AI for shops: customer messages, stock by size, returns and loyalty cards

The AI for shops is most useful outside the actual sale: messages arriving while you are at the till ("do you have size 38?"), replenishment by size and colour, exchange and warranty requests that get mixed up at the counter, and loyalty-card history sitting unused in the business software. These tasks already have data (stock levels, sales and commercial documents) and rules you can write down. An AI assistant connected to the business software can prepare or complete them according to your instructions.

The tax part of the cash, on the other hand, does not need artificial intelligence. The electronic recorder stores sales and transmits to the Revenue Agency the data of the daily fees, and from 2026 the instrument by which you accept electronic payments must be connected to the recorder (D.Lgs. 127/2015, art. 2, text in force). There useful automation is more modest: remembering deadlines and reporting the evenings when the accounts do not return.

Here we talk about physical stores with one or two clerks: clothing, footwear, housewares, hardware, perfumery. For those who sell especially online there is the guide to AI for e-commerce.

"Do you have the 38?": the answers on WhatsApp must read the warehouse

Many messages to a footwear or clothing store have the same shape: a screenshot from the Instagram profile and a variant, "this one in blue, number 38?". The answer lies in management software, where each article is a grid of sizes and colors. An AI assistant on WhatsApp or in Instagram messages only serves if it reads that grid.

With the link, the assistant searches for the article from the name, code or post from which the message arrives, checks the stock of the variant and responds with what is really there, including the near number or other color. If the customer wants to pass, he proposes to keep the pair aside until a time decided by you; those who ask for a sold out number leave the contact for notice on the next arrival. Channel and technical limits are explained in WhatsApp chatbots for businesses. Without connection, better a smaller assistant: hours, address, exchange policy, brands treated.

Before turning it on write down what it passes to a salesperson (defects, complaints, discount requests, orders of more pieces). And from August 2, 2026, the writer must know that he is speaking with an artificial intelligence system, as the guide to the AI Act for an SME.

Change, defect or return: three requests that are confused at the counter

Three different requests arrive at the counter that the customer calls all "returned". The first is the change: wrong size, color that does not convince at home, double gift. The right of withdrawal applies to distance contracts and those negotiated outside business premises (Consumer Code, Art. 52), so for a purchase made in store the change is your rule. It is advisable to write it, with time, condition of the item and document to carry.

The second is the defect, and here the rule does not choose it. The seller responds to the consumer of any lack of conformity that existed at the delivery and that manifests itself within two years (Art. 133). If the defect manifests itself within one year of delivery it is assumed that it already existed, unless proof to the contrary (Art. 135). The customer chooses between repair and replacement, unless the chosen remedy is impossible or involves disproportionate costs; if neither arrives, he can ask for a reduction in the price or, for minor defects, the termination of the contract (Art. 135-bisThese are the rules for contracts concluded after 1 January 2022.

The third is the return of courtesy out of every rule ("I put it once, it does not convince me"): the holder decides, on a case by case basis.

Here the AI works behind the counter. An internal assistant, whom the clerks question from the cash tablet, finds the sale by date, amount or item, says in which case of the table you are and prepares for the supplier the card with defect, photos and references of the sale. Repair, replace or refund is decided by who is at the counter.

Think of a household goods and small-appliance shop. A customer brings back a blender that no longer switches on and says she bought it last winter, paying by card. The sales assistant writes “blender, January, card” on the tablet; the assistant finds the sale in the management software, shows the sales document’s date, flags that the defect appeared within the first year and prepares a replacement request to the distributor. The sales assistant decides whether to give a new item from stock immediately.

Sort by size and color, not by model

In clothing and footwear the order of the season is usually done in advance, on the samples presented by agents and showrooms, and during the season you reassort what turns. For both decisions the total sold per model says little. It serves the sale by variant: which numbers ended in the first weeks, which colors remained whole at the end of the season, which sizes customers asked when they were not there.

The latter given the recorder does not see it: a missed sale does not produce a receipt. The messages instead see it: if the assistant records each "do you have the 42?" left unanswered, at the end of the season you have the list of lost requests, variant by variant. Placed next to the sold, it becomes an order proposal for the next season that you can take to the showroom and discuss with the agent. The decision remains yours; the forecasting methods are in AI for stock and inventory management.

The other useful job is the paper between supplier and store: order confirmations in PDF, transport documents, invoices. An automation reads them and compares them with what you ordered, as in AI for purchases and suppliers.

An example: a footwear store receives confirmation of the new season from the supplier. Automation compares it to the order signed in the showroom and signals that the 41 of the leather model has disappeared and that the 37 have doubled. When the goods arrive, he reads the transport document, loads the stocks and writes to the customers on the waiting list for that number, one at a time, with the time until which the pair stays aside.

Loyalty card: what you can do with the purchase history

The loyalty card collects just the data needed for a targeted recall: who buys what, how much, in which department. The reference to use them is the general provision of the Privacy Guarantor on fidelity card of 24 February 2005, written for large retail but also useful to a single store. It distinguishes three purposes: loyalty in the strict sense, i.e. points and discounts; profiling on consumption habits; direct marketing. For the latter two, a specific and distinct consent is needed, membership of the card cannot depend on that consent and the customer must have information and time to decide. The Guarantor cites the prompt subscription requested "while the customer is in line at the cash register".

The same provision puts a cap on the preservation of the detail of purchases of identifiable customers: twelve months for profiling and twenty-four for marketing, except for a real transformation into anonymous form. He adds that personal data should not be used if the purpose is achieved with anonymous data, and that sensitive data is not normally allowed for any of these purposes. It is a text written before the GDPR: times and legal bases should be reviewed with your privacy consultant, and the general picture is in GDPR and artificial intelligence in the company.

Within these boundaries AI works in two ways. For those who have given consent he writes the right call: the cream that buys back every autumn, the new collection of the brand that follows. For everyone else he works on aggregated data, without names: which departments grow, which items are bought together. In an herbalist or a health care professional the historian can say a lot about a person's health, and there profiling is to be excluded.

The birthday message with the discount is marketing. On WhatsApp it goes through a marketing category model, with the approval rules explained in WhatsApp Business templates, and leaves only to those who gave consent.

Telematics and POS recorder: what changes for those who automate the checkout

Paragraph 3 of Article 2 of Legislative Decree 127/2015, in the version applicable from 1 January 2026 as a result of the budget law 2025, requires that the instrument by which you accept electronic payments is always connected to the recorder, which records in a timely manner and transmits in aggregate form the fees and data of daily electronic payments. The connection is not a cable. According to the Statement of the Revenue Agency of 5 March 2026, is a combination that the operator, also through an intermediary, makes in the portal 'Invoices and fees' with the service 'Manage links', associating the registration number of the recorder with the payment instruments that result in his name.

For each new POS there is a precise window: the pairing is done from the sixth day of the second month following the one in which the tool is available, and within the last working day of that month. The example of the same communiqué: a POS that starts operating in March connects between 6 and 31 May. It is a rare deadline, so easy to forget: just a reminder related to the terminal's activation date.

The other sensible automation is the evening closure: a small stream compares the recorder total, POS receipts and cash counted, and only writes to you when there is a difference. For sums there is no need for a linguistic model; AI enters when there is to read a receipt, a poorly beaten return or the note of the clerk.

One note: in the version in force from 1 January 2027 those paragraphs of Article 2 are repealed by the VAT Consolidated Law (Legislative Decree 19 January 2026, No. 10), then ask your accountant where the rule from that date stands.

Where a store that starts with AI misses

• Let the assistant invent the exchange policy. Without your text, a linguistic model writes a plausible one, perhaps with a term you have never granted: it is the problem of AI Hallucinations.
• Use the exchange policy to deny the guarantee. "The goods on sale are not changed" can be your rule on the exchange. A lack of conformity, however, follows Article 133, which speaks of "any defect".
• Collect the consents of the loyalty card at the end of the checkout. It is the negative example chosen by the Guarantor: better a form to be completed calmly, with separate consents.
• Connect the assistant to an out-of-date warehouse. If the variants in management software are wrong, the AI repeats the error with more confidence: first the stocks, then the chat.
• Publish responses to reviews without reading them. AI prepares the draft and the holder approves it: in the demo for Reply to Google reviews, on negative reviews the system does not publish anything on its own and calls a person.

Where to start, shop by shop

Whatever the first flow, measure it before you touch it, for example by counting messages left unanswered at the end of the day. The picture for all retail is on the page AI for Retail. If you want to understand which of these flows connect first to the management software that you already use, tell us how your store works: Giallo Studio Build the assistant on the stocks, documents, and rules that you wrote.