Multilingual WhatsApp chatbots: a practical review guide

List the languages your team can maintain. For each, record who reviews policy text, who handles escalations and which templates exist. A model producing fluent text is not proof that the complete service works in that language.

Separate interface language, customer reply language and source-document language. A Hindi message may need an answer based on an English policy, with the policy meaning preserved.

A name, city or one-word greeting is weak evidence of a preferred language. Let the customer choose or continue in the language they used for the substantive request. Make switching simple.

Illustrative mixed-language request: "Order kal aayega?" A useful response checks the delivery record and answers the question. It should not invent an arrival date to avoid asking for clarification. Have a fluent reviewer assess any translated example before publishing it to customers.

Keep product codes, order references and payment amounts unchanged. Explain dates unambiguously. Translating a return policy must not turn "eligible for review" into "approved."

Use a terminology sheet with source term; approved translation; meaning; prohibited interpretation. Include cancellation, refund requested, payment pending and estimated delivery. These distinctions matter more than making every sentence sound polished.

The knowledge-base guide covers source ownership. Record which language version is authoritative when documents disagree.

When a customer requests a person, preserve their original message alongside any translated summary. Label the summary so staff know it is an interpretation. Do not silently replace the customer's words.

The handoff also needs to work in the language you offered. If nobody available can help in that language, explain the actual alternative instead of leaving the customer waiting for support your team cannot provide. Keep urgent support routes understandable without relying on another generated translation.

For each language, try a normal request, spelling variation, mixed-language message, correction, policy exception and staff request. Record whether meaning, amounts, dates, ownership and next action survived.

Ask reviewers to identify misunderstandings rather than rate fluency alone. A natural-sounding wrong answer is still wrong. Start with a narrow set of reviewed tasks and expand only when the team can maintain them. Do not describe arbitrary language detection or translated templates as verified platform features without checking the live setup.