AI chat in your store: a shopping assistant and a back office assistant, with limits

Artificial intelligence assistants have gone from curiosity to tool in a very short time. In an online store there are two places where they can work: front of house, helping customers find and decide, and behind the scenes, helping the team look things up and manage the store without clicking through ten screens. They are different uses, with different risks.

The assistant for customers

Where it helps

• Finding the product when the customer does not know what it is called: “something to hang bikes in the garage without drilling”.
• Specific availability: “do you have the M in blue?”, “is there any left in 128 GB?”. It is the question that sells, and many search boxes and filters do not answer it well.
• Alternatives when the requested item is out of stock: another variant of the same product or a similar one from the same category.
• Exploring the catalogue: new arrivals, best sellers, categories, price ranges.

Where it should not go

• Promises it cannot keep: exact delivery times, discounts that do not exist, compatibility that is not on the product page.
• Professional advice: health, safety, use of hazardous products. The prudent approach is for it to refer to the product’s official information or to a person.
• Complex complaints and returns: an angry customer needs someone from your team, not a language model.
• Other people’s order data: an assistant that can look up orders must do so only for the identified customer and only their own orders.

The technical key is for the assistant to answer with your store’s data, looked up at that moment (product, combination, stock, price), and not with what the model “knows” or assumes. A model that answers from memory makes things up; one that queries the store’s tools and is told not to invent does so far less.

The assistant for the back office

Here the user is your team. The value lies in saving navigation and filtering:

• “How much did we sell yesterday?”, “what is the average order value this month?”
• “Which products are at critical stock levels?”
• “Who are the customers who spend the most, and which have not bought for months?”
• “Show me this customer’s last order.”
• “Apply 10% off this category until Sunday.”

Queries carry little risk. Actions do: changing stock, creating discounts or disabling products from a sentence typed in a hurry can go wrong. A typical example: someone types “set the T-shirt stock to zero” thinking of one particular size, and the assistant takes it to mean every T-shirt in the catalogue. That is why a back office assistant must:

1. Ask for explicit confirmation before any change, showing exactly what it is going to do.
2. Respect the employee’s permissions: not do through chat what that person could not do from the menu.
3. Keep a log of every request and every action carried out, per employee.

Data: what leaves your store

An AI assistant sends text to a model provider (OpenAI, Google, Anthropic or others). It is worth knowing what is sent and under what conditions:

• Use your own API key under a business agreement with the provider, not a personal account. Check in their terms whether they use the data for training and where they process it.
• Minimise: answering about stock does not require sending customer names or addresses.
• State in your privacy policy that you use a third-party AI service in the chat, and treat it as a data processor.
• Do not ask for sensitive data in the customer chat: no cards, no documents, no health data.
• Keep spending under control: every conversation uses tokens. A usage log per employee or per day avoids surprises on the provider’s invoice.

How to start without risk

1. Back office first, in query-only mode. Your team learns what it does well and what it does not, without exposing anything to customers.
2. Actions with confirmation, once you trust the queries.
3. Customer chat on trial, for example only for identified customers or on a copy of the store, reviewing real conversations.
4. Your own instructions: tone, what not to say, when to hand over to a person, how to talk about delivery times.
5. Regular review of conversations: the questions the assistant cannot answer are valuable information for your product pages.

Current assistants require recent versions of PHP and PrestaShop. If your store is several versions behind, that is usually the first obstacle, and one more reason to keep it up to date.

What to measure

• Conversations that end in a cart or a visit to a product page, against those that end without a useful answer.
• Unresolved questions: group them by topic. If many are about shipping times or sizes, the information is missing from the store, not from the chat.
• Hand-overs to a person and the reason for them.
• Time saved in the back office: which queries the team makes through chat that used to take several screens.
• Cost per conversation and per month, to decide whether it is worth extending its use.

What it will not do

An AI chat will not fix a catalogue with poor product pages or stock that is not kept up to date: if the store’s data is wrong, the assistant will answer wrongly, and with great confidence. Nor does it replace human service where judgement is needed. Set up properly, it saves your team time and clears up the doubt that was stopping the customer from buying.

How to solve it in PrestaShop

Rekire PrestaChat adds a chat to the back office that queries and manages the store: stock, sales and orders, customers, temporary discounts and product activation. It works with your OpenRouter, OpenAI, Gemini or Anthropic key, with fallback models if one fails, takes into account the screen you have open, asks for confirmation in the chat before any change and logs requests, tokens and actions per employee. It can also enable a shopping assistant in the store, only for identified customers if you prefer. The Disponibilidad y Alternativas add-on teaches it to answer for a specific combination (size, colour, capacity) with units and price, and to suggest other variants or up to five similar products in stock when the requested item is sold out.