Before signing up, ask twelve questions in writing: the full first-year cost including setup fees, the contract term and exit, usage charges, price rises, which POS and booking systems it writes to, how menu and allergen changes reach it, what happens at Saturday peak, who owns guest data and recordings, and how a guest reaches a person.
How a vendor answers matters as much as what it says. Good answers are specific, arrive in writing and match the contract. Below, each question comes with why it matters in a restaurant, how to check the answer and what a red flag sounds like. After the twelve you'll find an email to send, a scoring sheet and five test calls to make before you commit. For questions that apply to any AI vendor in any sector, the general vendor question list covers the basics; this one is for restaurants.
Money: questions 1 to 4
1. What will the first year cost in total, setup fees included?
Why: Setup fees vary from nothing to several hundred dollars, and they're often missing from the pricing page. Hardware (a tablet for the host stand, a printer), integration add-ons and training can add more.
How to check: Ask for a written quote with every line: monthly price times twelve, setup, hardware, integrations, training, and anything charged per location.
Run the numbers: here is how an itemised quote for an illustrative 60-cover bistro compares with the headline it was sold on.
Headline on the pricing page: $299 a month
Itemised first-year quote:
Subscription, 12 x $299 $3,588
Onboarding and menu setup $350 (one-off)
Booking-system connector,
12 x $49 add-on $588
Host-stand tablet $0 (uses our own)
Staff training session $0 (included)
------
First-year total $4,526The real first year is about 26% above "$299 a month", and the $49 connector is the line that makes the tool do what the bistro bought it for. Without the itemised quote, it would have surfaced on the second invoice.
Red flag: "We'll sort out setup once you're on board."
2. How long is the contract, and how do I leave?
Why: Annual commitments are common, and yearly billing is usually the cheaper rate, so vendors push it. Automatic renewal with a long notice period can trap you for a second year.
How to check: Find the term, renewal and cancellation clauses in the contract itself, not the sales deck. Note the notice period in your calendar the day you sign.
Red flag: Auto-renewal with a 60- or 90-day notice window you only discover in the small print.
How this goes wrong: an illustrative pizzeria signed a yearly plan at $200 a month. The contract said: "This agreement renews automatically for further 12-month terms unless either party gives written notice at least 60 days before the end of the current term." The owner decided in month eleven that the tool wasn't paying its way and emailed to cancel, 30 days before the anniversary. The notice was too late, and the pizzeria was committed to a second year, $2,400 for a tool nobody wanted. A reminder set for month nine on the day of signing, 90 days before renewal, would have left time to decide, negotiate or leave.
3. What is charged per use?
Why: Per-minute, per-caller, per-order and per-text charges scale with your busiest weeks. Goodcall, for example, includes a set number of unique callers a month and charges $0.50 for each one beyond that; in a restaurant, every person ringing to ask about a table counts.
How to check: Give the vendor your call or order volume for your busiest month last year and ask what that month would have cost.
Red flag: "Unlimited" with no written definition of fair use.
Run the numbers: take Goodcall's published structure, where the $129 plan includes 250 unique callers a month and each extra caller costs $0.50. An illustrative 80-cover restaurant has about 320 different callers in a quiet February and 900 in December, with party bookings. February costs $129 plus 70 extra callers at $0.50, so $164. December costs $129 plus 650 extra callers, so $454. The plan looks like $129 a month on the pricing page; across that year it might average closer to $250. That's not a reason to rule it out, only a reason to price your busiest month before you sign.
4. What happens to the price after the first year or the promotion?
Why: An introductory rate that doubles at renewal changes the whole case.
How to check: Ask for the renewal price in writing and whether the contract caps annual increases.
Red flag: "We rarely raise prices" with nothing in the contract.
A good answer vs a weak one: say a vendor offers an introductory $149 a month for the first year, with a list price of $249. The weak answer is "renewal pricing will be discussed nearer the time", which means year two could cost $2,988 instead of the $1,788 the owner has budgeted, a $1,200 surprise. A good answer, written into the order form, reads something like "Year two at $179 a month; increases thereafter capped at 5% a year." If the vendor won't cap it, budget year two at full list price and decide whether the tool still pays on that figure.
Fit with your service: questions 5 to 8
5. Which POS and reservation systems does it write to, not just read?
Why: A tool that can see your diary but can't put a booking in it turns every call into a message someone has to action. Vendors usually name their integrations: Maple lists Toast, Square, Clover and SkyTab for orders plus OpenTable for reservations on its Pro plan; Slang.ai lists OpenTable, SevenRooms and Yelp among others.
How to check: Ask for a live demo making a booking or order into your actual system, or a reference customer using the same POS and booking platform.
Red flag: "We integrate with that" when bookings actually arrive as emails.
How this shows up in a demo: an illustrative 70-cover brasserie asked the vendor to book a table for four at 8pm through the AI, live, while the manager watched the reservation system. Nothing appeared. Two minutes later an email arrived in the restaurant inbox headed "New reservation request", which someone would have to type into the book by hand. That is "reading", not "writing". A genuine write-back looks different: the table appears in the book with the party size, time, phone number and any note ("highchair"), and the 8pm slot's available covers drop by four, so the next caller can't be offered the same table.
6. How do menu changes, sold-out dishes and allergens reach it?
Why: An AI that offers last week's special or a dish the kitchen has run out of does more harm than a missed call. Allergen questions carry real risk and should reach a person who checks with the kitchen.
How to check: Ask them to show you changing a dish in your POS and whether the AI picks it up automatically or needs a manual edit. Ask exactly what it says when a caller mentions an allergy; how AI should handle allergen questions sets out the safe answer.
Red flag: "It knows your menu" without explaining how it stays current, or any willingness to answer allergen safety questions itself.
Test it like this: at 4.10pm, mark the crab linguine as sold out in the POS. At 4.15pm, ring the AI and ask whether it's on tonight. If it still offers the dish, ask the vendor when menus sync; an answer of "overnight" means every sell-out during service is invisible to the AI until tomorrow. Then ask the allergy question and compare the transcript with these two versions:
Unsafe: "Our chocolate torte doesn't contain nuts, so it
should be fine for your daughter."
Safe: "I can't confirm allergens over the phone. Can I
take your number? The manager on duty will call
you back after checking with the kitchen."The unsafe reply may even be accurate about the recipe, and it's still the wrong answer, because it can't know about shared equipment or a substitution made that morning.
7. What happens at 8pm on a Saturday?
Why: Your busiest hour is when the tool matters most and when systems are most stretched. Some vendors state how many calls they handle at once; Loman's pricing page, for instance, lists up to 50 concurrent calls on its entry plan.
How to check: Ask how many simultaneous calls or chats your plan handles, what happened during their last outage, and where calls go if their system is down.
Red flag: No fallback. If their system fails, your phone should ring in the restaurant, not go silent.
A good answer vs a weak one: "We have 99.9% uptime" tells you nothing about your Saturday. A useful answer (illustrative) sounds like: "Our last incident was a 38-minute outage at our voice provider. Calls fell back automatically to ringing your restaurant line, and we emailed affected customers that evening with a list of missed callers." During the trial, test the busy-hour claim yourself: have three people ring the line within the same minute at 7.45pm on a Saturday and note whether all three get through, what each hears, and whether any booking lands twice.
8. Can I see and correct what it tells guests?
Why: You'll only catch wrong answers if you can read them. Call recordings or transcripts and summaries show what was actually said.
How to check: Ask to see the dashboard, find one conversation, and change an answer. Time how long the change takes to go live.
Red flag: Changes that need a support ticket and a few days' wait.
What a good fix looks like: reading transcripts in the first week, the owner of an illustrative family restaurant finds this exchange:
Caller: Do you do a children's menu on Sundays?
AI: Yes, our children's menu is available every day.In fact the children's menu runs only at weekend lunch. The owner edits the menu entry in the dashboard at 11.02am, rings the line at 11.07am, asks the same question and gets the corrected answer. Five minutes from finding the error to the fix being live scores a 2 on this question. If the same change had needed an email to support and a two-day wait, every Sunday-evening caller in between would have been told the wrong thing.
Guest data and your control: questions 9 to 12
9. Who owns the guest data, and can I export it?
Why: Names, numbers, booking history and preferences are valuable to you, and you'll want them if you switch vendors.
How to check: Read the data clauses and ask for a sample export. Ask what is deleted, and when, after you cancel. What to check in an AI tool's privacy terms lists the clauses to find.
Red flag: Export only as PDF reports, or "you can request your data" with no format or timescale.
What to look for in the sample: open the export in a spreadsheet and check the columns. A useful one (illustrative) has guest name, phone, email, number of visits, last visit date, notes such as "window table, anniversary in June", and a marketing-consent column. Picture a wine bar whose sample had names and phone numbers but no consent column: if it ever switched vendors, it couldn't tell which of its 2,100 guests had agreed to hear about events. The fix is to ask for the field to be added to the export before signing, and to get that promise in writing.
10. Are calls recorded, who handles consent, and for how long?
Why: Rules on recording calls and telling callers vary depending on where you and your callers are. You're the one guests will hold responsible.
How to check: Ask whether the system announces recording, how long recordings are kept, whether they're used to train the vendor's AI, and whether a data processing agreement is available. If you're unsure what your obligations are, ask your data-protection adviser.
Red flag: "We use recordings to improve our AI" with no way to opt out.
Example wording to ask for: a greeting that covers both points in one breath, such as "Thanks for calling. You're speaking to our automated assistant, and calls are recorded so we can check bookings." Then pin down the rest in the vendor's reply. A clear answer (illustrative) reads: "Recordings and transcripts are kept for 90 days, then deleted. They are not used to train models shared with other customers. A data processing agreement is attached." A vague one reads "we follow industry best practice", which answers none of the three questions.
11. How does a guest reach a person, and how quickly?
Why: Regulars, complaints, large parties and anything unusual should reach staff. A guest stuck in a loop with an AI is a guest you lose.
How to check: Ask what triggers a transfer, what the caller hears, and what happens if nobody picks up the transfer during service.
Red flag: No clear route to a person, or transfers that drop the call when staff are busy.
Two trial results side by side: an illustrative steakhouse trialled two vendors and asked each for the manager at 7.52pm on a Friday. Vendor A said "Let me put you through", rang the restaurant for 30 seconds, and when nobody answered said "The manager is with guests. Can I take your number? I'll text them now." The manager got a text with the caller's name, number and reason, and rang back at 8.10pm. Vendor B rang the restaurant, nobody picked up, and the call ended. Same question, same restaurant, same busy moment: one guest was looked after and the other was lost.
12. Who fixes it on a Friday night?
Why: Restaurants run evenings and weekends. Support that works office hours only isn't support for you.
How to check: Ask for support hours, the response time in writing, and whether you get a named contact.
Red flag: Email-only support with a two-day response, for a tool that answers your phone at night.
How this goes wrong: at 6pm on a Friday, after a vendor update, an illustrative seafood restaurant's AI starts telling callers the kitchen closes at 9pm instead of 10pm. The support ticket is acknowledged automatically and answered on Monday morning, after a weekend of callers told there was no late table. So ask a sharper version of the question: "If your AI starts giving wrong opening hours at 6pm on a Friday, what exactly do I do, and when will a person respond?" And know your own off switch: turning call forwarding off at your phone provider puts every call back on the restaurant line in minutes, whatever the vendor's support hours are.
Send the questions as one email
One written email makes vendors answer in writing, and lets you compare two or three vendors line by line.
Subject: Questions before we sign up - [restaurant name]
We're a [covers]-cover restaurant open [days/hours], using
[POS] and [booking system]. Busiest month: about [number] calls
and [number] takeaway orders.
1. Total first-year cost, itemised, including setup/onboarding,
hardware, integrations and training?
2. Contract length, renewal terms and cancellation notice?
3. Any per-minute, per-caller, per-order or per-text charges?
What would our busiest month have cost?
4. Renewal price after year one or any promotion? Any cap?
5. Can you create bookings/orders directly in [POS] and
[booking system]? Can we see a live demo in our set-up?
6. How do menu changes and sold-out items reach the AI?
What does it say when a caller mentions an allergy?
7. How many simultaneous calls on our plan? Where do calls go
if your system is down?
8. Can we read transcripts and change answers ourselves?
How fast do changes go live?
9. Who owns guest data? Export format? Deletion after we leave?
10. Are calls recorded? Retention period? Used for training?
Data processing agreement available?
11. How does a caller reach a person? What if nobody answers?
12. Support hours, response times, named contact?Score the replies
Give each answer 0, 1 or 2, then apply the stop rules.
| Score | What it means |
|---|---|
| 2 | Specific, in writing, and consistent with the contract |
| 1 | Given verbally, vague, or "we can arrange that" |
| 0 | No answer, a deflection, or one of the red flags above |
- Stop rule: a 0 on question 1 (cost), 5 (write-back), 6 (menu and allergens), 9 (guest data) or 11 (reaching a person) means don't sign until it's fixed in writing.
- 18 to 24: worth a paid trial on monthly billing.
- 12 to 17: negotiate the weak answers, or trial a competitor alongside.
- Under 12: walk away.
Here is how the scoring looked for an illustrative 90-cover restaurant comparing two phone-answering vendors:
| Question | Vendor A | Vendor B | Note |
|---|---|---|---|
| 1. First-year cost | 2 | 1 | B gave a monthly price but "setup depends" |
| 2. Contract and exit | 2 | 2 | Both monthly after a 3-month minimum |
| 3. Per-use charges | 2 | 1 | B's "fair use" limit not defined |
| 4. Renewal price | 1 | 2 | A: "discussed at renewal" |
| 5. POS and booking write-back | 2 | 0 | B's demo booking arrived as an email |
| 6. Menu, sold-out, allergens | 2 | 1 | B's menu syncs overnight |
| 7. Saturday peak and fallback | 1 | 2 | A's fallback is manual: staff switch off forwarding |
| 8. See and correct answers | 2 | 2 | Both under ten minutes |
| 9. Guest data and export | 2 | 1 | B exports without consent flags |
| 10. Recording and retention | 2 | 2 | Both 90 days, no shared training |
| 11. Reaching a person | 2 | 1 | B drops the call if nobody answers |
| 12. Out-of-hours support | 1 | 2 | A: chat support until 10pm only |
| Total | 21 | 17 |
On totals alone the two are four points apart, but the stop rule settles it: Vendor B's 0 on write-back means no signature until its bookings land in the reservation system. Vendor A goes to a monthly trial, with two things to settle in writing first: a renewal price, and exactly who switches forwarding off, and how, if its system fails on a Saturday.
If a vendor's cost answer seems high or low, compare it with published restaurant AI phone prices before you negotiate.
Five test calls to make during the trial
Call from a mobile, at a busy time, and write down what happens. These five cover the failures that cause most complaints, the ones described in the AI mistakes restaurant owners make with bookings and reviews.
- A complicated booking. A table for seven at 7.30pm on Saturday, with a highchair and one guest in a wheelchair. Check what lands in your booking system.
- A dish you removed yesterday. Ask whether it's on tonight. It should say no.
- An allergy. "My daughter has a nut allergy. Is the dessert safe?" It should pass you to a person or take details for a call-back, not reassure you.
- A change under a slightly different name. Cancel or move a booking giving the surname spelled differently. See whether it finds the right booking or creates a mess.
- Background noise and a request for a person. Call from somewhere loud, speak quickly, and ask for the manager. Time how long it takes to reach someone.
Keep the results in a short log, because the vendor conversation afterwards goes better with specifics. An illustrative 45-cover noodle and curry restaurant's log after one Friday evening:
Trial test calls - Fri 19:10 to 19:40, owner's mobile
1. Party of 7, Sat 19:30, highchair, wheelchair
PARTLY: booking in the system, highchair in notes,
wheelchair not recorded. Ask: can "access needs" be
a required field?
2. Massaman lamb (removed Thursday)
FAIL: offered it. Menu syncs overnight; ours was
changed after the sync. Ask: manual "86" button?
3. Nut allergy, dessert
PASS: took number, manager texted within a minute.
4. Move booking, surname given with two letters
different from the booking
FAIL: made a second booking; two tables held.
Ask: can it match on phone number instead?
5. Loud street, asked for manager
PASS: rang through in about 40 seconds.Two fails out of five isn't a reason to walk away. It is two written questions for the vendor, a retest of both after the fix, and a trial extended until they pass, with the double booking treated as the more serious of the two because it costs a table on a busy night.
When a test booking goes wrong, note exactly what was said and how it was recorded. What to do when an AI receptionist gets a booking wrong shows how to trace and fix the cause rather than just the booking.
Further reads
- How to Evaluate an AI Implementation Proposal or Quote — Judge a quote or proposal once the answers are in.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep your guest data and set-up portable if you switch.
- How Much Does AI Cost a Small Restaurant Each Month? — See where a new tool fits in a restaurant's monthly AI budget.
- How to Pilot AI in Shadow Mode Before Customers See It — Run the tool behind the scenes before guests hear it.
- How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — Decide keep, fix or cancel once the trial is over.
- Your First 30 Days of AI in a Restaurant, Week by Week — What to do week by week after you sign.
- What an AI Implementation Plan Looks Like for a Small Restaurant — One illustrative 48-cover restaurant's AI plan, start to finish: a week of tracking, three jobs chosen, four ruled out, and the day-90 numbers.
- AI Ordering vs Delivery Apps: What a Takeaway Actually Pays — Delivery app commission against AI phone and web ordering, line by line, with a break-even worksheet a takeaway can fill in with its own numbers.
- Where AI Actually Saves Time in a Small Restaurant — A task-by-task look at a small restaurant's admin week: where AI cuts real minutes, where it only moves them, and where it adds work.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Maple, Slang.ai, Goodcall and Loman product and pricing pages, checked September 2026. Vendor details are examples of what to ask about, not endorsements.