An AI chat in your store that does not stop at the answer

The answer is only half of the conversation

A customer asking the chat about trail running shoes does not want a lecture on tread patterns. They want to see two or three specific models, their prices and whether they come in their size. If the assistant replies with a neat paragraph and three links, the customer does the rest of the work alone, and quite often simply does not do it.

That is why our assistant can show a recommended product as a card rather than a link. The card has a thumbnail, the price, availability and a button leading to the product page. It is the same information the customer would see on a product listing in the store, only given in the place where they are holding the conversation.

A price that tells the truth

An AI assistant works on a knowledge base that is refreshed periodically. That is enough for descriptions, manuals and terms of sale, but not for prices. A sale switched on in the morning does not always reach the knowledge base straight away, and a price printed in bold next to a photo that does not match the product page costs trust faster than no answer at all.

So the product card reads the price and the stock level from the store at the moment it is shown, not from the knowledge base. The same rule applies to the answers themselves: the assistant can ask the store for the current price and availability, search the catalogue by category, price range and availability, check which payment and delivery methods are switched on right now and which promotions are running. A delivery method switched off yesterday is no longer offered to anyone.

Two small situations that used to spoil conversations in practice have their own fixes. A customer asking about a product code gets that product, because the code is looked up exactly rather than by similarity of meaning, which simply does not work for a string of digits and letters. A customer on a product page who only asks how much it costs gets an answer about the product they are looking at, and if the chat has nothing to relate the question to, it asks instead of guessing.

The cart without leaving the conversation

The shortest way from a question to an order runs through the Add to cart button on a product card. The store itself adds the product, using the same mechanism as the button on the product page and in the customer’s own session, so the mini cart updates at once and the store shows its usual confirmation. Next to it sits Add to wishlist for those who are still making up their mind.

The button appears only where adding makes sense: for a product without options that has a price and is in stock. A product with variants leads to its own page, because adding it without a choice would put a variant in the cart the customer never picked. There is also a rule we held to from the start: the assistant never adds anything to the cart by itself. Every such action is the customer’s click.

An order you can see, not just read about

A large share of store conversations is about orders that have already been placed. The customer wants to know where the parcel is. A text reply saying the order has been shipped does not help much if they then have to dig the tracking number out of their email.

So an answer about an order can be followed by an order card: the status, the date it was placed and each shipment with its status and tracking number, plus a link to the order details for a signed-in customer. The card is built by the store from its own data, not by the assistant from its answer, and it is shown only for an order the customer is entitled to see in that conversation: their own when signed in, or one confirmed with a code. A signed-in customer also gets an Order again button on the card, which moves the order items to the cart.

A conversation that starts by itself

A customer who opens the chat often does not know what to ask. They see a greeting and an empty box, while the assistant knows the whole catalogue and the whole knowledge base of the store. The most common questions can therefore sit under the greeting as buttons. The answer to such a question can come instantly from an FAQ entry, without calling the model and without using a request from the package, because the answer to the most frequent questions is always the same.

After an answer about the offer, delivery, payment or an order, the chat can suggest two or three natural next questions in the language of the conversation. A customer who finds speaking easier than typing can dictate a question with the microphone button and correct the recognised text before sending it.

The style of the answers adapts to the store. Short answers of about two sentences read better on a phone, while lists of products or delivery methods still stay complete. A gift shop can switch on emoji, which never appear next to a price, an order status or in an answer about a problem with a parcel. The assistant also replies in the customer’s language, down to their regional variety, instead of switching to the formal literary standard.

Does the chat pay for itself

This question comes up at every subscription renewal, and until recently it had no good answer. The panel could tell how many conversations had taken place, but not what they were worth.

Now the panel shows how many conversations ended in a purchase and what those orders were worth. We count conservatively on purpose: an order is credited to the chat only when it is placed during the same visit as the conversation, and cancelled or declined orders drop out by themselves. A customer who asked in the evening and bought the next day is not counted. Such a figure is on the low side, but it can be defended when someone asks where it came from.

The other side of quality is ratings. Under every answer the customer can mark whether it helped. The most dangerous answers are not the ones where the assistant admits it does not know, because those land on the list of knowledge gaps. The most dangerous one is confident and wrong. Answers marked as unhelpful can be filtered in the conversation history and are listed in the periodic email report next to the unanswered questions.

Keeping store conversations safe

Customers type their email address into the chat when they check an order, and paste a bank account number when they ask about a refund. When the conversation is saved, that data is masked: an address keeps its first letters and the domain, a card or account number its last four characters. Product codes, EAN numbers and parcel numbers stay untouched.

A public chat also attracts people who test whether it can be made to reveal its instructions or change its behaviour. Such attempts are caught and collected on a separate screen in the panel. Detection blocks nothing, because a simple rule can also match an innocent question, and the real protection works independently of it.

Where to start

Each of the features described here is switched on separately in the add-on settings, so there is no need to change everything at once. A sensible order is product cards and starter questions first, because they change the first impression of the conversation, then the cart button and the order card, and after a few weeks a look at the number of conversations that ended in a purchase and at the answers marked as unhelpful.

The full description of the assistant, including the choice between the cloud and your own servers, is on the AI assistant page, and how the assistant learns from the questions it could not answer is covered in the article on unanswered questions. If you want to see it in your own store, get in touch.