Multilingual Customer Support Chatbot: 8 Practical Tests

1. Question language versus website language

Leave the website interface in Finnish, then ask a question in English or Swedish. A multilingual assistant should address the actual question rather than repeating a canned Finnish answer. Evaluate language detection, the displayed interface and the final response separately.

2. Ask the same service question three ways

Try “Vaihdatteko jarrupalat?”, “Byter ni bromsbelägg?” and “Do you replace brake pads?”. If the auto shop offers that service, the factual answer should agree across all three languages. If the service is unconfirmed, the bot should admit uncertainty instead of inventing a yes.

3. Keep one approved source of truth

Prices, opening hours and service boundaries must survive translation. If the business supplies a tyre-change price, its Swedish and English responses must preserve the amount and conditions. Offering tyre replacement does not prove that the garage also offers tyre storage.

4. Preserve context across follow-up questions

Ask “How much is a service?” followed by “Entä kauanko se kestää?”. The bot should understand the same service is being discussed, but must not invent its duration when only the price is documented. Try short availability follow-ups without connecting a real calendar.

5. Test country-specific ecommerce policies

Ask an online store for shipping to Finland and then to Sweden. Free domestic shipping does not imply free international shipping. Check delivery restrictions and returns, and ensure a visitor never sees another customer's order or tracking details.

6. Include typos, abbreviations and mixed languages

Real customers write “can u book huolto?”, “byt däck?” or “paljo maksaa”. A useful bot can understand common wording or request clarification. A confident but unsupported price or appointment is worse than asking one sensible follow-up.

7. Validate handoff and customer-data separation

Ask for an unlisted repair price and check that the bot offers a contact request or a human rather than guessing. For multi-business platforms, run similar questions for two different companies; neither a new chat nor an older follow-up may expose the other company's products or contact details.

8. Build a repeatable scorecard

Record the question, language, approved source fact, expected answer, actual response and one outcome: correct, clarification needed or incorrect. Fix invented prices, unsupported services, broken links and failed human handoffs first. Rerun the cases after each significant knowledge-base update.

Frequently asked questions