not a verdict. The per-minute builds are cheapest when calls are short and your menu is simple, and they rise when callers are chatty. The flat plan costs the same on a quiet month and a record one, which suits a shop whose volume swings. The Vapi range uses its published 1,000-minute estimate scaled up, so the real figure depends on the model and voice you pick.
The cheapest line in the table is not the cheapest system. A per-minute build with no till connection means a staff member re-types every order. If that takes two minutes per order, 600 orders cost 20 hours a month, which at a $15 hourly wage is $300. That is more than the flat plan you thought you were avoiding.
Missed-call maths on a Friday peak
Payback comes from calls nobody answers. The numbers below are invented to show the method; replace each with your own from last month's phone log.
Say a single shop is busiest on Friday and Saturday from 5 to 9 pm, with 70 calls an evening. Everyone is boxing pizzas, so 12% go unanswered or hang up on hold: about 8 calls. Seven in ten of those callers wanted to order, which is 5.6. Half of them ring a competitor and never return, which is 2.8 lost orders an evening. At an average ticket of $32, that is about $90 an evening. Across 8 peak evenings a month it comes to $717 of lost sales.
Sales are not profit. If food, packaging and driver costs take 45% of the ticket, you keep about 55%, or $394 a month. Against a $220 plan that leaves about $174 a month before the cost of any set-up time, and the system pays back. Against a $350 monthly plan it does not, on the same numbers. This is why billing frequency and the real volume of lost calls decide the outcome, not the headline price.
The break-even rule is simple. Divide the monthly cost by the profit on one order. $220 divided by $17.60 ($32 at 55%) is 12.5 orders. $350 divided by the same is about 20. If you do not recover that many orders a month from calls that used to go unanswered, the plan does not pay back on call recovery alone, though it may still save staff time on quiet-hour calls.
What the saving looks like when the shop is quiet
The other side is time. A person answering routine calls ("What time do you close? Do you do gluten-free?") spends maybe 30 seconds each. If 40% of calls are such questions, a flat plan that handles them frees a few hours of attention a week, but unless you cut a shift it does not appear as cash. Count it as a bonus, not as the reason to buy.
Costs that never show on the pricing page
- POS connection. Check the integration with your exact POS and version. Without it, orders arrive as emails or texts someone has to key in.
- Set-up and menu loading. Maple Pro mentions white-glove onboarding, but a setup fee, if any, is not on the page I read. On a per-minute build, count your own hours. At 15 hours of unpaid owner time, valued at $25 an hour, you have spent $375 before the first order.
- Payment handling. Taking card details over the phone brings processing fees and compliance tasks. Many shops prefer a text with a payment link, so check which your vendor offers.
- Call forwarding and number costs. You may need a new number or a forwarding rule from your existing line. Retell lists phone numbers at $2 a month; another provider may charge more.
- Transfers and failures. When the AI cannot handle a call and transfers it, a person picks up mid-rush. Budget for that.
- Review time. Someone has to listen to calls or read transcripts weekly. One hour a week is a realistic starting figure.
Twenty Friday-style test orders before you commit
Run 20 calls, ideally from phones other than your own, with a mix of the awkward cases a Friday brings. Score each against the same sheet.
Call 1: Large pepperoni, thin crust, extra olives. Pay on delivery.
Call 2: Half margherita, half ham and pineapple. Ask for no onions.
Call 3: Ask whether the dough is vegan. Ask about a nut allergy.
Call 4: Order, then change address after the order is confirmed.
Call 5: Loud background noise; give a street name with an unusual spelling.
Call 6: Ask for a refund on last night's order.
Call 7: Order for collection in 20 minutes; ask for a time change.
Pass mark: correct order in the POS on 18 of 20; transfers to a person
on every allergy and refund call.A condensed, illustrative failure looks like this: the system hears "half ham and pineapple" and enters a whole ham and pineapple pizza at the full price. The order text goes out, the customer does not notice, and the shop discovers it when the customer rings. The fix is a confirmation step: the system reads the order back by text with the split shown, and the customer replies to confirm. Ask any vendor to show you that step in a live test.
Use a short script for the human side too. When a call transfers, the person should hear a one-line summary before speaking, such as "Caller wants a refund on order 214, items arrived cold". If the vendor does not provide a summary, that missing detail is a cost.
Three shop sizes, three different answers
The same price list gives different verdicts depending on the shop. These are illustrative profiles, and the ticket and margin figures carry over from the sum above ($32 average ticket, 55% kept).
The quiet neighbourhood shop. About 250 calls a month, one person on the phone and the counter, and perhaps 10 missed calls on a busy Friday. If 4 of those would have been orders, that is roughly 4 times $17.60, or $70 of profit a month. No ordering plan clears that. A cheaper fix is a voicemail greeting that gives the online ordering link, plus a text-back to missed calls. Cost it before buying a phone system.
The busy two-oven shop. Around 900 calls a month and a clear Friday and Saturday jam. Lost sales on the earlier sum were $717 a month, about $394 of profit. Maple Pro on annual billing clears that with $174 to spare, and the monthly price does not. This shop should test on monthly billing for a month, then move to annual if the missed-call count falls.
The late-night shop with a long tail. Calls arrive between 10 pm and 2 am, when nobody wants to staff the phone. Here the comparison is not against lost calls but against a wage. A closing-time employee on the phone for four hours at $15 an hour is $60 a night, or about $1,800 a month if that is fully attributable. An answering plan costing $350 against that saving is an easy decision, provided the orders come through correctly. Compare it with the person doing other jobs at that time, though; the wage is only saved if the shift changes.
Questions to put to a vendor in writing
A flat monthly price can hide limits. Put these to any vendor and keep the answers by email.
- What is included each month: calls, minutes or customers, and what happens above that?
- Is there a setup, menu-loading or integration fee, and is it one-off?
- Which POS does it connect to, and does the integration cover modifiers such as half-and-half or extra cheese?
- What happens to a call it cannot handle, and who pays for the transfer?
- Can I export call recordings and transcripts, and for how long are they kept?
- What is the notice period if I cancel on an annual plan?
- Does the system tell callers they are speaking to an AI? If you sell to customers in the EU, transparency duties for chatbots and voice systems may apply, so ask a qualified adviser what wording you need.
A sample reply is worth reading closely. One illustrative answer to question 1 was "unlimited calls on all plans". That sounds final, but the follow-up showed a fair-use clause that capped monthly minutes. The vendor, not the pricing page, held the real figure. Get the clause in writing.
How to tell in a month whether it is paying
Choose four numbers on day one and read them again after 30 days. Pull them from the phone provider's call report and your POS, not from the AI vendor's dashboard alone.
| Measure | Before (illustrative) | After 30 days (illustrative) | What it tells you |
|---|---|---|---|
| Calls unanswered at peak | 8 an evening | 1 an evening | Whether the rescue actually happens |
| Phone orders per month | 380 | 402 | Extra orders to count against the plan |
| Orders needing a fix after the call | Not measured | 14 of 402 | Accuracy, aiming under 5% |
| Calls transferred to staff | Not measured | 96 | Whether staff time is really saved |
On those figures the extra 22 orders at $17.60 profit each is $387, minus a $220 plan, leaves about $167, and 14 fix-ups at five minutes each is a bit over an hour of staff time. Notice that the accuracy line is the one most likely to catch you out: 14 of 402 is 3.5%, under the 5% target, but a rate of 10% would mean a refund and an angry customer on every tenth call.
The 90-day method for measuring an AI receptionist's return covers the longer version of this check, including what to do if the numbers fall short.
Ways to lower the bill without losing orders
- Route only the peak. Forward calls to the AI only during rush hours or when the line has been ringing for 20 seconds. On a flat plan this changes nothing; on a per-minute or per-customer plan it cuts the bill directly.
- Move simple questions to text. Put hours, address and the online ordering link on the website and in a text reply. Fewer calls means fewer minutes or customers billed.
- Trim the menu the AI must know. Cut rarely sold items and one-off specials from the voice menu. Fewer options mean shorter calls and fewer errors.
- Use the annual price after the test month. It saves about $130 a month on Maple Pro, which is $1,560 a year.
- Cap call length. On per-minute platforms, a maximum of about 8 minutes stops a stuck call from running up charges.
Picking a plan by call volume
| Your situation | Starting point | Reason |
|---|---|---|
| Under 150 calls a month, nobody misses them | No AI; fix hold music and staffing | Monthly cost exceeds recoverable profit. |
| 300 to 800 calls, peaks of missed calls | Flat restaurant plan with POS orders, annual only after test month | Predictable bill, fast set-up. |
| Many calls, mostly questions | An answering-only plan | Cheaper than one that takes orders. |
| Technical owner, unusual menu | Per-minute platform | Cheapest if you can build and maintain it. |
| Several shops | Quote from the vendor's volume tier | Group pricing is custom. |
If your problem is orders leaving through delivery apps, rather than missed calls, compare the numbers in the tutorial on AI ordering versus delivery apps. For the wider cost picture of restaurant phone AI, the AI phone answering cost for restaurants tutorial covers sit-down venues, and the pizza delivery software tutorial covers what happens once the order is in.
Four mistakes that inflate the bill
Paying for ordering when you need answering. A shop whose calls are mostly opening-hours questions buys the top plan and uses the bottom one. Count call types from a week of logs first.
Signing annually before testing. The annual discount is real, but a year of the wrong tool costs more than the discount saves.
Ignoring minutes on a per-minute build. One looping call can cost cents but a misconfigured agent that re-dials or waits on hold adds up. Set a maximum call length.
Counting every missed call as a lost order. Spam, wrong numbers and people who ring back are not lost sales. Use your 70% and 50% figures from real logs, not the ones in this tutorial.
If you want your own phone log turned into this sum with someone checking the assumptions, my AI implementation consultation covers that; the place to start is the contact page.
Cost questions before you buy phone AI
Is a cheaper per-minute build better than a flat plan?
Only if you can connect it to your point-of-sale system and maintain it. Per-minute platforms supply the voice and the model, not the menu logic or the till connection. A flat restaurant plan that already integrates with your POS usually costs more per month but far less in set-up time.
What happens to my bill if Friday is busier than usual?
On per-minute pricing it rises with every minute. On a flat plan it stays put unless the vendor caps usage, so ask in writing whether there is a call allowance, an overage rate or a fair-use limit. Goodcall-style plans bill by unique customers instead.
Should I pay annually to get the lower price?
Not for the first month or two. Pay monthly while you run test calls and watch real orders, then switch to annual billing once the call log shows the system takes orders correctly. Maple Pro's annual price is about 37% below its monthly one.
Do I still need staff on the phone?
Yes, for transfers. Any system should hand the call to a person for allergies, complaints, large catering orders and anything it cannot match to the menu. Cost the transfer routine into your plan; it is part of the system.
Further reads
- AI Phone Ordering for Takeaways: A Test Checklist for Menus and Postcodes — A test checklist for menus and delivery addresses on phone AI.
- How to Set Up an AI Phone Line for Takeaway Orders — Walks through setting up an AI line for takeaway orders.
- AI Restaurant Order Taking: Who Takes the Order, Who Texts a Link — Separates vendors that take the order from those that text a link.
- AI Startups for Restaurants: Protect Your Number, Data and Exit — Protects your number and data if the vendor fails.
- Is Yelp Host Worth It for a Small Restaurant? — Assesses one restaurant phone option on its own terms.
- AI Receptionist or Answering Service for a Booking-Based Business? — Compares AI with a human answering service on cost.
- Pizza Shop Marketing: Fill Slow Nights With AI-Written Offers — Fills the quiet nights the phone system cannot.
- Popmenu AI Answering: Does It Take Orders or Only Send Links? — Popmenu AI Answering sends links instead of taking orders. A test-call script, a link-drop-off sum and what to put in place before a takeaway relies on it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Maple, Goodcall, Retell and Vapi pricing pages, read in October 2026; the order-value, margin and call-volume figures in the worked example are illustrative.