Six tasks that really take hours in a venue
This article is part of the series "AI for business by industry" and answers the same five questions: what, with what, how much, what not to do and where to start. It covers only tasks that exist in a café or restaurant and make no sense in an online shop or a clinic.
- Menus and dish descriptions. A new seasonal item, an English translation of the menu, a description for the website and the delivery service. Plus the list of allergens that the law requires a venue to give its guests.
- Recipe cards and food cost. The chef changed a recipe or the supplier raised the price of cream, and the cost of a dozen dishes has to be recalculated.
- Replies to reviews on Google Maps. Twenty new reviews in a week, three of them negative, and each one needs a reply that is not a template.
- Shift rota. Eight people, seven days, morning and evening shifts, someone studies, someone wants Friday off.
- Sales analysis from the till. Which dishes sell and how much the venue earns on each, not just the day's revenue.
- Dish photos. For Maps, the website and the delivery service. Separate rules apply; see below.
What to use
| Task | Tool | How to check the result |
|---|---|---|
| Menu, descriptions, translation | An assistant: "here is the composition from the recipe card, write a 25-word description" | Every ingredient in the description is on the recipe card; allergens come from the card, not from the description |
| Recipe cards and food cost | An assistant with a table, or spreadsheet formulas the assistant helped write | Recalculate one dish by hand; the price per kilo is multiplied by the gross weight |
| Review replies | An assistant with a short brief: tone, what may be promised, when to call the manager | A person reads every reply before it goes out; no discounts for changing a rating |
| Shift rota | An assistant with a table of wishes and constraints, or the rota module in the POS system | The four questions in the rota section below, for every employee |
| Sales analysis | An assistant with a sales export by dish and the cost from recipe cards | Total revenue in the analysis matches the till report for the same period |
| Dish photos | A phone camera and basic editing; image generators only for decoration, not for dishes | The dish in the photo is the one that is served, in the same portion |
If you get many messenger questions about bookings and opening hours, how to choose between a button bot and an AI assistant is covered in the article on chatbots for business.
What it costs
For a venue with one or two locations, AI spending means an assistant subscription for one or two people: the owner and the chef or the manager. According to the ChatGPT pricing page on 13 October 2026, Plus costs $20 a month, Business $25 per person billed monthly or $20 billed annually, from two people. At the National Bank of Ukraine rate for 13 October 2026 (UAH 44.9567 per dollar), Plus is about UAH 899 a month and Business for two billed monthly about UAH 2,248. On Business, content is not used to train models by default; on Plus you switch this off in settings yourself. Prices of other assistants are in the series hub.
A venue of this size usually needs no API, batch generation or automation: it has dozens of descriptions, not thousands. The real cost is the time spent checking, and the sections below show where that cannot be skipped.
Recipe cards: where AI gets food cost wrong
A recipe card lists each ingredient's net weight (what ends up on the plate) and gross weight (what has to be bought, allowing for peeling, trimming and waste). The supplier sells by the gross kilo. If you ask an assistant to "calculate the cost" and give it only net weights, it will multiply the price by the net weight and understate the cost. The mistake is invisible: the number looks plausible.
A hypothetical example: pasta with chicken and mushrooms, menu price UAH 285.
| Ingredient | Net, g | Waste | Gross, g | Price, UAH/kg | Cost from net, UAH | Cost from gross, UAH |
|---|---|---|---|---|---|---|
| Chicken breast | 120 | 10% | 133 | 220 | 26.40 | 29.33 |
| Dry pasta | 100 | — | 100 | 90 | 9.00 | 9.00 |
| Cream 20% | 80 | — | 80 | 110 | 8.80 | 8.80 |
| Parmesan | 20 | — | 20 | 900 | 18.00 | 18.00 |
| Mushrooms | 60 | 20% | 75 | 120 | 7.20 | 9.00 |
| Oil, spices | fixed | 5.00 | 5.00 | |||
| Total | 74.40 (26.1% of price) | 79.13 (27.8% of price) | ||||
The difference is UAH 4.73 per portion. It looks small, but at a hypothetical 600 portions a month that is about UAH 2,840 missing from the mark-up calculation. And this is one dish with two ingredients that have waste; on a menu with many of them (fish, meat on the bone, herbs) the gap is wider.
What this means for working with an assistant:
- Give it gross weights, not just net. Or give it the waste percentage and ask for gross weight in a separate column.
- Take waste percentages from your own weighing or your own recipe cards, rather than asking the AI to "use a standard one". It will use a number whose source you cannot check.
- Recalculate one dish by hand after every change to the formulas. If it matches, spot checks are enough for the rest.
Menus and allergens: what the law says
Dishes a venue sells to guests fall under Ukraine's Law "On Information for Consumers about Food Products". Part 9 of Article 6 requires that, for non-prepacked food, the venue must provide the product name and every ingredient from the allergen list that is used in preparation and remains in the finished dish, even in altered form. The venue chooses how: on the menu, on a separate card, on request. The list is in Annex 1 to the law and has 14 groups: cereals containing gluten, crustaceans, eggs, fish, peanuts, soy, milk (including lactose), tree nuts, celery, mustard, sesame, sulphur dioxide and sulphites above 10 mg per kg or litre, lupin, molluscs. Article 4 of the same law requires the information to be accurate and reliable.
That sets the limit for AI. An assistant can describe a dish well, translate the menu and bring names to one style. But it does not know that the cook thickens the sauce with flour and adds mustard to the dressing. Ask it to "write a description and mark the allergens", and it will infer allergens from the dish name, which means guessing.
The working order is this: the chef records allergens for every dish on the recipe card, including sauces, breading and frying oil. The assistant gets the recipe card and writes a description in which allergens are carried over from the card unchanged. When a recipe changes, the card is updated first, then the menu.
Reviews and photos on Maps
AI helps well with review replies: read the complaint, pick out the point, draft a reply in the venue's tone. The limits here are set not by AI but by Google's content policy for Maps.
| What you might want to do | What Google's rules say | Conclusion |
|---|---|---|
| Reply to every review with an assistant's draft | The reply is published in the venue's name, so the venue is responsible for what it says | Yes, if a person reads every reply |
| Give a discount or a dessert for a review | Incentives (payments, discounts, free goods or services) for a new review are prohibited | No |
| Offer compensation if the guest changes their rating | Incentives for editing or removing a negative review are prohibited | Compensation only with no condition attached to the review |
| Ask waiters to collect 10 reviews per shift | Requiring staff to collect a set number of reviews is prohibited | No; inviting guests to leave a review without a quota is fine |
| Write reviews in the names of friends or staff | Reviews must be based on real experience; conflict of interest is grounds for removal | No, with or without AI |
The assistant does not handle complaints about feeling unwell after a meal at all. That reply is written by the manager, because it may become part of an inspection or a dispute.
Dish photos. Google's photo requirements for Business Profiles say a photo should be in focus, well lit, without significant alterations and with minimal filters, in other words it should represent reality. An AI-generated "perfect pasta" does not meet that requirement, even if it looks like yours. The same logic holds for delivery services: the guest compares what arrived with the photo. An image generator is fine for a poster or a post background, but the dish itself is better shot on a phone in daylight, on the same plate it is served on.
Shift rota: what to check against the Labour Code
Building a rota is a task an assistant handles quickly: give it the list of people, opening hours and days-off wishes, and it lays out the shifts. But it optimises what you asked for, and it knows about labour rules only what you put in the brief. If staff are employed under employment contracts, the rota is checked against four provisions of Ukraine's Labour Code.
- Does anyone work two shifts in a row?Article 59 prohibits putting an employee on two consecutive shifts, for example a morning and an evening shift on the same day when these are two separate shifts. Yes → redo the rota
- Is the break between shifts at least twice the previous shift?Under the same Article 59, the break is counted from the length of the previous shift including the lunch break: after 12 hours, 24 hours of rest. A "two on, two off" rota of 12-hour shifts with only 12 hours between two day shifts does not square with a literal reading of this rule. No → show the rota to an HR specialist or lawyer
- No more than 40 hours a week?Article 50. If the venue has introduced cumulative working-time accounting (Article 61, agreed with the trade union body), the norm is counted over the accounting period rather than each week. More → redistribute shifts or record overtime
- Does everyone get uninterrupted weekly rest?Article 70 requires at least 42 hours. For the period of martial law, Article 6 of Law No. 2136-IX allows it to be cut to 24 hours. Less → move the days off
The simplest approach is to write these four rules straight into the assistant's brief: then it takes them into account when building the rota, and a person only checks the result. For people working as sole proprietors under a civil-law contract these Labour Code provisions do not apply, but then a shift rota for them is a matter of agreement, not an order.
What not to do
- Do not let the assistant decide the allergens. Only let it carry them over from the recipe card into the menu and the translation.
- Do not paste bookings with names and phone numbers into the assistant. A name and phone number are personal data under the Law "On Personal Data Protection". To analyse how busy you are, the date, time and number of guests are enough. Where to switch off training on your data is in the article on data for an AI assistant.
- Do not publish review replies unread. The assistant may promise things the venue does not do: delivery, a banquet hall, a gluten-free menu.
- Do not put a generated photo in place of a dish photo. Not on Maps, not in the delivery service.
First step for one week: a margin analysis of the menu
Of the six tasks, this one gives a result in hryvnias. Every till program has revenue by dish, recipe cards have the cost, but the two are rarely put together. Together they show which dish sells often and earns little, and which earns well but nobody orders.
- Export sales by dish for three months from the till: name, quantity, revenue. No guest data.
- Add the cost per portion from the recipe cards, calculated from gross weight as in the section above.
- Ask the assistant to calculate for each dish the margin per portion (price minus cost) and the margin for the period (margin per portion × quantity), sort them and split them into four groups: sells often and earns well; often but little; rarely but well; rarely and little.
- Check the total: the revenue in the analysis must match the till report for the same period. If it does not, the assistant has lost rows.
- Pick two or three actions: review the price or the recipe of "often but little" dishes, make "rarely but well" dishes more visible on the menu, drop one or two "rarely and little" items that take up kitchen space.
The work takes a day. A month later, compare margin per receipt with the previous three months. If it has not changed, your problem is not the menu but the number of guests, and that needs a different analysis.
Our position
For a café or restaurant with one or two locations, one assistant subscription is enough. Start not with social media texts but with a margin analysis of the menu and a recalculation of recipe cards. Turn on texts and review replies later, with the rule that a person reads every reply. Allergens and the shift rota the assistant only formats: the recipe card and the Labour Code decide.
This position does not hold for a venue that has no recipe cards, or whose cards do not match what the kitchen actually cooks. Then AI will multiply the inaccuracy, and the place to start is the recipe cards, weighing ingredients by hand, not the assistant.
ChatGPT prices were checked on 13 October 2026 on the ChatGPT pricing page; allergen requirements in the Law "On Information for Consumers about Food Products" (version of 16.09.2026); working-time rules in the Labour Code (version of 31.07.2026) and Law No. 2136-IX; review and photo rules in Google's help pages; the exchange rate from the National Bank of Ukraine for 13 October 2026. The recipe card example is hypothetical, and the ingredient prices in it are not market data.