Hiring someone to build a chatbot that answers in several languages

The short answer

Making a chatbot reply in another language is the easy part; the AI models behind most chatbots already do it. What makes a multilingual bot a bigger job is everything around the reply: which language it chooses, what it answers from in each language, what it does with a language you do not support, and who takes over when a person is needed and does not speak that language.

Write those four decisions into the brief, and test each language separately before you pay. The chatbot brief has a “Tone and languages” field. This guide is about what to write in it.

Why “it speaks Spanish too” is not a spec

A seller’s demo will show the bot answering a question in English, then the same question in a second language, correctly. That proves the model can translate. It does not prove:

• the bot gives the same facts in both languages (the same price, the same return window)
• it answers from your documents in the second language rather than from general knowledge
• it stays in the customer’s language for the whole conversation, including after a handoff message or an error
• it refuses the same topics in every language
• the person who takes over can read the conversation

Each of these can fail without anyone noticing, because the people checking the bot often read only one of the languages.

There is a quieter layer underneath, too. Most of what my own content pipeline handles is not in English, and the failures I have had with non-English text were never loud. A watchdog of mine once reported an empty queue when one item was waiting: it looked for a phrase in a script’s non-English output, the text arrived in a different encoding, and the pattern simply failed to match. No error, just a wrong answer. A script saved in one encoding and read in another turned its non-English text into garbage and broke its quoting. A nightly job’s instructions were cut short at a double quote, which is why I now test anything that passes text around with quotes and non-English characters in it. A chatbot passes text through more places than any of my scripts: the widget, the AI provider, your documents, the handoff inbox, the logs.

The four decisions to make first

1. Which languages, and how good. List them, and for each one say whether it is full (answers, actions and handoff), answers only, or redirect (a polite message in that language pointing to a supported one). Name the regional variant where it matters: Brazilian or European Portuguese, Latin American or European Spanish.

2. How the bot picks the language. Options: reply in the language of the customer’s message; use the website’s language setting; ask at the start. Then decide the edge cases:

• a customer who writes in two languages in one message
• a one-word first message (“hola”, “price?”)
• a language you did not list

Ask what the bot does when it cannot tell. A bot that silently falls back to English is the multilingual version of a job that silently uses a default. In my pipeline, a missing setting now stops the run instead of quietly using the default, because quiet defaults are the failures nobody sees.

3. What it answers from in each language. If your help pages exist only in English, the bot will answer other languages by translating them on the fly. That is often fine for how-to answers. It is risky for prices, policies, warranties and anything legal, where a translation that is slightly off becomes a promise. Decide which documents need a reviewed translation, and say who updates each language when the English changes. Chatbots on your own documents covers the knowledge side in general.

4. Who takes over, in which language. If the bot hands a conversation to a person, that person needs to read it. Options: route by language to a speaker; translate the conversation for the agent and the agent’s reply back; or tell the customer, in their language, when a speaker will reply. Put the choice in the brief, because it changes the build.

What changes in the price

Every extra language adds work in a few predictable places, so expect the quote to move with:

• the number of full languages (each needs its own test set and review)
• translated knowledge that must stay in sync with the English source
• language-aware handoff to your team or helpdesk
• channels, since a website widget, WhatsApp and email each pass text differently
• usage cost, which the AI provider bills per amount of text, and some languages use more of it per sentence than English does. AI tokens and usage costs explains why

The chatbot job page lists the usual starting range and notes that model usage is billed separately and should be on your account.

The brief you can copy

Accept it language by language

The chatbot acceptance lines start with 30 real questions. For a multilingual bot, build that set once in English with the expected answers, then:

Test	Per language	Pass
Same facts	The 10 questions about prices, policies and deadlines	Same numbers and terms as English
Own documents	5 questions only your documents can answer	Answers from your material, no invented details
Never-do list	5 refused topics	Refused or handed over every time
Language switching	Start in one language, switch mid-conversation	Follows the customer, stays consistent
Unclear or unlisted	One-word message; an unsupported language	Asks, or gives the agreed redirect
Handoff	Ask for a person	A person receives it, readable, with history
Text survives	A name with accents, a non-Latin address	Unchanged in the reply, the logs and the handoff

Have a native or fluent speaker judge each language, and score each language on its own. A bot that passes in English and fails in Portuguese has failed in Portuguese. Testing a chatbot before accepting covers how to set a pass mark before you start, and chatbots that make things up covers what to do with invented answers.

Who to ask

Look for a seller who asks which languages and what happens at the handoff before quoting. On the chatbot page, Bensbahou is listed for bots that answer from your documents, and Luka Abramovic builds customer-support bots with a ticket inbox across channels, which is where language-aware handoff lives. If your customers are on WhatsApp, Hadj H. is listed for WhatsApp bots.

Wrap-up

A multilingual chatbot fails quietly: a price that drifts in translation, a fallback to English, a handoff nobody can read, a name mangled in the logs. Decide the languages and their level, how the bot picks one, what it answers from in each, and who takes over. Then test every language on its own with someone who reads it, before the final payment.