Broekmans: AI Drinks Adviser for a Beverage E-commerce

An Adviser for 8,000 References

Drankenshop Broekmans is a Belgian drinks specialist based in Zolder, trading since 1989. Its webshop carries more than 8,000 references: whisky, rum, gin, wine, champagne, cocktails, and alcohol-free alternatives.

A catalogue that size creates a simple problem. A visitor who knows exactly which bottle they want finds it in a few clicks. A visitor looking for “a good whisky as a gift, around this budget” faces hundreds of listings and filters that do not speak their language.

We built an intelligent assistant embedded in the Broekmans site, designed to play the part of the in-store adviser.

Understand Before Recommending

The assistant does not answer the first question with a product list. It frames the need the way a shop assistant would: category, budget, flavour direction, and whether the bottle is a gift. One focused question at a time, never a questionnaire.

Customers describe taste in plain language: “more on the fruity side”, “not too peaty”, “something along the lines of a bottle I already enjoy”. The assistant turns that wording into search criteria against the catalogue’s tasting profiles. The customer never sees a numeric scale and is never asked to rate an aroma.

When the request is already specific, a reference, a brand, a factual question, the assistant answers directly instead of forcing a discovery interview.

Connected to the Real Catalogue

Broekmans runs its own PHP webshop rather than an off-the-shelf platform, so there was no connector to install. We built the integration against their catalogue directly. Two mechanisms carry the core of it:

1. Structured search runs over synced product metadata: category, country, brand, bottler, cask type, tags, and tasting profiles. This is what narrows 8,000 references down to a handful of credible candidates.
2. Live verification queries the product record before anything is recommended, to get the exact price, stock, availability, image, and URL.

The split is deliberate. Synced data exists to explore a large catalogue quickly, never to serve as proof of price or availability. No bottle is presented without being re-checked live, and both price and availability are resolved for the customer’s market, Belgium or the Netherlands.

A further layer completes the picture: semantic search across descriptions and tasting notes, used for the nuance structured metadata cannot capture, such as a request for something similar to a reference bottle.

Catalogue syncing is incremental. Commerce data refreshes without re-indexing the full set of product records.

Three Options, Explained

A completed recommendation normally presents three verified bottles as clickable cards inside the conversation: a safer or lower-priced fit, the strongest overall match, and a more premium or more adventurous option. Fewer are shown only when fewer genuinely fit, or when the customer asked about one exact bottle.

Each choice is justified from the criteria the customer supplied, trade-offs included when no bottle ticks every box. A stated budget is respected, and going over it is only ever offered as a clearly labelled option, never in place of one that fits.

Multilingual and Bounded

The Broekmans site is available in Dutch, French, and English. The assistant replies in the customer’s language, using natural Flemish Dutch for Dutch-speaking customers.

The same assistant on broekmans.be, broekmans.be/fr and broekmans.be/en: product cards in Dutch and French, the welcome screen in English.

Its scope is explicitly bounded, which matters especially for alcoholic drinks: no alcohol recommendations to a customer who indicates they may be under the legal purchase age, a shift to alcohol-free alternatives when driving, pregnancy, or medication comes up, and no artificial scarcity or urgency claims.

The assistant never invents a product, a price, a vintage, an ABV, an allergen, or a delivery promise. When something cannot be verified from the connected sources, it says so and points to the official contact route.

The In-Store Adviser, Online

What Broekmans gets out of it is the part of the physical shop that the webshop could not carry over: someone who asks what the bottle is for before pointing at a shelf. Every bottle the assistant puts in front of a customer has been checked against the live product record first, so a recommendation is one that is actually in stock, correctly priced, and available in that customer’s market.