An AI Store Assistant That Answers From Live Stock, With Staff on Hand

A multi-location retailer wanted shoppers to get real answers at any hour: is this item in stock at my store, when do you close, do you deliver, where is my order. A generic chatbot could not answer any of that, because the answers live in the store's own catalog, stock and opening hours. We built a customer chat for their Mage-OS and Hyvä store that answers from that live data, and puts the store team in the loop when a person is needed.

The challenge

The retailer sells across many stores, and the same product can be in stock at one store and only on display at another. Prices, hours, phone numbers and delivery rules differ by location too. Shoppers asked the same handful of questions again and again, and most of them came in when the stores were closed or busy.

What they needed was not a FAQ widget. They needed an assistant that:

knows which store the shopper picked and answers for that store only;
shows real products with the right price and stock status, and never promises stock that is not there;
keeps out of the way when the question is not about shopping (a complaint, a delivery question, a warranty claim);
hands the conversation to a store employee when the customer asks for a person;
never slows the storefront down, even when many people chat at once.

What we built

We built a complete chat system inside Magento, with a customer side, a staff side and a quality side.

For shoppers

A chat launcher on every page, with a short teaser bubble that matches the page ("Need help choosing?"). On a category page with an active promotion, the bubble alternates with the deal itself.
Answers with clickable product cards: image, price and stock status for the shopper's own store, plus Add to Cart (with the optional protection plan offered at the same moment).
Promotion quick answers: tapping the deal in the teaser opens the chat with a ready answer and the qualifying products. No AI call happens at click time, so it appears instantly.
Store answers that come straight from the store's own data: opening hours, phone number as tap-to-call, and a reminder to call now when same-day delivery is close to the cut-off.
Order status lookup after the customer confirms their order details, for web orders and for orders held in the retailer's ERP.
Chat history that survives: guests keep their past chats on the device, and logged-in customers see the same history on any device and can pick up an old conversation.
A callback request or warranty claim form inside the chat when the assistant cannot help, sent to the right store.

For store staff

A live chat monitor in the Magento admin: a badge shows how many customers are chatting at your stores right now, and managers only see their own locations.
Read-along transcripts, with the pages the visitor viewed and which product cards they clicked.
Live hand-off: a customer who asks for a real person is flagged, and an employee can offer a live chat. When the customer accepts, the AI pauses and the employee answers in the same window.
A short AI summary at the top of a chat (what the customer wants, what was discussed, next step), so nobody has to read a long transcript before calling back.
A grid of leads and claims, and a searchable history of every past conversation by store, customer or anything that was said.

For the business

An admin knowledge base of rules and answers (delivery, financing, assembly, payment, hiring) that staff edit without a developer.
Per-store rollout: chat can run at selected stores only, or for staff only, before it opens to everyone.
Reports on topics, answer quality and cost per conversation.

How it works

The shopper asks a question. The chat knows the current page and the store they picked.
Our code works out the intent first. Is it a product search, a store question, a delivery question, an order lookup, a complaint, or a request for a person? Requests for a person and order lookups are handled without the AI at all.
Our code gathers the facts. Products from the catalog search, with typo tolerance, so "haedboards" still finds headboards, plus the store's hours and phone, the matching knowledge-base rules and the shopper's earlier turns. Stock status comes with plain rules, for example "in store only" is never described as available to buy online.
Personal data is removed. Card numbers are blocked in the chat box, and phone numbers, emails and similar details are removed from the copy that goes to the AI provider.
The question joins a queue. The web request returns at once, and a background worker sends it to the language model, which only phrases an answer from the facts it was given.
The reply comes back with product cards only when the shopper was actually shopping, and with the store's phone number when a call is the best next step.
Staff can watch and step in. The conversation shows up in the live monitor for that store, and an employee can take over at any point.

Under the hood

Mage-OS (Magento 2) runs everything: the chat is a set of custom modules, with its own database tables for sessions, messages, leads, knowledge rules and reports.
Hyvä storefront: the launcher and chat panel are Alpine.js components, light enough to ship on every page and cached with the page. Time-sensitive parts such as a promotion's end date are checked in the browser, so a cached page never shows an expired deal.
OpenSearch powers the product search, with a fuzzy fallback and "did you mean" suggestions drawn from the catalog itself.
A hosted large language model writes the replies. We benchmarked several providers on real conversations and kept the one that stayed faithful to the facts it was given. The provider can be changed in configuration.
A job queue and a background worker handle the AI calls, so slow model responses never hold up the web servers.
Integrations with the retailer's ERP (order lookups and promotions), their store and warehouse data (stock, hours, phone numbers), their email marketing platform (an opt-in community sign-up inside the chat) and their site analytics (page trails in transcripts).
Cron jobs for nightly regression tests, weekly quality snapshots and keeping product summaries up to date.

What it changed

We only list numbers we measured ourselves; we don't quote revenue or sales results.

The storefront stays fast while people chat. In a load test of 100 simultaneous chats against a simulated slow AI, the storefront's 95th-percentile response time during the run was 168 ms, against a 182 ms baseline without chat (measured on the development copy).
Shorter, faster replies with the same facts. After we tuned reply length, a 10-question test set run through the real pipeline went from an average of 7.7 s to 5.2 s per reply (worst case 19.5 s to 9.0 s), and all 10 answers still contained the facts the test required.
Quality is watched, not hoped for. A nightly test run alerts the developers only when an answer that used to pass starts failing. A weekly report tracks how often customers have to ask the same thing again, by topic.

The business can now:

answer stock, hours and delivery questions for each store around the clock;
see live who is chatting at each location, and step in with a real person;
turn chats into callbacks and warranty claims that land with the right store;
change answers and rules from the admin, and roll chat out store by store.

Could this work for you?

It fits any Magento or Mage-OS store where the right answer depends on your own data rather than general knowledge:

retailers with several stores or warehouses, where stock, prices and hours differ by location;
stores with an ERP that holds orders, layaways or promotions customers ask about;
businesses that want an AI assistant and a way for real staff to take over a chat.

We build this kind of integration as custom work, usually on a Hyvä storefront. See our services, or, if your store is still on Magento 1 or an older Magento 2, start with a migration to Mage-OS and Hyvä. Then tell us about your store and what your customers keep asking.