Key takeaways
- Phone orders are direct orders, which matters when many operators say third-party delivery orders earn them no profit.
- Decide the level first: text an ordering link, capture the order for staff, or send a ticket straight to the POS.
- Most errors come from modifiers, nicknames and items that are out of stock, so prepare the menu the way guests say it.
- Audit a sample of orders every day for the first month, comparing the call, the ticket and what left the kitchen.
Why phone orders deserve attention
Takeout is no longer a side business. Mobile ordering is mainstream, with 57% of adults having recently used it[1], and 51% of Gen Z and millennial adults say picking up takeout or using a drive-thru is essential to their lifestyle[2]. Yet the old-fashioned phone order still matters for a simple reason: it usually comes to you directly, without a marketplace in between.
The economics of third-party orders are tough. In a May 2026 survey cited by the National Restaurant Association, 47% of full-service and 43% of limited-service operators said they do not make a profit on third-party delivery orders, and two in three restaurants said their average platform fee falls between 15% and 29.9% of the transaction[3]. Every takeout call you answer and get right is an order without that fee.
Source: National Restaurant Association, 2026[3]
Show the data
| Item | Value |
|---|---|
| Full-service | 47% |
| Limited-service | 43% |
Three levels of AI phone ordering
| Level | How it works | Accuracy risk | What to check first |
|---|---|---|---|
| Text the ordering link | The agent offers to text your official online ordering link and the guest orders themselves | Low: the guest builds the order on your menu | Correct link, mobile-friendly page, consent to text |
| Capture for staff | The agent takes the order by voice and sends a written summary for staff to enter | Medium: errors in hearing and re-entry | Readback of every item and modifier; clear summary format |
| POS ticket | The agent builds the order against your menu data and sends a ticket to the POS or kitchen display | Highest: mistakes go straight to the line | Menu mapping, modifiers, 86 list, prices and taxes |
Many restaurants start with the link because it is safe and quick to launch, then move up a level once they trust the system. Ava, for example, offers the link route today: when a caller wants pickup or delivery and agrees to a text, it sends your official ordering link. Whichever level you pick, the checks below apply to the parts you use.
Prepare the menu the way guests say it
Your POS menu was written for servers and cooks, not for callers. Guests say “the chicken sandwich with no mayo”, not “CHX SAND - MOD: NO MAYO”. Before go-live, someone who knows the menu should build a spoken version.
- List every item's common nicknames: “the big burger”, “the spicy one”, “the usual kids' meal”.
- Mark required choices: size, temperature, side, sauce, bread. The agent must ask for each one.
- Mark optional modifiers and which ones carry an upcharge.
- List items that cannot be modified, such as sauces made in batches.
- Note items sold by weight or count, such as wings by the dozen.
- Connect the 86 list, or decide who updates the agent when an item runs out during service.
Modifiers are where orders break
In most audits, the item is right and the modifier is wrong. A guest asks for a burger medium-rare with no onion and extra pickles, and the ticket shows a burger with extra onion. Test the combinations that matter most to your kitchen and your guests.
- Removals: “no onion”, “hold the cheese”, “without the bun”.
- Substitutions: “sweet potato fries instead”, “gluten-free bread”.
- Additions with charges: “add bacon”, “extra avocado”.
- Split instructions: “half the wings buffalo, half garlic parmesan”.
- Temperature and doneness for proteins.
- Order-level notes: “cutlery for four”, “please don't ring the bell”.
Require a full readback of every item and modifier before the order is placed, and require the guest to confirm it. A readback adds seconds and saves remakes.
Pickup times and payment
Quote pickup times from the kitchen's real load, not a fixed number. If the agent always says “about 20 minutes”, Friday guests arrive to cold bags or long waits. Ask the vendor how quotes are calculated and whether the kitchen can adjust them during service.
Avoid having guests read card numbers aloud to any system unless it is certified for card handling. Safer options are paying at pickup or receiving a secure payment link by text. Check that whatever method you use matches how your POS closes the ticket, so the order is not left open or charged twice.
Allergies and special instructions
US law recognizes nine major food allergens, with sesame added as the ninth from January 1, 2023[4]. A takeout guest who mentions any of them should not get a yes or no from an automated system. Set the rule: the agent records the allergy on the order, flags it clearly on the ticket, and hands the guest to a manager or kitchen lead for confirmation before the order is fired.
Questions for the vendor about ordering
- Which POS systems can you send tickets to today, and which are planned?
- How do you import our menu, and who maps modifiers and upcharges?
- What happens when we change prices or add a special: is it automatic or manual?
- How does the agent learn an item is out of stock during service?
- Where do orders go if the POS connection fails mid-shift?
- Can we see transcripts next to the tickets they produced, for audits?
- How are allergy notes shown on the ticket?
Ask for a live test with your own menu loaded, not a demo restaurant. Order the five most modified items on your menu and check the tickets.
Suggestions and upsells, used carefully
An AI agent can suggest a drink, a side or a dessert on every call, which is more consistent than most busy staff. Used lightly it raises check size; used on every call it makes the order longer and more irritating. Pick one natural suggestion per order, tied to what the guest already chose, such as fries with a burger or a second sauce with wings. Skip suggestions entirely when the guest is in a hurry or the order is large and complicated.
Never let a suggestion change what the guest asked for. If the guest says no, the agent moves on without asking again.
The first-month accuracy audit
Pick a daily sample
Choose a fixed number of AI-taken orders each day, spread across lunch, dinner and late night.
Compare three things
The call transcript, the ticket that reached the kitchen, and what was actually packed. Note every difference.
Classify each error
Wrong item, wrong modifier, missing item, wrong quantity, wrong pickup time, wrong price, or a note that was lost.
Find the cause
Mishearing, missing nickname, bad menu mapping, out-of-stock item, or a kitchen mistake that would have happened anyway.
Fix and retest
Update the spoken menu, modifiers or 86 process, then call in test orders for the items that failed.
Watch complaints
Track remakes, refunds and complaints about phone orders against the weeks before launch.
Example: a fictional pizza shop sets its own target that no more than 2 in 100 audited orders have an error caused by the phone system, and keeps auditing daily until it holds that level for two straight weeks, then moves to a weekly sample.
Guests who would rather tap than talk
Not every takeout caller wants to order by voice. A majority of full-service customers say they are comfortable placing an order with a smartphone app (63%)[5]. Offer both: let the agent take the order for those who want to talk, and offer the link to those who would rather see the menu with photos and prices.
Example: a two-week launch
Example: a fictional noodle bar starts with capture-for-staff mode. In week one the manager audits every phone order. Most errors involve spice levels, because guests say “mild”, “medium” and “Thai hot” while the POS uses numbers. The manager adds those words to the spoken menu and the errors stop.
Example: in week two the audit finds two orders for a sold-out dumpling. The team adds a step to the pre-shift checklist to update the agent's 86 list. After a clean month, the owner moves to POS tickets for lunch only. Before expanding, the team reruns the matrix in how to test an AI phone agent before going live and reviews the handoff rules for staff.
Frequently asked questions
Can AI take takeout orders over the phone accurately?
It can, if the menu is prepared the way guests speak, every modifier is read back, and you audit a daily sample of orders in the first month.
Should AI phone orders go straight into the POS?
Only after the menu mapping, modifiers and out-of-stock process are tested. Many restaurants start by texting an ordering link or capturing orders for staff to enter.
How should an AI agent handle allergies on takeout orders?
Record the allergy on the order, flag it on the ticket and hand the guest to a manager or kitchen lead to confirm before the order is made.
Is it safe to take card payments by voice?
Avoid it unless the system is certified for card handling. Pay at pickup or a secure payment link by text are safer options.
References
National Restaurant Association (2025). From Trend to Transformation: Off-Premises Dining Now Essential for Restaurant Consumers, Operators (opens in a new tab) ↑
National Restaurant Association (2025). Report: Takeout, drive-thru, delivery are more popular than ever (opens in a new tab) ↑
National Restaurant Association (2026). National Restaurant Association comments to the FTC on the Food Delivery Fees ANPRM (Project No. P267101) (opens in a new tab) ↑
US Food and Drug Administration (2023). Food Allergies (opens in a new tab) ↑
National Restaurant Association (2024). Restaurant Technology Landscape Report 2024 (opens in a new tab) ↑