Treat each language as its own capability. An agent can answer in a customer's language at depth 3 only where its certification included real cases in that language. Regulated wording, such as disclosures, fees and complaint rights, comes from approved translations, not live translation. Every reply is filed in both the customer's language and the source language.
What is multilingual support, and who owns it?
Customers write and call in the language they are most comfortable with, and many firms serve markets where several languages are common. A multilingual agent detects or asks for the customer's language, answers from sources that may exist only in one language, and hands over to a person who can continue in that language or with a translation.
Customer support owns the service and decides which languages are offered. Compliance owns any wording with legal effect and its approved translations. Language specialists, in-house or contracted, review quality. The agent's role is bounded by what it has proven in each language, the same principle that runs through our guide to the AI support agent that knows when to say no. For the short answer to whether this works at all, see can an AI agent give support in several languages?
What are the steps, and who acts at each one?
- Person, support lead: chooses the languages the agent may use, based on customer need and on whether reviewers exist for each one.
- Person, compliance: approves translations of regulated wording: disclosures, fee tables, complaint rights, key terms.
- Agent, depth 1: detects the customer's language and offers a choice where it is unclear. The customer's choice is recorded and used for the whole conversation.
- Agent, depth 1: retrieves the answer from approved sources, in whatever language they are held.
- Agent, depth 3: replies in the customer's language, using approved translations for any regulated passage and its own translation for ordinary explanation. This applies only in certified languages.
- Agent, depth 1: switches to transfer in languages where it is not certified, or where the customer's meaning is unclear after one clarifying question.
- Agent, depth 2: writes the handover note in the language of the receiving team, with the customer's original words attached.
- Person, adviser: continues the conversation, in the customer's language or through an approved interpreting route.
- Person, language reviewer: samples conversations in each language and scores them.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Languages | Nothing | none | Choosing languages |
| Translations | Nothing | none | Approving regulated wording |
| Detect | Identifies or asks language | 1 | Nothing |
| Retrieve | Finds approved answer | 1 | Nothing |
| Reply | Answers in that language | 3 | Uncertified languages |
| Transfer | Hands over when unsure | 1 | Nothing |
| Note | Writes handover with originals | 2 | Adviser continues |
| Review | Nothing | none | Sampling and scoring |
Which controls matter, and what evidence does each leave?
Certification on cases in each language. An agent certified on cases in one language is not certified in another. Attest scores each language on cases in that language. Evidence: the agent's certificate and its cases, by language.
Approved translations for regulated wording. A fee or a cooling-off right translated on the fly can change meaning. Evidence: translation version cited in each reply that contains regulated wording.
Both texts are filed. The record holds what the customer saw and the source-language equivalent, so a reviewer who does not speak the language can still check the substance. Evidence: paired texts in the append-only record.
Unclear meaning triggers a question, then a transfer. The agent does not guess at an ambiguous phrase. Evidence: clarifying turns and transfer records.
The binding language is stated. Where contracts say one language prevails, the agent says so when quoting terms. Evidence: the statement in the reply.
What should you measure?
- Resolution and transfer rates by language. A large gap between languages shows where the agent or the sources are weak.
- Reviewer scores by language, for accuracy, tone and clarity.
- Clarifying questions per conversation, by language.
- Complaints mentioning misunderstanding.
- Gaps in approved translations: regulated passages requested in a language with no approved version.
Where does it usually go wrong?
Fluency hides error. An answer can read naturally and still be wrong. Reviewers must check substance against the source, not only the language.
Formality is off. Some languages expect formal address with customers; a casual tone reads as rude. Set the register per language in the agent's brief.
Sources exist in one language only. If the help centre is written in one language, the agent translates every answer and errors compound. Translate the most-used articles properly first, and let the agent's own translations cover the long tail.
Code-switching confuses detection. Customers mix languages within a message. The agent should reply in the language the customer chose, not switch each turn.
Angry customers get literal translations. Idioms and frustration do not translate word for word. See can an AI agent deal with angry customers? for the escalation rules, which apply in every language.
Channels differ. Short messages on chat apps carry less context than email. The approach in answering customers on messaging apps with an AI agent matters more when language is a second barrier, and voice adds accent and recognition problems, covered in answering customer calls with a voice agent.
Where Pryme fits
Pryme certifies the agent on cases in each language it answers, so an agent at depth 3 answers in its certified languages and only reads and routes in others. Regulated wording comes from sealed, approved translations in the policy library, and every conversation is filed with both texts.
Questions people ask
Should the agent offer languages that no person on the team speaks?
Only if you have a route for handovers in that language, such as an interpreting service. Otherwise customers reach a dead end when the agent cannot help.
Who checks the quality of translations the agent makes itself?
Language reviewers, on a sample, scoring both accuracy and tone. Ordinary explanation can be translated by the agent; legally significant wording cannot.
Does certifying more languages take much longer?
Each language needs its own set of real cases and a reviewer to score them. Start with the languages most of your customers use and add others as evidence allows.