AI Consulting for Restaurants: What to Ask and What to Expect

Ask a restaurant AI consultant which one job they will fix first, what the fee covers, which software costs sit outside it, where your sales and supplier data go, and who runs it afterwards. In month one, expect a written scope, one tested process, a data note and a hand-over guide, not a transformed restaurant.

Consulting pages aimed at restaurants lean on big promises: lower food cost, smarter menus, full automation. A first engagement can deliver one of those, partly, if it is picked well. The questions below are designed to make a consultant name that one thing, price it, and show how you will test it.

Follow me on Instagram@sagnikteaches

Start from a sample scope of work

Instead of a plan day by day, a restaurant is better served by a scope of work: the paper that says what is in, what is out and what you will hold at the end. Here is an illustrative one for a 40-seat bistro, with invented figures, so you can see the level of detail to ask for.

Connect on LinkedInSagnik Bhattacharya

SCOPE OF WORK (illustrative)
Client: 40-seat bistro, one site, owner-run, head chef plus 9 staff
Process in scope: weekly supplier-invoice price check
Why this one: weekly food spend about $4,200; owner suspects slow price
  creep but checks invoices by eye on Sunday nights
Out of scope: menu redesign, website chatbot, phone answering, rota
Inputs: last 12 weeks of supplier invoices (photos or PDFs), the ordering
  spreadsheet, the POS weekly sales export
Deliverables:
  1. Ranked list of 8 candidate jobs with hours a week and risk rating
  2. A spreadsheet that logs unit price per item per week and flags any
     item up more than 5 percent against its 12-week average
  3. A prompt that reads an invoice and fills the log; tested on 12 past
     invoices, errors recorded
  4. Data note: what may go into the AI account (invoices, yes; staff
     records and guest details, no)
  5. One-page guide and a 20-minute recorded walkthrough
  6. Day-60 review call
Fee: $2,400 fixed, two payments (after deliverable 3, after 5 and 6)
Tools, billed to the restaurant directly: business-grade assistant seat,
  list price checked on vendor page
Change control: work outside this scope needs written approval and a price

Notice how readable this is for a busy owner. A consultant who sends you something this concrete after a short chat is probably the kind who will deliver. A consultant who sends a glossy proposal with no inputs list has probably not asked how your Sunday nights actually go.

Subscribe on YouTube@sagnikteaches

The numbers behind that scope

Check whether the project can pay for itself before signing. Using the invented bistro figures: weekly food spend of $4,200 is about $218,400 a year. If unnoticed price creep adds 3 percent across supplier lines, that is $126 a week, about $6,550 a year. Catching half of it, because a flag prompts the owner to challenge a supplier or switch, is about $3,275 a year. Against a $2,400 fee and roughly $240 a year for one assistant seat at the annual business rate of $20 a month, the first year nets about $635, and later years about $3,000. The owner also stops spending a few hours every Sunday on invoice checking, which is a benefit that is real but not in the sum.

If the sum shows a loss in the first year, it is not necessarily a reason to refuse: perhaps the real value is a habit of checking. But it should be your decision, made with the numbers on the table, not something discovered in month six. The monthly cost of AI for a small restaurant gives running costs to place beside the fee.

Questions to ask, grouped by restaurant job

Consultants speak about "use cases". Owners think in jobs. Ask the consultant to map their proposal onto the jobs below and tell you which they would leave alone.

Purchasing and cost control

Ask: "How would you read supplier invoices, and what happens when a photo is blurry or a unit changes from case to kilo?" The honest answer includes human checking of the first few weeks and a rule for failures. Reading invoices is a promising job; the failure mode is a plausible but wrong price being logged. Insist on a test against twelve or more real invoices.

Menu and sales analysis

Ask: "What will you do with our POS export, and what will you not claim?" AI can summarise which items sell by day and hour, and flag slow lines. It cannot tell you why a dish sells badly without your knowledge of the kitchen. A sensible consultant is cautious about recommending price changes from a few weeks of data.

Guest communication

Ask: "Who reviews guest-facing text before it is sent, and what is the rule for allergens?" The rule should be that allergen information comes only from the signed-off matrix and that an AI-drafted answer never goes out unreviewed. Read whether a chatbot should answer allergen questions before the consultant proposes one.

Staff and rotas

Ask: "What data do you need about staff, and where is it stored?" Rota drafting from sales forecasts can help, but staff names, pay rates and availability are personal data. Keep names out of the first test by using initials or roles, and read how to build a rota from AI demand forecasts for the second stage.

Phones and bookings

Ask: "How is this priced per call, and what happens at peak?" Per-call pricing makes the sum easy to check. As an illustration from one vendor's published price, Smith.ai's AI receptionist is free for 25 calls, and its paid Pro plan starts at $150 a month with per-call rates of about $2 at the 75-call bundle (check the page for the current bundles); a restaurant taking 400 calls a month would price that against a staff member's time. The phone question has its own page: AI phone answering for restaurants: cost and payback.

What stays out of the consultant's fee

The fee pays for judgement, set-up and hand-over. These items should be listed separately, in the restaurant's name:

• Assistant seats. List prices shift, so check the vendor page the week you buy. As a reference point, ChatGPT Business Standard is listed at $25 a user a month, or $20 a user billed annually; Claude Team Standard has the same list price.
• Phone or booking tools. Priced per call, per cover or per seat depending on the vendor.
• Your POS vendor's own AI features. Ask your POS provider what is included in your current plan before paying anyone to replace it.
• Staff time. About 8 to 10 hours of owner or chef time in month one for testing is a realistic estimate; write it down so nobody is surprised.

The principle is simple: when a monthly tool renews, it should renew in your name, with your card, so that you can cancel it without asking the consultant.

What month one should leave on your desk

If the failure log is empty, be suspicious. A process tested on twelve invoices will always fail on some, and a consultant who reports none either did not test or did not look.

A sample exchange: weak answer against strong

This is invented, but you will hear versions of it. The question is "How will we know it's working?"

Weak answer: "You'll see the difference quickly. Our clients usually love the results." Nothing to measure, no date, and "clients usually" is unprovable.

Strong answer: "On day 60 we compare three numbers from your own records: minutes spent checking invoices on Sundays, the number of items flagged and challenged, and any price corrections you recovered. If Sunday time has not fallen by a third, we look at why."

One realistic mistake: the plausible wrong number

Imagine the invoice reader logs a delivery of olive oil at $64 a case when the invoice said $64 for a pack of two cases, so $32 each. Nothing looks absurd, and the price-increase flag silently stops working for that supplier. Nobody notices for five weeks, until the owner spots a figure that cannot be right. The fix is a unit field the prompt must fill, a rule that any line without a unit is flagged for a person, and a weekly spot check of three lines. This is the kind of detail a good consultant volunteers when you ask what could go wrong.

A second mistake: nothing written down about the data

A manager pastes a staff availability sheet, with names, phone numbers and notes about someone's health, into a free consumer chat tool to draft a rota. There is no policy, so there was no rule broken, and also no protection. The data note exists to close this gap in plain words: "Rota drafting uses initials only. Phone numbers, health notes and pay rates never go in. Use the restaurant account." If you serve guests or employ staff in the EU, data-protection law applies to this kind of data, and a qualified adviser should look at the final arrangements.

Choosing the first job: a quick scoring round

Ask the consultant to score candidate jobs with you rather than for you. A simple scheme works: rate each job from 1 to 5 on hours saved, on how reusable the data is, and on how bad a mistake would be (5 means a mistake is harmless), then add the three. Here is an illustrative round for the bistro.

The sales summary wins on this scoring, even though the invoice check has the clearer dollar prize. That is a legitimate disagreement worth having with the consultant: a safe, quick win in week one builds the team's trust, whereas the higher-value job may suit week five. Allergen answers score lowest because a mistake there is the least harmless of all, which is consistent with the rule that they stay with a person. What matters is that the scoring is visible and that you can overrule it.

What the owner actually does in the first week

People assume a consultant works while the owner carries on. In practice, the first week asks for about three hours of you: a 45-minute conversation about how invoices, orders and reviews are handled today; half an hour to export the files; and an hour of reading the ranked list and picking the job. If you cannot find those three hours, say so at the start. The hours are the cheapest part of the project and the one most often missing.

The consultant should also ask to sit in on one ordinary Sunday-night routine, or whichever routine you have chosen. Watching the real job turns up small habits that a questionnaire never would: the supplier who writes prices in the margin, the delivery notes kept in a different folder, the chef who corrects the order sheet by hand. Good scopes are built from those details.

Reading the day-60 review properly

The review call is where you decide whether to keep going, so prepare for it. Gather the three numbers before the call: Sunday-night minutes this month against the earlier count, the number of price flags raised and how many led to action, and the number of times the process failed and was caught. Then ask four questions. Who has actually used it, and who has quietly stopped? What did the failure log add since hand-over? What would you stop doing if you had to cut it back to the minimum? And what is the next job, with its own scoring and its own fee?

An honest outcome might be "keep the invoice check, drop the review replies, and pause everything else for a quarter". That is a successful engagement. The aim of the first project is to learn what works in your restaurant at a cost you chose in advance, and the second project, if there is one, should be cheaper to scope because you now know how your own team behaves.

Sample test for the invoice prompt

When the consultant demonstrates deliverable 3, ask to see one invoice go in and the log row come out. The shape below is what you are looking for; the figures are invented.

Input (from a supplier invoice, one line):
  "Olive oil EV 5L x2   $64.00   line total $128.00"

Expected log row:
  Date | Supplier | Item        | Pack  | Units | Unit price | Flag
  3 Mar| Supplier A | Olive oil EV| 5L    | 2     | $64.00     | up 6.7% vs average, CHECK

If the pack size or unit is missing from the invoice, the row must read
"UNIT UNCLEAR, CHECK BY HAND" and not guess.

The last rule is the important one. The point of testing on twelve invoices is to see how often the prompt guesses instead of asking for help; a good build has a clear refusal rule and a count of how many times it fired.

Fee models, briefly, for a restaurant

• Fixed fee per process is the right start: you know the maximum, and the deliverables are checkable.
• Day rate with a cap suits an uncertain audit, for instance when you do not know which of your tools already has AI features.
• Retainer comes later, for tuning and a second job.

For context, Contra's hiring page for automation freelancers shows $500 to $1,500 for simple automations and $2,000 to $10,000 or more for complex ones, and retainers of $1,000 to $5,000 a month. Clutch's AI pricing guide (September 2026) shows typical reviewed projects between $10,000 and $49,999, and its directory lists firm minimums of $10,000 and up. A $2,400 first process is therefore at the low end for an independent and well below a firm's usual minimum; that is not a problem if the scope is small and honest. How the different suppliers compare is covered in consultant, agency or freelancer.

Review replies: a before and after

If the consultant's first job touches online reviews, ask to see the tone work. Here is an illustrative before and after for a three-star review about a slow Saturday service.

Before (typical unprompted draft): "Thank you for your feedback! We are sorry to hear you did not have a five-star experience. We strive for excellence and hope to see you again soon!" It is generic, it dodges the complaint, and regulars can spot it.

After (with your facts and a house-voice sample in the prompt): "Thank you for telling us about Saturday. Forty minutes between starters and mains is longer than we aim for, and we were a chef short that night. We have changed how we stagger the early tables. If you come back, ask for [your name] and we will look after you." The second reply states a fact the owner supplied, makes one promise the restaurant can keep, and avoids inventing anything. The rule that goes with it: the owner reads and sends every reply, and never lets the AI invent a cause. The tutorial on answering restaurant complaints with AI goes deeper.

Signals that a consultant understands kitchens

• They ask when your busiest service is, and plan testing around it, not through it.
• They ask who is allowed to change a recipe card or a price in the system.
• They are cautious about food-safety information and refuse to automate it without a named reviewer.
• They propose measuring in minutes and dollars you already track, not new dashboards.
• They have a view on what to leave manual, and say it unprompted.

An owner who prefers a wider map before choosing a consultant can read what an AI implementation plan looks like for a small restaurant; it describes the longer road that month one is the first step of. If you would rather work through your own scope with someone, that is what my AI implementation consultation covers, through the contact page.

Owner questions about hiring AI help for a restaurant

Should an AI consultant touch our allergen information?

Only to organise it, never to answer guests unchecked. The allergen matrix should stay owned and signed off by a named person in the kitchen. A consultant can help you format it and keep it current, but any guest-facing AI answer about allergens needs a human review route or a plain refusal to answer.

Can a consultant work from our POS exports alone?

Usually yes, and it is the safest start. A weekly sales export as a spreadsheet gives them item sales, times and totals without a live connection to your till. Ask for read-only access later, only if the manual version proves its value and your POS terms allow it.

Is a retainer sensible in the first month?

Rarely. A fixed fee for one scoped process gives you a clear test of whether the consultant is worth keeping. Consider a small retainer only after the first process is running and you have a second job ranked, such as ordering or rota drafting.

What if my team is too busy to test anything?

Then shrink the scope, not the testing. Choose a job the owner or head chef does anyway, such as invoice checking, and ask the consultant to build around it. A process nobody has time to review is the commonest reason these projects fade.

Further reads

• Restaurant AI Tools: 12 Questions to Ask Before You Sign Up — Questions for software vendors, as opposed to consultants.
• How to Forecast Covers and Cut Food Waste With AI — A likely second project after month one.
• What Are Diners Really Saying? Using AI to Analyse Your Reviews — A low-risk first job if reviews are your pain point.
• AI Software for Restaurants: POS-Built AI or an Add-On Tool? — Check what your POS already includes before paying for more.
• Toast IQ Team or Toast IQ Grow: Which Suits a Small Restaurant? — One POS-built AI option, priced by tier.
• AI Startups for Restaurants: Protect Your Number, Data and Exit — Protect your data and exit rights with any tool a consultant picks.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Smith.ai pricing page (September 2026) for the per-call example, OpenAI business plan pricing, Contra and Clutch published rate guides (September 2026).