How does AI handle real-time translation in customer service?

Real-time translation lets a human agent speak in one language and the customer hear it in another within a one or two second delay. AI does the translation well at the technical level. The harder questions are about whether the cultural register survives the translation, whether the customer is told the conversation is being translated, and whether the company is on the hook for the AI's choice of words.

An English-speaking agent in Manila is handling a call from a German customer. Before AI, this required a German-speaking agent or a delayed translator on the line. With real-time translation, the agent speaks English, the customer hears German, and the conversation runs at near-normal pace. The technology works. The questions it raises are not all technical.

What people in the field are saying

Service Matters touches the orchestration of cross-language support in "Demystifying orchestration: the key...": handling languages used to be a routing decision (route to a speaker of the customer's language); with translation, it can be a conversation-layer decision.

What does real-time translation do well?

The literal translation, on common-language pairs (English-Spanish, English-French, English-German), to a level close to a fluent human translator. Latency low enough that the conversation feels natural. Vocabulary that handles routine support topics (orders, returns, accounts) well.

Where does it fall short?

Register: the polite-formal level of one language is not the polite-formal level of another, and direct translation can land too casual or too stiff. Idiom: language-specific phrases that do not translate literally. Sensitive topics: complaints, anger, regulated information; the translation may be technically correct and culturally off. Low-resource languages: quality drops sharply outside the common pairs.

What about disclosure?

The customer should know the conversation is being translated, especially because it changes who the company is responsible for. If the AI translates "I can offer you a credit" into the local language using a word that implies a different kind of compensation, the company is still responsible for what the customer heard. Disclosure upfront helps both sides.

Where does this go wrong?

The agent assumes the customer is hearing what the agent said. The customer assumes they are talking to a German-speaker. Both are technically wrong. When the conversation gets sensitive, the gap between what the agent meant and what the customer heard becomes a real problem the company has to manage.

What is the practical pattern?

Use real-time translation for routine contacts where the cost of misalignment is low. Route sensitive contacts (complaints, regulated topics, emotional cases) to a human who speaks the customer's language natively. Disclose the translation. Log both sides of the conversation (what the agent said, what the customer was told) for audit and for translation-quality review.