Multilingual customer support playbook for a small team and an AI agent

A multilingual customer support playbook a small team can run this week: decide per language, give the AI agent language rules, hand over honestly when nobody speaks the language, and test the same 10 questions in every language each month. Filled example, Word and Excel.

1. Language coverage matrix: decide per language

Fill first

This multilingual customer support playbook starts with one decision per language. The AI agent replies in 95+ languages, so every language gets an answer. The matrix decides which answers you stand behind, who checks them and what happens when they fail.

Three coverage levels

LevelWhat the customer getsWhat it needs
Supported: AI plus peopleAI answers, and a person who speaks the language takes handovers.A teammate who reads and writes it well, a reviewer, a passed test, a fallback.
AI onlyAI answers. A person replies with a translation tool, more slowly.A reviewer for about 1 hour a month, approved policy texts, a passed test, a fallback.
Not supportedAI answers simple questions. Binding terms come in main language, with an honest note.An honest message and a reply time in main language.

Matrix row (one per language)Language: language, with the region if it matters
Share of chats: % last month
AI answers: Yes / No
People who speak it: names, or none
Reviewer: native or fluent speaker
Fallback: what happens when quality fails or nobody speaks it
Status: Supported / AI only / Not supported
Owner and next action: name, action, by date

Decide the level

  1. Count real chatsShare per language over the last month. Also read 20 chats: what counts is the language people write in.
  2. Over 3% of chatsAt least AI only: find a reviewer and write the fallback.
  3. Someone speaks itSupported, if that person covers your support hours or has a backup.
  4. Under 3%, no reviewerNot supported for now. Look again every 3 months.

Completion checkA language counts as supported, AI only or AI plus people, only when it passed this month's test with a reviewer and has a working fallback. Until then, treat it as Not supported.

For small sales and support teams

Answer in 95+ languages and check every language you promise.

In CRMsoftware.pro the AI agent replies in the visitor's language from your Knowledge Base and hands chats to your team in Live Support. On Growth, Auto translate lets a teammate reply in a language they do not speak. Scoring answers is your monthly test, not an app feature.

  • AI replies in 95+ languages on every plan, from your Knowledge Base
  • Human handover in Live Support from Launch, Auto translate on Growth
  • Growth, $49 a month billed yearly for 20 users, adds the multilingual widget

Free plan: 3 users, AI chat and a basic knowledge base, no credit card.

Filled example

An online shop in 5 languages: coverage and the October test

Sagebrush Leather, a fictional online shop in San Diego, sells leather bags to the US, Mexico, Canada and Brazil. Daniela, the owner, speaks English and Spanish; Tom, in support, speaks English. In September, 41% of 610 chats were not in English. Here a customer in Brazil asks for a person in Portuguese.

Sagebrush LeatherFictional chat, translated from Portuguese
CustomerHi! I ordered the Ranger Tote 6 days ago and it has not arrived in São Paulo. Can I talk to a person?
AI agentI am sorry for the wait, Marina. Delivery to Brazil takes 8 to 15 business days, so your order is still on time. Our team works in English and Spanish and answers Portuguese messages with a translation tool, within 1 business day. Shall I pass you to them? Please send your order number.
CustomerYes, please. Order SL-4821. Will I have to pay import tax?
AI agentThank you. For Brazil, import duties and taxes are paid on delivery and set by customs, so I cannot tell you the amount. Tom from our team will reply here.
Handed to Tom in Live Support, Auto translate on
Tom (support)Hi Marina, Tom here. Order SL-4821 left our warehouse on 1 October and is now with Brazilian customs. Your tracking link is in the shipping email. I will check it again on 12 October and email you the status.

Language coverage matrix, October 2026

English, 59% of chatsSupported. Reviewer: Tom. Fallback: Tom in Live Support, 2 business hours.
Spanish, 24%Supported. Reviewer: Daniela. Fallback: Daniela, or Tom with Auto translate.
Portuguese (Brazil), 9%AI only. Reviewer: Lucas Almeida, freelance. Fallback: Tom with Auto translate, 1 business day.
French (Canada), 6%AI only. Reviewer: Amélie Roy, freelance. Fallback: Tom with Auto translate, 1 business day.
German, 2%Not supported. No reviewer. Fallback: honest message, email in English in 1 business day.

Monthly answer evaluation, 5 October 2026

LanguageScoreResult
English30 of 30 (100%)Pass
Spanish29 of 30 (97%)Pass
Portuguese28 of 30 (93%)Pass
French27 of 30 (90%)Fail: 1 policy error
GermanNot testedNo reviewer

What failed, the fix and the completion check

French, question 3Policy error: promised a free return label to Canada. Fix: a regional policy note in French, approved by Amélie.
French, questions 1 and 8Converted USD 189 to about CAD 259 by itself; used 'tu'. Fix: rules 5 and 7.
Spanish and PortugueseTranslated the name Ranger Tote; wrong delivery time to Brazil; European Portuguese wording. Fixed on 6 October.
Completion checkEnglish and Spanish count as Supported, Portuguese as AI only. French counts as Not supported until the retest on 12 October.

Why it matters

Why multilingual answers need a test in every language

More than 1 in 5 people age 5 and older in the United States, 22%, spoke a language other than English at home in 2017 to 2021; about 61% of them spoke Spanish1. An AI agent answers those chats whether or not anyone on the team can check the answer.

Fluent is not the same as correct, and quality is not the same in every language. When OpenAI machine-translated the same multiple-choice exam into other languages, GPT-4 scored 85.5% in English, 84.0% in Spanish and 62.0% in Telugu2. In a 2021 study of Google Translate for emergency department instructions, the meaning survived in 82.5% of 400 translations, from 94% in Spanish to 55% in Armenian, and the authors found the results inconsistent between languages3.

These numbers come from 2021 and 2023 tools and will change. The method will not: the same questions in every language, checked by someone who reads it, with a fallback for the languages that fail.

22%

of US residents age 5 and older spoke a language other than English at home, 2017 to 2021

85.5% vs 62.0%

GPT-4 on the same exam in English and translated into Telugu

94% vs 55%

meaning kept by Google Translate in Spanish and in Armenian, 2021 study

Excel evaluation

Multilingual answer evaluation template in Excel

Three sheets: Coverage matrix (one row per language), Evaluation (the example shop's October test: 4 languages, 10 questions each) and Summary. Shaded columns fill in by themselves. Opens in Excel, LibreOffice and Google Sheets.

ColumnWhat it does
Coverage: language, share of chats, AI answers, people who speak itOne row per language. The share comes from Chat History or Analytics.
Reviewer, fallback, fallback tried, statusWho reviews, what happens when quality fails, Yes once tried this month, and the level from a list.
Test result, check, counts as AutoThis month's result, what is missing, and what the language really counts as by the completion check.
Owner, next action, dueWho does what next for this language, by when.
Evaluation: language, question, expected facts, three checksOne row per language and question, with the reviewer's Yes or No and a note for every No.
Score Auto1 point per Yes, 0 to 3. Empty until all three checks and the reviewer are filled.
Summary: score, policy errors, result AutoPer language for the month you choose: Pass at 90% or more with no policy error. The threshold is editable.

The Summary sheet shows each language's score for the month, Pass or Fail against the threshold, and the languages missing a reviewer or a working fallback. In the sample, French scores 90% but fails on one policy error, and German has no reviewer.

How to use

How to use the multilingual customer support playbook

  1. 1

    Fill the coverage matrix

    Share of chats, people who speak it, a reviewer, a fallback and a level per language (tab 1).

  2. 2

    Set the greeting and rules

    First message from tab 2, language rules from tab 8, approved texts and the terminology list in the Knowledge Base.

  3. 3

    Agree the handover

    Who replies in which language, by when (tab 3). Try one handover per language.

  4. 4

    Run the monthly test

    The same 10 questions in every language, scored by the reviewer in the Excel file.

  5. 5

    Fix and decide

    Fix the source, retest failures within 7 days and update each level.

Set it up in CRMsoftware.pro

  1. Find your languages: Chat History shows a Language under Visitor Info for each chat, and Analytics, where your plan includes it, shows Language Distribution under Demographics & Traffic.
  2. My Bots > choose the bot > Basics tab: paste the language rules into Instructions (Shared tab), add approved texts and the terminology list to the Knowledge Base, write the First Message. Save Bot.
  3. Set Widget Language for buttons and labels, or turn on Auto-detect website language; the language you pick is the fallback.
  4. Run the 10 questions in each language with Try Bot, with the reviewer typing.
  5. Handovers arrive in Live Support as Awaiting agent. On Growth, turn on Auto translate to read the customer's messages in your language and reply in theirs; Show original shows their own words.
  6. Tag chats by language: Settings > Conversation Tags > New tag, for example lang-fr. On Growth, save localized replies: Settings > Saved Replies > New reply, Shortcut /refund-fr.

CRMsoftware.pro does not score answers or route chats by language: the matrix and the test are your documents. AI chat and the Knowledge Base: Free, 3 users. Human handover: Launch, $19 a month billed yearly for 8 users. Auto translate, saved replies and the 17-language widget: Growth, $49.

Monthly language review, 30 minutes

  • Every language marked Supported or AI only passed this month, with a named reviewer.
  • Every fallback was tried once this month.
  • Every No in the test has a fix, an owner and a retest date.

FAQ

Questions about multilingual customer support

What is a multilingual customer support playbook?

A short operating plan for serving customers in more than one language: which languages you support and how, what the agent and the team say in each, and how you check quality. This one adds a monthly test with the same 10 questions in every language.

What are multilingual customer support best practices for a small team?

Decide per language instead of promising to be multilingual. Use approved translations for policies and prices, keep product names untranslated, say honestly which language a person will reply in, and test every language monthly with a native or fluent reviewer.

How do you evaluate a multilingual chatbot's answers?

Ask the same questions in every language, with the expected facts written down from your source. A native or fluent reviewer scores each answer for correct facts, tone and policy. In this template, a language passes at 90% or more with no policy error; test monthly and after any change to prices or policies.

Is there a multilingual answer evaluation template in Excel?

Yes. The free Excel file has a coverage matrix, an evaluation sheet with a score formula per language and question, and a Summary with Pass or Fail per language and the languages missing a reviewer or a working fallback.

Can I machine-translate my refund policy and terms?

Only as a draft. A fluent reviewer checks it against the source and the owner approves it before the agent quotes it. In a 2021 study, Google Translate kept the meaning of emergency instructions in 94% of Spanish translations but only 55% of Armenian ones3.

What should the chatbot say when nobody speaks the customer's language?

Before the handover, name the language your team works in, the language of the reply and the time, for example: "Our team works in English and Spanish, so a person will reply in English within 1 business day." Never promise a person who speaks their language.

Which languages should a small business support?

Count a month of chats by language. Languages above about 3% get at least AI only, with a reviewer and a fallback. A language someone on the team speaks can be Supported. The rest stay Not supported, with an honest message.

Does CRMsoftware.pro answer in the customer's language?

Yes. The AI agent replies in the language the visitor writes in, in 95+ languages, on every plan. For the widget's buttons, Auto-detect website language matches your website's language, with Widget Language as the fallback; the 17-language widget comes with Growth.

Can my team reply in a language they do not speak in CRMsoftware.pro?

Yes, on Growth, $49 a month billed yearly: turn on Auto translate in Live Support to read the customer's messages translated and reply in their language; Show original shows their own words. For refunds and legal terms, still paste the approved translation.

Answer every language. Promise only the ones you test.

Start free with 3 users: AI chat in 95+ languages and a basic knowledge base. Launch, $19 a month billed yearly for 8 users, adds human handover. Growth, $49, adds Auto translate, saved replies and the 17-language widget.