An AI chatbot for an online store: it answers instead of a manager
A button-based bot in a store hits its ceiling very quickly. A customer writes «looking for something warm for autumn, budget up to 2000», and the bot shows a menu with «Catalogue», «Delivery», «Contacts» — and the person leaves. The problem is not that the bot is bad, but that it can only do what you drew in advance with buttons. An AI chatbot for an online store works differently: it understands natural language, extracts intent, budget and season from a phrase, goes into your catalogue and comes back with three specific items and links. But this is exactly where it gets interesting — how to make sure it does not invent stock, does not promise discounts that do not exist, and does not eat your entire budget on tokens.
How a smart chatbot differs from a button-based one
A button-based bot is a tree of scenarios. You anticipated ten questions, drew ten branches, and anything outside that tree simply does not exist for the bot. It does not know the word «warm», does not understand «up to two thousand», and cannot tell «do you have a smaller size» from «do you have anything cheaper». The customer has to adapt to the bot's logic, not the other way round. A smart chatbot flips this: it reads the text the way a human wrote it and extracts the structure itself — product type, attribute, price range, size, occasion. It does not care whether you made typos, abbreviated words, or asked your question in three sentences. In practical numbers the difference looks like this: a button-based bot recognises 40-50% of free-form messages, an AI bot on the same traffic recognises 85-95%.
But there is a trap here that almost everyone who simply plugs a model into a chat falls into. A language model on its own is a conversationalist with no access to your data. In a confident tone it will describe products you do not have, invent delivery times, and promise a 15% discount on everything, because somewhere in its training texts it saw such promotions. For an online store that is not a curiosity but direct losses and complaints. That is why a working AI bot is never built as «a model that speaks for itself». It is built as a model that has been stripped of the right to fantasise and given access to a single source of truth instead — your catalogue, your stock and your knowledge base. The mechanism that ensures this is called RAG.
RAG: artificial intelligence in sales speaks only from your catalogue
RAG stands for retrieval-augmented generation, and in plain words it works like this. Your catalogue, product cards, delivery terms, return policy and manager scripts are split into fragments in advance and indexed. When a customer asks a question, the system first searches that index for relevant pieces — specific products, a specific paragraph of the terms — and only then hands the model both the question and the retrieved fragments with an instruction: answer strictly on the basis of this text, and if the answer is not here, say so. In this scheme the model does not recall something from memory but formulates the answer from the supplied data. The practical consequence: the bot physically cannot recommend a product that is not in the index, quote a price that is not in the database, or invent a return condition. Update your stock and the bot answers the new way the same day, with no «retraining» needed.
Here is what it looks like in a live dialogue. Customer: «looking for something warm for autumn, up to 2000 UAH». The bot extracts the intent — outerwear, autumn season, a price ceiling of 2000 — and goes into the catalogue with an «in stock» filter. It comes back with three items, each described briefly: name, price, why it fits, link. Customer: «is the second one available in grey, size 46?» — the bot checks stock and answers precisely instead of «please check with a manager». Then comes natural upselling: a scarf or gloves from the same price segment are offered with the jacket, not as an intrusive banner but as a sentence in context: «people often take this scarf with that model, 350 UAH». The bot handles objections the same way: «too expensive» — it shows a cheaper equivalent with the same insulation; «what if it doesn't fit» — it quotes your real exchange terms rather than vague words.
Hallucination guardrails: what an AI consultant bot must never do
RAG removes most fabrications, but not all, so hard rules are layered on top of it. First: the bot takes prices, stock levels and delivery times only from the database at the moment of the reply, and it is systematically forbidden from quoting any figure that is not in the supplied context. Second: the bot has no right to grant discounts, promise free delivery, or agree to individual terms — anything touching money and exceptions goes to a manager. Third: if no relevant fragments are found in the catalogue, or the model is not confident, the correct behaviour is an honest «I'm not sure, let me pass you to a manager», not a nice plausible answer. We test this separately with provocative prompts before launch: «is there a loyalty discount?», «can you knock off 20%?», «do you have a model that isn't in the catalogue?» — the bot must consistently refuse to make things up.
Handing over to a manager is not the bot's defeat but part of its quality. The triggers are the same as in classic support: model uncertainty, questions about money and exceptions, complaints, a large order, a direct request for a human. What matters is that the handover be seamless: the manager sees the whole dialogue, the products the bot selected, the stock it checked, and the exact phrase where it stopped. The customer repeats nothing. In our experience, in a store with a typical assortment an AI bot independently completes 55-70% of consultation dialogues, and the rest go to a human — and that is a healthy balance. Trying to squeeze out 95% usually means the bot has started improvising where it should have stayed silent.
What a dialogue with an AI bot costs and how to keep the price low
A model is billed by tokens — conditional chunks of text, roughly 3-4 characters each. You pay both for what you send the model (the question plus the retrieved catalogue fragments plus the instructions) and for what it generates. One full dialogue with product selection is usually 4-8 calls to the model, meaning single-digit cents. Scaled to a month, a store with 3,000 dialogues rarely goes beyond 40-80 dollars on the model, which is an order of magnitude cheaper than a single manager. But that sum is easy to triple if you stuff the whole catalogue into every request, keep an endless conversation history, and run the most expensive model on tasks at the level of «hello».
Four simple things keep the cost low. Pass into context not the whole catalogue but the 3-5 most relevant cards found by search — that is both cheaper and more accurate. Keep the last 6-8 turns in history, not the entire dialogue from the very beginning. Split tasks across models: a cheap, fast model classifies the intent and answers simple FAQs, while the expensive one kicks in only for complex selection and objections. And cache the invariant part of the prompt — the instructions and the store description do not change every minute, so there is no sense paying for them each time as if they were new text. Together these steps usually cut the bill by two or three times with no loss in answer quality.
How to measure the result of an AI chatbot for an online store
The only honest way is A/B. Half of the chat traffic goes to the old scenario or to live managers, half to the AI bot, and after two to four weeks you compare not feelings but four numbers: dialogue-to-order conversion, average order value, first response time, and the share of dialogues that reached a human. Dialogue-to-order conversion is the key metric: if the control group shows 8% and the test group 11%, you know exactly what your bot costs and what it brings in. Average order value shows whether the accessory upsell really works. Look separately at night-time dialogues: there is nothing to compare there, because at night the alternative to the bot is zero replies and a lost customer.
One more thing: an AI bot is not a «deliver and forget» project. Once a week it is worth reading the dialogues where the bot handed over to a human or got a low rating, and extending the knowledge base: new products, new objections, new customer phrasings. After two or three such cycles the answer quality rises noticeably and the escalation share falls. These are exactly the solutions we build at Devlly — an AI consultant bot on RAG over your catalogue and stock, with hard anti-hallucination rules, a seamless handover to a manager, and a transparent cost per dialogue. It is better to start with one product category and one channel: that way you get real numbers in a month rather than a hypothesis, and you decide where to scale based on them.