Multilingual e-commerce support: the right answer needs more than a translation

A customer asks about shipping in German. Your chatbot replies in fluent German, gives the domestic delivery price and links to a Czech product page. The translation sounds right. The answer still sends the customer in the wrong direction.

Multilingual customer support has two jobs: understand the language and apply the right business facts. Breezaro can converse in multiple languages using your supplied knowledge. Preparing that knowledge, deciding which countries you serve and arranging human cover remain work for your team. Here is a practical way to approach it without rewriting the whole shop at once.

First separate language, destination and currency

English does not mean a UK delivery address. German does not necessarily mean Germany. A Czech-speaking customer may be ordering a gift to Vienna. The browser language is a useful clue for wording, but it does not establish where a parcel should go.

Before publishing an answer about delivery, make sure the relevant source distinguishes the destination country, delivery service, price and currency, dispatch time and estimated transit time. State any conditions, such as a minimum order value or an excluded product type, beside the rule they qualify.

For a Shopify or WooCommerce shop selling into several markets, compare that source with the actual checkout for each market. Add the relevant delivery rules to the chatbot's sources: settings in your shop administration are not automatically part of its knowledge base.

One question in three languages, one set of facts

Consider a fictional shop with these deliberately invented rules: it ships an in-stock item to Germany for EUR 6.90, normally dispatches within two working days, and quotes three to five working days in transit after dispatch. These are demonstration values, not Breezaro terms, carrier rates or a customer result.

The same question might arrive as:

For that fictional shop, these are sample target answers for a test, not captured chatbot responses:

Czech: „Doprava do Německa stojí 6,90 EUR. Zboží skladem obvykle odesíláme do dvou pracovních dnů. Po odeslání počítejte s přibližně třemi až pěti pracovními dny přepravy.“

English: “Shipping to Germany costs EUR 6.90. We normally dispatch in-stock items within two working days. After dispatch, allow approximately three to five working days for transit.”

German: „Der Versand nach Deutschland kostet 6,90 EUR. Lagernde Artikel versenden wir in der Regel innerhalb von zwei Werktagen. Nach dem Versand beträgt die Transportzeit voraussichtlich drei bis fünf Werktage.“

The factual checks are identical: Germany, euros, a dispatch period and a separate transit estimate. None of the answers promises a particular arrival date. Then remove “to Germany” from the question. If the destination is unknown, your desired behaviour is to ask for it before quoting a country-specific rate.

Give each important fact a maintained source

You do not necessarily need three separately written versions of every internal note. You do need clear information that the assistant can retrieve, and customer-facing pages people can understand when they follow a link.

Start with shipping, payments, returns procedures, availability and human support. For each topic, record the source URL or document, the markets it covers, the person responsible and the last review date. The FAQ template helps you organise the initial answers; the knowledge-base guide covers feeding them into Breezaro.

Watch for a common conflict: the English shipping page was updated yesterday, but an old German PDF still contains last year's rate. Refresh or remove obsolete material. Do not leave both versions in the knowledge base and expect the model to infer your intended policy from the language or filename.

When testing reveals a contradiction, fix the underlying source first. For an unresolved case, the recommended answer is to explain that the rate cannot currently be confirmed and offer a human check. That is an operating rule to establish and test, not a guarantee that every conflict will be detected automatically.

Check product links and prices as carefully as the wording

A translated product name is not a reliable identifier. Match the same product and variant across markets, then check the URL a visitor receives. Does it open the intended product? Is the relevant variant still available? Does the page preserve the market and currency the customer needs?

In a test sheet, an example such as shop.example/de/products/mug is only a placeholder. Replace it with a real product URL that you have opened. Do not ask the assistant to manufacture a translated path by adding /de/; multilingual stores do not all organise URLs the same way.

Keep currency explicit. A German reply can correctly quote CZK if that is how the shop charges. It must not silently convert that number to EUR or substitute a different market's price. Check decimal separators, tax labels and variant prices against the store's current customer-facing source. Localised wording cannot resolve inconsistent prices.

Live stock or a particular order's status is a separate issue. A downloaded feed or crawled page is a snapshot. Fresh data requires an appropriate configured integration. See the WooCommerce order-status example for that distinction; adding a language does not create an order integration.

Plan the handover in the languages your people actually cover

An assistant able to reply in German does not mean a German-speaking colleague is available around the clock. Write down who can handle each language, during which hours, and what happens outside them. Include this information in your support materials.

For a team that handles complex cases only in English, an illustrative notice could say: “Our support team handles detailed order enquiries in English during [hours and time zone].” Have the local-language version reviewed before publishing it, and replace the bracketed fields. Avoid promising an immediate native-language callback unless you provide that service.

Breezaro supports operator takeover, but handing over a conversation is not automatic translation of every human message. Rehearse the complete exchange, including a customer changing languages halfway through and asking about an existing order. Make sure the person receiving it can understand both the request and the answer already given.

Start with ten questions per market

For each target market, test the same ten useful questions in the languages your visitors use. Include an unspecified destination, an unavailable product, a contradictory source, a price question and a request for a person. Record the expected facts, actual answer and destination of every link. Ask a competent speaker to review the important answers.

Repeat these checks when shipping rules, prices or source pages change. The chatbot testing checklist gives you a repeatable format. Begin with one additional market and the sources you can keep accurate. Once the facts, links and human handover work together, multilingual support becomes useful beyond sounding fluent.