AI Automation for Restaurants: Back-of-House Jobs to Start With

Start with prep lists, then supplier order drafts, and leave rota drafts until last. Prep lists come from sales data you already hold and a wrong one costs little; orders need written par levels first; rotas touch pay, breaks and people's lives. In every case the AI drafts and the chef or manager approves before anything is sent.

Much of this work is arithmetic on your sales and stock, so a spreadsheet does the sums and the AI handles the untidy parts: a supplier's emailed price list, a chef's notes, a booking comment. That is why the split between rule-based automation and AI matters in a kitchen. Everything below is back-of-house. Bookings, phone answering, reviews and marketing are separate jobs with their own tutorials.

Follow me on Instagram@sagnikteaches

Why prep lists come first, orders second and rotas last

The order comes from two questions. What does the AI need that you already have? And what does a wrong draft cost before someone catches it?

Connect on LinkedInSagnik Bhattacharya

Prep lists fail softly and the evidence appears the same night. Orders fail in a day or two, when the delivery arrives. Rotas fail slowly and publicly, and a published rota is awkward to retract. Moving in that order means each stage trains you to review AI output before the stakes rise.

Subscribe on YouTube@sagnikteaches

Write down what the kitchen already knows

Nothing here works until the numbers living in people's heads are on paper. The head chef knows that Friday needs 6 kg of chicken thigh and that the butcher cuts off orders at 14:00 on Wednesday. The AI doesn't, and neither does a new sous chef. Start with a par sheet for the 30 or so lines you order every week. This is a filled-in extract (illustrative).

A par is the quantity you want on the shelf at the start of a service period; the order is par minus what you've counted. The tutorial on par levels and orders with AI goes through how to set them from your own history, and the sheet above is the minimum to start. Add allergen notes for anything a supplier might substitute, because you will need them in stage two.

No par levels yet? Mine them from old invoices

If nobody has written par levels down, your invoices already hold a first guess. Take eight weeks of lines from one supplier and ask for the range of weekly quantities, not a recommendation (illustrative):

Here are eight weeks of invoice lines from our butcher (CSV: week, item,
quantity, unit). For each item give the weekly minimum, median and
maximum quantity. Mark any item whose maximum is more than double its
minimum. Do not guess units: if a unit is missing, list the line under
"needs checking".

Chicken thigh, boneless (kg): min 11, median 14, max 19
Beef short rib (kg): min 8, median 12, max 12
Duck leg (each): min 0, median 0, max 24 -> max more than double min
Lamb shoulder (kg): unit missing in weeks 3 and 6 -> needs checking

Invoices show what you bought, not what you needed. A week with 19 kg of thigh might be a week you over-bought and binned the rest, so use the median as a first par and ask the chef which weeks were abnormal. The duck line is a typical catch: it was ordered for one event, and a par built on it would stock duck for ever. The missing units are the other typical catch, and the instruction not to guess is what surfaced them.

Stage one: a Friday prep list from four weeks of sales

Here is a 45-seat restaurant that preps its Friday by gut feel. The head chef agrees to test a list built from the last four Fridays, plus pre-orders for a 14-person booking. The first step is getting the item sales out of the till. In Toast the route is Reports, Menus, Product mix, which exports to CSV or Excel and compares periods, although it can't be emailed on a schedule. In Square it is Reports, Custom, Custom item report, and renamed items appear as separate entries, so tidy item names before you export. Then give the assistant the data and the rules (illustrative):

You are helping the head chef of a 45-seat restaurant plan Friday prep.
Data: portions sold on the last four Fridays, plus pre-orders for
tomorrow's 14-person booking.
Rules:
- Forecast = average of the last four Fridays plus pre-orders, then trim
  10% because the average already includes some large-party nights.
- Portion specs: chicken thigh 180 g raw boneless; risotto rice 90 g,
  stock 350 ml, mushrooms 80 g.
- Round raw ingredients up to the nearest 0.5 kg or 0.5 litre.
- Flag anything that needs a decision. Do not invent dishes, specs or
  stock levels.

dish,fri1,fri2,fri3,fri4
Chicken thigh,31,27,35,29
Braised short rib,22,25,19,24
Mushroom risotto,18,21,17,20
Pre-orders: chicken 5, short rib 5, risotto 4

The arithmetic the AI should reproduce is plain. Chicken averages 30.5 over four Fridays, plus 5 pre-orders is 35.5, and a 10% trim gives 32 portions. Short rib averages 22.5, plus 5 is 27.5, trimmed to 25. Risotto averages 19, plus 4 is 23, trimmed to 21. Here is what a typical assistant returns:

FRIDAY PREP (forecast portions)
Chicken thigh: 32 portions = 5.76 kg raw boneless, prep 6.0 kg
Braised short rib: 25 portions (braise the usual batch of 30)
Mushroom risotto: 21 portions
  rice 1.89 kg (prep 2.0 kg), stock 7.35 litres (prep 7.5 litres),
  mushrooms 1.68 kg (prep 2.0 kg)
FLAGS
- Chicken marinade needs 6 hours: start by 14:00 for the 20:00 peak.
- Risotto pre-order of 4 is for the 14-person table: confirm vegetarian
  stock.

The numbers are right. Two lines are not from your data: "the usual batch of 30" and "the marinade needs 6 hours". Nobody gave the assistant a batch size or a marinating time, and both sound plausible, which is exactly the danger. The chef deletes them or supplies real figures. The trim also deserves a decision. Adding pre-orders on top of an average that already contains large-party nights can double count, and only the chef knows whether the last four Fridays included a similar booking. That judgement is why the output is a draft.

Run this beside the chef's own list for two Fridays without acting on it, then compare dish by dish. If the AI is within about 10% on most dishes and the chef can explain the gaps, move to live use. The wider forecasting method, including weather and events, is in forecasting covers and cutting food waste with AI.

Stage two: supplier orders from par levels and a closing count

Orders are prep lists with money attached, so they start only after the par sheet exists and someone does a closing count. Keep the count to the 30 or so lines you order weekly; it takes about eight minutes on a phone form, and a shorter list gets done every night. The AI's job is to subtract the count from the par, round up to the supplier's pack size, check the price against the last invoice and write the message. This is a Thursday-night produce order drafted that way, using weekend pars that cover Friday to Sunday rather than the single-day Friday pars on the sheet above (illustrative).

PRODUCE ORDER - delivery Friday morning, cut-off 18:00 Thursday
Pars cover Fri-Sun
Chestnut mushrooms: par 6 kg, counted 1.2 kg, need 4.8 kg -> 3 trays (6 kg)
Lemons: par 12 kg, counted 4 kg, need 8 kg -> 1 case (15 kg)
Flat-leaf parsley: par 10 bunches, counted 2 -> 8 bunches
Baby spinach: par 3 kg, counted 0.5 kg -> 3 bags (1 kg each)
PRICE FLAGS
- Chestnut mushrooms: last invoice $5.40/kg, current list $6.35/kg (+17.6%)
CHECK
- Lemons: one case is 7 kg over par; confirm you want the full case.

Four things to check on every draft. Units: kilograms against trays against cases is where order drafts most often go wrong. Price movement: the 17.6% rise on mushrooms is the sort of thing an agent can flag in a second and a busy chef misses for a month. Substitutions: if a supplier swaps a product, the allergen profile can change, and that check stays with a person. Minimums and cut-offs: the draft doesn't know that this supplier has a minimum order value unless the sheet says so.

A realistic mistake shows why the unit check matters. A par sheet said "double cream: 6", meaning six litres. The supplier sells cream in cases of six 2-litre tubs, the draft ordered "6" in the supplier's own unit, and 72 litres arrived on Friday. Nobody looked at the unit column because the number looked sensible. The fix was cheap. Every par line now carries its unit and pack size, and the AI's last step restates each line in the supplier's own unit with the litres or kilograms that equals, so a human reads "6 cases = 72 litres" before anything is sent.

The message to the supplier should be boring: lines, quantities, delivery day, and a request to confirm stock and price. If five suppliers used to take 12 minutes each, that was 60 minutes a week. With drafts ready, the review takes about 20, a saving of 40 minutes a week, nearly three hours a month. For price comparison across suppliers, comparing prices and drafting supply orders goes further.

If you prefer ready-made software, MarketMan listed Starter at $249, Growth at $299 and Enterprise from $449 a month at the last check, and its purchasing page names Square, Toast and Lightspeed as point-of-sale partners. Its pages don't say which plan includes suggested orders from par levels, so ask before you buy. Square's Order Guide, on its Restaurants Plus and Premium plans, exports PDF, CSV or XLSX and carries a 750-item limit; Square's release notes label it beta, so test it with a small supplier first. Toast IQ's base assistant comes with the Toast point-of-sale system at no extra cost.

Stage three: a kitchen rota draft that respects the rules you give it

A rota is where an AI draft most often looks finished and isn't. It will balance hours against a forecast with ease. It won't know your local rest and break rules, who is allowed to run the grill, or that two of your cooks can't work the same shift, unless you tell it. So the input is a constraint list, not just a spreadsheet of names:

Build a draft kitchen rota for Mon-Sun.
Forecast sales by day: Mon 2,800 / Tue 3,100 / Wed 3,600 / Thu 4,700 /
Fri 8,600 / Sat 9,200 / Sun 6,000 (dollars).
Kitchen labour target: 12% of weekly sales.
Staff (initials only), role, contracted hours, availability: [table].
Rules: minimum 10 hours between the end of one shift and the start of the
next; at least one person who can run the pass on every service; nobody
over their contracted hours without a flag; list every rule you could not
satisfy.

Suppose the draft comes back costing $5,040 against weekly sales of $38,000. That is 13.3% and the target is 12%, which is $4,560, so the draft is $480 over. At an average of $17 an hour that is about 28 hours to remove or justify. The draft also puts one cook on a close on Tuesday and an open on Wednesday with a seven-hour gap, breaking the ten-hour rule it was given, and it hasn't listed that as unmet. Both are normal first-draft faults. The manager removes the hours that serve no peak, fixes the gap, and asks the AI to check the revised rota against every rule, listing any it still fails. The 10-hour gap is an example; set yours from your own contracts and local rules, and ask an adviser where you're unsure, especially for staff under 18.

Two practical notes. Decide whether labour percentage uses total sales or net sales before you start, because vendors differ (Toast's own labour guide divides by total revenue) and a mixed definition makes every comparison meaningless. And treat the staff list as personal data: initials, no pay rates the AI doesn't need, and a business plan whose terms say your content isn't used for training by default. The deeper method sits in building a restaurant rota from AI demand forecasts.

Café, pub or brigade kitchen: how the rollout order shifts

The order above suits a restaurant with a changing menu and a handful of suppliers. Other kitchens shift the emphasis.

A café kitchen with 20 seats and daily deliveries has little to order and a lot to bake. Its first job is the morning bake and prep list, because over-baking is where its waste sits. Orders can stay manual for weeks, since most lines arrive daily and a par sheet adds little. The rota is usually the owner's own, so a draft saves little.

A pub kitchen with a Sunday roast has one service that dwarfs the others. Its forecast is for a single day, driven by bookings made earlier in the week, so the prep list should be rebuilt on Thursday and again on Saturday from the booking count. Meat is ordered once, on Thursday, and a wrong order can't be topped up, which makes the order draft the stage to watch.

A 120-cover restaurant with a brigade is where the rota earns its place. Kitchen wages are its biggest controllable cost, and a draft by section (pass, grill, larder, pastry) is worth the setup. For a kitchen like that I'd run stages one and three together and add orders afterwards, the reverse of the small-kitchen order, provided every station's skills are written into the constraint list.

Running the weekly routine without opening a chat window

Do the whole routine by hand, with pasted data, until it has run about ten times. Automating a flawed routine only makes the flaw arrive on time. Once it's stable, a scheduled flow is simple: Sunday at 17:00, read the sales export from a shared sheet, ask the AI for Monday to Saturday prep, and email the result to the head chef.

The cost is small. On Make, a weekly flow of four or five steps uses a few credits a run. AI steps can use more than one credit, but the whole month stays far below the free plan's 1,000 credits. On Zapier, a flow with a sheet lookup, an AI step and an email can't run on the Free plan, which allows two-step Zaps only and 100 tasks a month. Professional costs $19.99 a month on annual billing ($29.99 monthly) for 750 tasks, and a weekly flow uses a handful. Triggers don't count as tasks, and each successful action step does.

The weakest link is still the data going in. A till export can be scheduled in some systems and must be pulled by hand in others, and the closing count has to be typed or tapped in by a person. A flow that runs perfectly on stale counts produces confident, wrong orders, so put the count date at the top of every email the flow sends.

Five numbers that show whether the drafts are earning their place

Review weekly and keep the sheet to five measures. The thresholds are my starting points; adjust them once you have your own baseline.

Take week five: 142 portions prepared for a Friday and 128 sold, so 14 over, which is 9.9% and acceptable. The same week there is one run-out, on short rib, and 4 order lines edited out of 22 drafted (18%). Nothing in that week says stop; the short-rib run-out says add the pre-order trim to the rules for large bookings.

Back-of-house jobs to leave off the list for now

• Temperature logs and food-safety records. AI can tidy and summarise them, but a named person signs them. A record that the AI "completed" is not evidence that anyone checked a fridge. If you want help with the plan itself, treat the AI output as a draft to verify line by line, not a finished plan.
• Allergen information. Keep the matrix under human control and update it when a supplier changes a product. The AI can flag a changed product code; it must not declare a dish safe.
• Recipe changes. An AI that rewrites a recipe to save cost changes taste, yield and allergens at once.
• Supplier negotiation and choice. The draft compares prices. Deciding that a more expensive butcher is worth it for consistency is yours.
• The physical count. Software can prepare the sheet; someone still opens the walk-in.

A ten-week rollout calendar

If a stage stalls, stay on it. The first stage is the cheapest place to learn how your kitchen's data behaves, and piloting AI in shadow mode explains the habit of running a draft beside the real thing, which is the safest way through weeks one to four.

Doubts kitchen managers raise

Do I need restaurant software to automate prep and ordering?

No. A weekly sales export from your till, a par sheet and a chat assistant cover the first two stages. Dedicated purchasing software such as MarketMan starts at $249 a month at list price, so earn the habit with a spreadsheet first and buy software only when the weekly routine has proved itself.

Can AI order from my suppliers without me checking?

Not safely at the start. An order commits money, and early drafts need corrections for pack sizes, price rises and substitutions. Let the AI draft, have the chef or manager review each line, and consider sending automatically only after weeks of drafts that needed almost no edits.

Is it safe to put staff names and availability into an AI tool?

Treat rota data as personal data. Use initials or staff numbers, remove pay rates you don't need, and use a business plan whose terms say your content isn't used for training by default. Consumer plans usually let you switch training off in privacy settings, but check the setting yourself.

How do I know the prep list is actually better than the chef's?

Run both for two weeks without acting on the AI version, then compare portions prepared, portions sold, waste and run-outs for each dish. If the AI list is within about 10% of the chef's on most dishes and the chef can explain the rest, move to live use.

Further reads

• AI Food Insights Tool: Toast IQ, Square AI or a Spreadsheet? — Pull the sales data from Toast, Square or a sheet.
• Recipe Costing Spreadsheet for Restaurants: Build It With ChatGPT — Cost each dish before you forecast it.
• AI in Hospitality Management: Catch Supplier Price Creep on Invoices — Catch supplier price rises on invoices.
• Restaurant SOP Template: Opening and Closing Checklists With AI — Write the routines the drafts depend on.
• Free AI Shift Scheduling for Restaurants: Where Free Ends — What free scheduling tools leave out.
• Restaurant Labour Cost Percentage: Test Rota Changes With AI First — Test rota changes against labour cost first.
• Can ChatGPT Write a HACCP Plan for a Restaurant? What to Verify — Why food-safety records stay with a person.
• Should an AI Chatbot Answer Allergen Questions for Your Restaurant? — Allergen answers are a front-of-house risk.
• AI for Restaurant Order Management: One Queue for Every Channel — One queue for every channel: choose where it lives, map each menu with AI, test phone ordering, and run the checks that catch dropped or doubled orders.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Toast and Square help pages for item sales reports and Square Order Guide; MarketMan pricing and purchasing pages; Zapier and Make pricing and credit documentation; ChatGPT, Claude and Gemini plan privacy pages.