How to Use ChatGPT for Restaurants: Five Jobs to Hand It Each Week

Hand ChatGPT five drafting jobs across the week: Monday's draft order, Tuesday's prep list and costed special, Wednesday's rota draft and pre-shift note, Thursday's weekend promotion, and Sunday's one-page review with replies to reviews. Keep allergen answers, food-safety records, the final order and anything with guest names in human hands.

The rule that makes this work: ChatGPT drafts, you decide, and it never sees a guest's name. It doesn't know your walk-in, your covers or your supplier's cut-off time, so every job starts with numbers you hold and ends with a check you do. What it removes is the blank page, not the responsibility.

Follow me on Instagram@sagnikteaches

Set up one Project so Monday starts in two minutes

Create a ChatGPT Project for the weekly routine and keep every job inside it, so the instructions and files are there each time. Use a Project rather than a custom GPT: OpenAI has stopped new custom GPTs, and existing ones are scheduled to end on 11 December 2026.

Connect on LinkedInSagnik Bhattacharya

Put four things in the Project's files or instructions:

Subscribe on YouTube@sagnikteaches

• The ordering sheet: each item with its supplier, case size, typical weekly use, safety stock and unit price.
• The supplier calendar: delivery days, order cut-off times and minimum orders. Without it ChatGPT will happily schedule a delivery for a day nobody comes.
• The menu with prices, with no allergen column. Keeping allergens out stops a draft from quietly repeating them.
• A house-voice note: three sentences on how you talk to guests, plus words you never use.

Add standing instructions along these lines: use metric units and the spelling on my menu, show arithmetic line by line, use only figures I provide, never state allergen or dietary information, and say "I need X" when an input is missing. OpenAI lists a file limit per plan (as of October 2026, five files on Free and 25 on Go and Plus), so one spreadsheet with a tab per list is easier to keep inside it. If you want a nudge, a ChatGPT scheduled task can post the week's input checklist at 7:30 each Monday; Plus users can keep five tasks active and Free or Go users three, according to OpenAI's help page, although its pricing page lists scheduled tasks only from Go up. It can remind you, but it can't count your stock.

Monday: turn counts and covers into a draft order

The inputs are Sunday-close counts, the ordering sheet, the supplier calendar and the week's forecast covers, taken from the booking book as numbers only (60 on Friday, 85 on Saturday). Add anything unusual, such as a party of 30 or a closed day. The sum behind every line is the same: typical use x (1 + uplift) + safety stock - on hand, rounded up to whole cases. In a spreadsheet that is one formula, =CEILING(MAX(0, use*(1+uplift)+safety-onhand), case), and both Excel and Google Sheets understand it. For a long list I'd ask ChatGPT to write that formula and let the sheet do the arithmetic; for a short one, the prompt below works.

Draft Monday's order for a 45-seat bistro. Use only the figures in the attached ordering sheet and the notes below.
Counts at Sunday close are in the on_hand column.
Forecast covers: Fri 60, Sat 85, Sun 50 (a normal week is 50, 70, 45).
Saturday has one party of 30 on the set menu, so use these uplifts: chicken thighs 15%, double cream 10%, salmon 10%.
For each item: need = typical_use x (1 + uplift) + safety_stock - on_hand. Round UP to whole cases. Show the arithmetic on one line per item.
Group by supplier. Check each supplier's delivery day and cut-off in the supplier calendar and flag any order that would miss its cut-off.
Do not invent prices. List any item with missing data under QUESTIONS.

Illustrative output for three lines, with my corrections:

Three habits keep this job honest. Make the model show its arithmetic, because you can scan a line in five seconds. Round to case sizes with the formula, not the chat. And compare the order's total cost with last week's invoice: a swing of more than about 15 percent (my own rule of thumb) usually means a wrong input, so look before you send anything. You place the order and you check the delivery. The tutorial on par levels and orders with AI goes further on the ordering half of the job.

Tuesday: prep quantities and one costed special

Prep is the same sum in a different place. Give ChatGPT Saturday's forecast (85 covers), last Saturday's mix of mains from the till report and each batch's yield. Say braise accounts for 40 percent of covers and a batch makes 12 portions: 85 x 0.40 = 34 portions, 34 / 12 = 2.83, so make 3 batches. Do the same for the roast-chicken jus, the salmon dressing and the vegetarian base, rounding up, and you have the prep list in one pass. Check it against what the kitchen actually ran out of last week, since a mix from one Saturday can mislead.

The second Tuesday task is a special built from stock that needs using. Suppose the walk-in holds 2 kg of ricotta with a Sunday use-by and 6 kg of very ripe tomatoes, so the special has to sell by the end of the weekend. Give the model those quantities, your menu style and your own unit costs, and ask for two ideas with per-portion quantities, the plate cost and a menu price at your food-cost target (30 percent here; use yours).

Illustrative result for the better idea, baked ricotta with roast tomatoes and grilled bread. Ricotta at 120 g and $9.00 a kg is $1.08, tomatoes at 200 g and $3.20 a kg are $0.64, oil and herbs $0.30, bread $0.40 and garnish $0.18. The plate cost is $2.60. Divide by 0.30 and the price is $8.67, so I'd charge $9.00, which is a 28.9 percent food cost. At 120 g a portion, 2 kg of ricotta gives 16 portions, which is the number Thursday's promotion can honestly offer.

The first draft contained a mistake worth recognising. It had added pine nuts to the dish, so the sum carried a $0.35 cost line nobody supplied, and the special had silently gained a tree-nut allergen. The fix is to delete the pine nuts and tell the model to use only the ingredients and costs you list. The special's allergen line is then written by a person from the supplier specification sheets, as covered further down, and the dish isn't announced until it exists.

Wednesday: a rota draft with the hours sum, and the briefing note

Give ChatGPT availability by initials, forecast covers per shift, minimum staffing for each station and a labour budget expressed as hours, using one blended hourly cost instead of anyone's real pay. Then ask it to draft the grid and, more usefully, to flag where the draft breaks your own rules. The flags are the point, because any tool can fill boxes.

Here is the sum for a Friday dinner (illustrative). Forecast 60 covers at an average spend of $38 is $2,280. A 28 percent labour target is $638, and at a blended $16 an hour that buys about 40 hours. The draft staffs four servers, three cooks, a kitchen porter and a bartender for six hours each, which is 54 hours and $864, or 37.9 percent of forecast sales. A good draft says so plainly. Your options are then yours to weigh: cut a server's first or last hour, stagger starts, question the forecast, or accept a higher labour percentage on your busiest night.

What stays with you is the rule check. ChatGPT doesn't know which working-time rules apply to you, what your contracts say about breaks, rest between shifts or younger staff, or who swapped with whom on Tuesday. Read the finished rota against those before it goes up. The tutorial on building a restaurant rota with AI demand forecasts covers the forecasting side.

The second Wednesday output is the Friday pre-shift note. Paste the numbers (60 booked, two tables of eight at 19:00 and 20:15), the special and its quantity, the 86 list, and the allergen line copied from your signed-off matrix. Ask for five short bullets. Illustrative draft, with the fix: it added "dairy-free" beside the sorbet, which the matrix doesn't say, so that came out. The note is only as good as the matrix lines you paste, which is why the model never writes them. For using till data to feed this note, see pre-shift meeting notes drafted from POS data.

Thursday: one weekend promotion, cut four ways

Write the facts once and let ChatGPT reshape them: the dish, the price, the quantity, the days and how to book. Ask for an Instagram caption, a two-sentence blurb for the mailing, a Google Business Profile post and a counter card. Paste only the blurb text into the email tool yourself. The mailing list never goes into ChatGPT.

A before and after shows what editing means here. The first draft read: "Treat yourself this weekend! Indulge in our mouth-watering baked ricotta, made with the freshest local ingredients." Every claim in it is unprovable. The edited version: "Saturday and Sunday only: baked ricotta with roast tomatoes and grilled bread, $9. Sixteen portions, then it's gone. Walk in, or book on our website." The second gives a guest the price, the limit and the next step, and it contains nothing you can't defend.

Check each version against this list before it goes out:

• Price, days and quantity match Tuesday's costing.
• No claim like "locally sourced", "vegan" or "gluten-free" unless the matrix and your supplier paperwork support it.
• Instagram allows at most five hashtags a post (the cap since December 2025), so cut any list of fifteen a model hands you.
• A Google Business Profile post framed as an event needs a start and an end date.
• Any photo is yours or licensed, and the post says how to book.

Run it every Thursday and the routine stays small: one set of facts in, four pieces of copy out, five checks before anything is posted.

Sunday: the week in one page, and replies to reviews

At close, take the week's totals from the till report. Totals only: no covers broken down by guest, no loyalty data. Illustrative week for the 45-seat bistro: sales $16,800, 442 covers, average spend $38.00, labour 33 percent, food cost 31 percent, voids $146, waste log $212. Paste them with last week's figures and ask for a one-page note.

Here are this week's and last week's totals for a restaurant, plus the five best and five worst sellers by margin.
Write a one-page note for the owner: what changed, three questions to ask the team, and one thing to try next week.
Use only the numbers I gave you. If a cause isn't visible in these numbers, don't offer one; label any hunch as a guess.
Keep the tone plain. No headings longer than five words.

Illustrative reply: covers rose 6 percent (442 against 417), but average spend slipped from $39.40 to $38.00. Questions: did the Saturday party's set menu pull the average down; which dish moved most between the best and worst lists; and why did voids reach $146. The draft also said the dip was "probably due to the weather", which no figure supports, so cut it. That is the usual fault with this job: a model with a pile of numbers wants a story. Insist on questions, not causes. If your till already offers its own assistant (Toast IQ's base assistant is included with Toast POS), try it on the same week and see which note you'd rather read.

The second Sunday task is review replies. Paste the review text without the name, and give the facts you've verified yourself. A review says: "Waited 40 minutes for mains and nobody told us why." First draft: "We're so sorry! Dinner is on us next time, just show this message." Problems: the free dinner is an inducement that may break Google's review rules, which bar offering anything for posting, revising or removing a review, and it invites a screenshot. It also admits nothing specific. The version I'd post: "Thank you for telling us. Forty minutes is too long, and not being told why is on us. Please email the restaurant with the date and we'll look into what happened that evening." It uses no names, mentions no booking details and moves the conversation offline. For complaints that are really about illness or injury, don't reply with a draft at all until you've recorded the facts. The tutorial on answering restaurant complaints with AI covers the tone in more depth.

What stays with a person, and why

Handing over a draft is different from handing over the decision. These jobs stay manual, with the part ChatGPT can still help with.

The tutorial on which hospitality jobs to delegate and which stay with staff takes the same question across hotels and bars too.

Allergen information: ChatGPT never holds the pen

Which allergens must be declared, and how, depends on the rules that apply to your business, and ChatGPT doesn't know which rules those are. It also can't see your supplier's specification for today's delivery, which is where the real answer sits. So the chain is: spec sheet, then a matrix a named person maintains, then a sign-off by the head chef after every recipe or supplier change, then staff trained to read it, then the guest. ChatGPT isn't a link in that chain.

Here is how the shortcut fails (illustrative). Asked to "make the allergen sheet for the new dessert from the recipe", a model returns "contains milk, eggs, wheat". The recipe listed "gluten-free flour blend", and the supplier's sheet shows that blend includes almond flour, so the dessert contains tree nuts. The draft was tidy and confident, and nothing in it flagged the gap. Only the comparison with the supplier's paperwork caught it.

If you do use ChatGPT to turn a finished matrix into a printable sheet, treat the result as untrusted: a second person checks every cell against the source, and the file stays out of any prompt that creates guest-facing wording. Never ask it whether a dish contains something, and never let a front-of-house message answer an allergy question from a chat. The tutorial on whether an AI chatbot should answer allergen questions sets out the same line for guest-facing bots.

Guest data: what to paste instead

The rule is easy to state and easy to break by accident, because guest data hides in ordinary exports. Swap each of these for something safe:

• The reservation book: paste covers per time slot and the number of large parties, never names, numbers or notes. Notes often carry allergies, which are health data.
• Reviews: paste the text with the name deleted. Don't add the booking you think it refers to.
• Delivery-app exports: paste totals by day. The rows carry names and addresses.
• The mailing list and loyalty data: keep them in the email tool. Paste only the copy you've written.
• Staff: use initials, with no pay rates, absence reasons or personal circumstances.

On a personal plan, find the model-training setting in privacy settings and switch it off before you start. Business workspaces are kept out of model training by default, though ChatGPT Business needs at least two seats. Neither setting makes it acceptable to paste guest names: if you can do the job without them, do.

Four weeks in: did the five jobs earn their time?

Keep a one-line log each Sunday: minutes spent on the five jobs, the number of corrections you made to the order, the waste figure, and one thing the model got wrong. An illustrative week is about 145 minutes: 25 on Monday, 30 each on Tuesday and Wednesday, 20 on Thursday and 40 on Sunday. After four weeks compare against the four before: waste, order corrections per week, and whether anyone stopped using a job. Drop any job you skipped three times, because it didn't fit your week, and tighten the inputs on any job whose corrections aren't falling.

On cost, a free plan is enough to test every prompt here, although data analysis on spreadsheets is limited on Free according to OpenAI's pricing page. For these five jobs, a paid individual plan (Go at about $8 a month or Plus at about $20, list prices) chiefly raises those limits rather than adding anything new. The routine pays for itself only if you act on what it surfaces, such as the Friday labour overshoot, the salmon box size or the dip in average spend. If the notes pile up unread, the fix is a shorter note, not a better prompt.

Questions managers ask before handing work to ChatGPT

Can ChatGPT answer a guest's allergy question?

No. Allergen answers must come from supplier specification sheets, a matrix that a person maintains and signs off, and staff trained to use it. ChatGPT cannot see your recipes or deliveries and sounds confident when wrong. At most, use it to lay out a blank template, then have a second person check every cell.

Is it safe to put rota details into ChatGPT?

Use initials instead of names, leave out pay rates and personal circumstances, and keep contracts and absence reasons out. Treat the draft as a suggestion and check it against your working-time rules before publishing. Business plans don't train on content by default; on a personal plan, switch off the model-training setting.

Which ChatGPT plan does a restaurant need?

Start on a free plan to test the prompts, then consider a paid individual plan for higher limits and data analysis on spreadsheets. ChatGPT Business suits several managers sharing a workspace but needs at least two seats. Plans and limits change, so check OpenAI's pricing page before paying.

What if ChatGPT's draft order is wrong?

Expect some lines to be wrong. Have it show the arithmetic per line, round up to supplier case sizes yourself, and compare the total with last week's invoice. You place the order, so a mistake costs a correction rather than a delivery. Log errors for four weeks, and tighten the inputs if they don't fall.

Further reads

• Restaurant Supply Ordering: Compare Prices and Draft Orders — Compare supplier prices and draft orders in more depth.
• How to Forecast Covers and Cut Food Waste With AI — Forecast covers and cut food waste before you order.
• Restaurant Google Business Profile: AI-Drafted Menu, Hours, Photos — Draft Google Business Profile posts for the weekend.
• Restaurant Email Marketing: A Weekly Send Written by AI — A weekly email written with AI from your own list.
• AI Food Insights Tool: Toast IQ, Square AI or a Spreadsheet? — Compare Toast IQ and Square AI with a plain spreadsheet for Sunday's numbers.
• Recipe Costing Spreadsheet for Restaurants: Build It With ChatGPT — Build the costing sheet behind Tuesday's special.
• ChatGPT Prompts for Restaurants: 10 for Menus, Rotas and Replies — Ten more prompts for menus, rotas and replies.
• Write a Good Bio for a Restaurant Owner With ChatGPT: About Page and Press — An interview-style ChatGPT prompt turns a restaurant owner's own facts into three bios, for the About page, press requests and social, with no invented awards.
• Can ChatGPT Write a HACCP Plan for a Restaurant? What to Verify — What ChatGPT drafts well in a HACCP plan, what it guesses, and the checks your food safety authority still makes: limits, logs, thermometers and training.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT pricing page (checked October 2026), OpenAI help text on Projects and scheduled tasks, OpenAI's custom GPT retirement FAQ, Google Business Profile and review policy pages.