Yes, language by language. An agent can support customers in each language it has been tested in on your own cases, answering from approved sources. Regulated wording, such as key terms, complaint responses and legal notices, should use approved translations rather than the agent's own. Where meaning is unclear, or the language is not certified, the agent hands over.
Why is each language a separate decision?
Because an agent's quality varies by language, and you cannot assume it from the language you tested. An agent that handles refund requests well in English may misread a polite indirect complaint in Japanese, or miss the urgency in a short message in Portuguese. Sarcasm, formality and idiom all shift.
So treat each language as its own capability. Test it on real conversations in that language, scored by someone fluent, before it goes live. The step-by-step method is in how to answer customers in several languages with an AI agent.
Where do the answers come from?
Most businesses keep policies and help content in one language. An agent answering in another is translating as it goes, and every translation risks drift: "may" becomes "will", a condition drops out, a fee becomes optional.
For everyday answers, that risk is manageable with testing. For wording that carries legal or regulatory weight, it is not. Contract terms, fee disclosures, complaint outcomes and data protection notices should come from translations the business has approved, and if none exists, the agent should say so and offer a person or the text in the original language.
What may the agent do, and what stays with a person?
| Situation | The agent | A person |
|---|---|---|
| Everyday question in a certified language | Answers from approved sources | Samples |
| Regulated wording, approved translation exists | Uses the approved text | Maintains translations |
| Regulated wording, no approved translation | Says so; offers original text or a person | Answers or arranges translation |
| Uncertain meaning or mixed languages | Asks one clarifying question, then hands over | Takes the conversation |
| Language not certified | Sends an approved message offering supported languages or interpretation | Arranges an interpreter |
| Complaint or vulnerability in any language | Hands over with a summary in both languages | Responds |
What does this look like in practice?
Take an international money transfer firm serving customers in English, French and Spanish. The agent is certified in all three, and its role covers fee disputes in English only.
A customer writes in Spanish asking where a transfer is. The agent answers in Spanish from the payment record. The same customer then disputes a fee. Fee disputes in Spanish are outside its role, so the agent says, in Spanish, that a colleague will handle the dispute and gives the expected response time. It hands over with a summary in English and the original Spanish messages attached. A second customer writes in a language the firm does not support. The agent recognises it, replies with an approved short message in that language offering English, French or Spanish, or a call with an interpreter.
Which controls make it safe?
- Certification on cases in each language, scored by fluent reviewers.
- Approved translations for regulated wording, held in the policy library with version numbers.
- Handover summaries that keep the customer's original words.
- Escalation and complaint rates tracked by language. A language with unusually low escalation may be one where the agent misses signals; see how high an AI support agent's escalation rate should be.
- No switching language mid-conversation unless the customer does.
Where Pryme fits
In Pryme each language is covered by the agent's certification, by replaying Attest against your own conversations in that language. An agent at depth 3 can answer transfer status in Spanish while its role leaves fee disputes in Spanish to people, and the minimum chain applies the lower bound in each case. The wider approach is in the guide to AI customer support agents.
Questions people ask
Does multilingual support mean fewer support staff?
It usually changes what staff do more than how many there are. The question is covered in whether an AI support agent will replace human support staff.
Should customers be told the reply was machine-translated?
They should know they are dealing with an AI agent. Many businesses also say that a person is available in the customer's language, or through an interpreter.
Can the agent translate documents customers send?
It can produce a working translation for staff. Decisions based on a document should rest on the original or a certified translation.