Українська

AI

AI for a corner shop: how to calculate a supplier order, what to do with repricing and what an assistant will not write about beer and cigarettes

A corner shop owner does the same thing every week: looks at the shelves, recalls how last week went and dictates quantities to the supplier. An assistant helps here, but not where it is usually put. A formula calculates the quantity, while the model writes the message, checks price tags after repricing and replies to reviews. It writes nothing about cigarettes, and that is how it should be.

Updated 2026-11-11 8 min Figures carry a check date
Illustration generated with AI (OpenAI gpt-image-2.5-flare) and labelled with IPTC metadata.

Five tasks of a corner shop

  1. A supplier order. How much milk, bread and eggs to take until the next delivery. A mistake costs either an empty shelf or written-off stock.
  2. Repricing. The supplier has sent a new price list, and someone has to find which price tags to reprint.
  3. Reviews on maps. Reply to a complaint about an expired yoghurt so that the next customer sees the shop responds.
  4. Announcements on social media and messengers. A new delivery, a promotion, holiday opening hours.
  5. Checking delivery notes. Whether the quantity and price on the note match the order. How to recognise documents and what stays with a person is covered in the article on AI for primary documents.

This article covers the first four. They share one thing: the model writes text and compares lists, while the number in the order, the price on the shelf and the receipt are decided by a spreadsheet, the owner and the till.

Supplier orders: a formula, not a model

Ask an assistant 'how much milk should I order' and it will name a number without showing where it came from. The only way to check such a number is to recalculate it yourself. So the quantity is better calculated in a spreadsheet with a visible formula:

Order = average daily sales × (days until the next order + delivery days) + safety stock − stock on hand − stock in transit.

Safety stock covers a week that turns out stronger than average: 1.65 × the spread of daily sales × √(days until the next order + delivery days). The factor 1.65 means that with an ordinary spread of sales the stock will last in roughly 95 cases out of 100. How to choose this level and what a mistake costs in each direction is covered in the article on stock for the December peak.

Products with a short shelf life have a second limit: you cannot order more than the shop will sell before the batch spoils. Cap = average daily sales × (delivery days + shelf life) − stock on hand − stock in transit. The smaller of the two numbers wins.

Product (illustrative)Sales over 4 weeks, pcsDeliveriesShelf life, daysOn handBy formulaShelf-life capOrder
Milk 2.5%, 0.9 l84, 91, 78, 95every 3 days, 1 day in transit72237.177.438
Wheat bread140, 152, 133, 147every 2 days, 1 day in transit21852.243.343
Eggs C1, 10 pcs38, 41, 35, 44weekly, 2 days in transit253028.0122.429
Buckwheat, 1 kg12, 9, 15, 10weekly, 3 days in transit36525−3.4579.60
Instant coffee, 100 g6, 8, 5, 7every 2 weeks, 5 days in transit540912.2497.113

All figures in the table are illustrative. The spread of daily sales is estimated from weekly figures: the standard deviation of the four weeks divided by √7. Fractional results are rounded up, and down where the cap applied.

Bread shows why the second limit matters. The formula asks for 53 loaves (52.2 rounded up) because the next delivery is three days away. But bread keeps for two, so only 43 will sell in time. That is a signal not for the model but for the owner: bread needs more frequent deliveries, because a bigger order goes to write-off. Buckwheat shows the opposite: stock on hand is enough, and there is nothing to order.

The model does three things in this process:

  • Writes the order text from the ready numbers in the format the supplier expects: a table, a messenger message, lines with item codes.
  • Flags anomalies. A product selling twice as fast as usual, or an item that has not sold for three weeks. The owner draws the 'why'.
  • Reminds you of what is not in the data: holidays, road works, a competitor opening next door. The formula sees only past weeks.

The model takes the quantity in the message from the spreadsheet rather than calculating it. If you do ask a model to calculate, the article on AI for sales spreadsheets explains the check: the model writes the formula, a person compares it with a control total.

Repricing and price tags

The Procedure for Trading Activity (Cabinet of Ministers Resolution No. 833, paragraph 20) requires prices to be shown on labels (price tags) or price boards. Article 15 of the Law 'On Consumer Protection' treats price as part of the product information a buyer must get before buying. If the seller finds that product information is inaccurate, it brings the information into line within a week.

In practice repricing breaks in one place: the price in the till software has changed, but the tag on the shelf is still the old one, or the other way round. Here the model is useful as a comparer of two lists:

  1. Export the old and new supplier price lists and the current prices from the till software.
  2. Ask the model to find the items whose purchase price has changed and to make a list of tags to reprint, with the shelf or section.
  3. You set the new retail price. The model can show how the markup changes, but the price is your decision. How many customers you can lose on a price increase without losing profit is calculated in the article on raising prices; how to build the exchange rate into the price of an imported product is in the article on currency risk.
  4. After reprinting, a shop assistant checks the tag against the till price for a few items in each section.

If the shop writes 'was — now' on promotional tags, the rules for a crossed-out price are covered separately: crossed-out price and discount.

Alcohol and tobacco: what the assistant will not write

For a corner shop this is the most sensitive part. A model will happily write a post 'Fresh beer at a discount', because it does not know that in Ukraine a separate article regulates this. Article 22 of the Law 'On Advertising' (version of 31.07.2026, status 'in force') treats alcohol and tobacco differently.

What the assistant may prepare for excise goods under Article 22 of the Law 'On Advertising'
A post on Instagram, Facebook or a Telegram channel✗ any online advertising is banned (part 1 item 8)≈ only with a warning on 15% of the area and no images of drinking (parts 3, 7)
A discount, 'second pack free', an exchange✗ any compensation for a purchase is banned (part 1 items 10–11)≈ as advertising: with the same limits of parts 3 and 7
Mailing by post, e-mail, to phones✗ banned to an indefinite circle of persons (part 6 item 8)✗ the same ban (part 6 item 8)
Product information inside the shop≈ not advertising under Art. 8 part 7, but sales promotion is banned✓ information at the point of sale is not advertising (Art. 8 part 7)

✓ — may be prepared; ≈ — may be prepared under conditions a person checks; ✗ — do not prepare. Parts of Article 22 in brackets unless stated otherwise.

A few details the model does not know without a prompt:

  • The alcohol warning. The text 'Excessive alcohol consumption is harmful to your health' takes up at least 15% of the advertisement's area, black on white (part 7). Alcohol advertising also may not show the act of drinking, use people under 18 as models or create the impression that most people drink (part 3).
  • Mailings. The ban in part 6 item 8 covers messages to 'an indefinite circle of persons'. Whether your messenger subscriber base is a defined circle is a matter of interpretation, and the text of the law gives no direct answer. It is safer to keep alcohol and tobacco out of mailings altogether.
  • Outdoor and indoor alcohol advertising is banned (part 2 items 5, 8). But the law does not count information about goods sold at a place of trade as indoor advertising (Art. 1). A shelf of wine with a price tag is not advertising; a beer banner on the facade is.

A practical rule for the assistant: state in its instructions that it does not prepare posts, promotions or mailings about tobacco, e-cigarettes and liquids, and prepares alcohol posts only with the warning and a human check. Without such an instruction the model will not bring it up. This is not legal advice: for an alcohol promotion that goes beyond an ordinary post, ask a lawyer.

Receipts and the till: no assistant here

The Law on Cash Registers (Article 3) requires settlement for the full amount through a cash register or software cash register, creating a settlement document and handing it to the buyer, including on returns. The till creates the receipt at the moment of payment, and no model takes part in this chain: it does not create a receipt, correct it or decide what to refund.

One detail concerns a shop with excise goods specifically. The same article lets a sole trader on the single tax who is not a VAT payer put a generalised product name on the receipt, showing the product group, but not for excise goods. So if the assistant helps maintain the product catalogue for the till, the names of alcohol and tobacco in it must be full, not shortened to 'drinks' or 'tobacco products'.

What it costs per month

An illustrative shop: three suppliers, an order once a week, so 13 messages a month (13,000 input tokens — the formula table and rules, 1,500 output). Four repricings (8,000 and 3,000), 40 review replies (700 and 200) and 8 announcements (1,000 and 400). The volumes are illustrative; Ukrainian tokenises less efficiently than English.

Prices from the OpenAI API pricing page on 11 November 2026 (standard processing, short context, per million input and output tokens): gpt-6-luna — $0.10 and $0.50, gpt-6.1-terra — $1 and $5, gpt-6.1-sol — $2 and $10, gpt-6-astra — $10 and $50. The National Bank rate on 11 November 2026 is UAH 45.5013 per dollar.

Monthly cost of 13 orders, 4 repricings, 40 replies and 8 announcements by model, UAH
gpt-6-lunaUAH 2
gpt-6.1-terraUAH 20
gpt-6.1-solUAH 41
gpt-6-astraUAH 205

Calculation: (13 × (13,000 × input + 1,500 × output) + 4 × (8,000 × input + 3,000 × output) + 40 × (700 × input + 200 × output) + 8 × (1,000 × input + 400 × output)) ÷ 1,000,000 × 45.5013. Volumes illustrative; OpenAI API prices and NBU rate on 11.11.2026.

One order on the most expensive model costs UAH 9.33; on the cheapest, 9 kopecks. For a corner shop the price of requests is no argument either way: a seat on a business plan of a chat assistant (ChatGPT Business or Claude Team, $20 a month billed annually) costs UAH 910, more than the whole monthly volume through the API even on the most expensive model. A chat, on the other hand, needs no setup. How to choose between a chat, a builder and custom development is in the article on chatbots for business.

What to give the model and what not

Should this shop task go to a model
  1. Is the result a number in an order?A quantity the shop will buy and pay for. Yes → a spreadsheet formula; the model only writes the message from ready numbers
  2. Is the result a shelf price?Retail price, markup, promotion. Yes → the owner decides; the model lists the tags to reprint
  3. Does the text mention tobacco, e-cigarettes or liquids?A post, a promotion, a mailing. Yes → do not prepare it at all
  4. Does the text advertise alcohol?A post or announcement outside the shop shelf. Yes → only with a warning on 15% of the area and a human check; not in mailings
  5. Does it concern a receipt or a refund?A settlement document, an amount, a return. Yes → the till and the shop assistant only

The model drafts review replies; the owner publishes them. What you may ask customers for and what you may not promise for a review is in the article on customer reviews. The customer's name and phone number, if they left them in a complaint, stay out of the request: the reply does not need them. General rules are in the article on data for an AI assistant.

A first step for a week

  1. Export from the till software the sales for the last four weeks for the 20 products that most often run out or get written off.
  2. Calculate the order in a spreadsheet with the formula above, using each product's delivery schedule and shelf life.
  3. Put next to it the quantity you would have ordered by eye. Where the difference is more than a third, think about who is right and write down the reason.
  4. Ask the assistant to write the order text for each supplier from the spreadsheet and to flag products with unusual sales. A week later compare write-offs and empty shelves with the previous week.

Our position

In a corner shop the model should not decide how much and at what price. A formula that can be seen and checked calculates the quantity, the owner sets the price, the till issues the receipt. The model is useful where there is a lot of text and comparison: a message to the supplier, a list of tags after repricing, a draft reply to a review. A separate rule for excise goods has to be written into the assistant's instructions, because it does not know the law on its own.

The position does not hold for a chain store or a franchise: there orders and prices are usually set by a central system, and your own calculation may contradict it. The order formula also works poorly for products that sell less than once a week and for seasonal items: on such data the four-week average and spread say nothing.

Checked on 11 November 2026 on zakon.rada.gov.ua (Law 'On Advertising' Arts. 1, 8, 22, version of 31.07.2026; Law 'On the Use of Cash Registers…' Art. 3; Law 'On Consumer Protection' Art. 15; Procedure No. 833 para. 20), on the OpenAI API pricing page, the ChatGPT and Claude pricing pages and at the National Bank of Ukraine. The arithmetic was recalculated in code. Sales, stock, delivery schedules and tokens are illustrative.

Still not sure this applies to you?

Describe the situation. The first reply says whether this is our work and how long the diagnostic takes.

Request a diagnostic