Multilingual and Real-Time Translation Bots
Assistants that work properly in each language you support, with quality measured language by language and a native reviewer signing off before any of them goes live.
Adding languages to an assistant looks like a configuration change and behaves like a new product per language. Retrieval quality differs, model fluency differs, your content probably only exists in one language, and the failure mode is that a customer receives a confident answer in Portuguese that your team cannot read to check.
Multilingual conversational AI is the design and engineering of an assistant that operates across several languages: detecting the user's language, retrieving from content that may exist in only one of them, generating a natural reply in the user's language, and reporting quality separately for each language rather than as an average.
Three ways to do this, and when each is right
| Approach | How it works | Best for |
|---|---|---|
| Translate in and out | User message translated to your base language, answer translated back | Quick coverage of long-tail languages |
| Native generation over translated retrieval | Retrieve in the base language, generate directly in the user's language | Most deployments; the usual default |
| Fully native | Content, retrieval and generation all in the target language | Priority markets with real content investment |
Most organisations end up with a mixture: fully native for the two or three markets that matter commercially, native generation for the next tier, and translation for the tail. Deciding that deliberately is cheaper than discovering it after a complaint.
What multilingual actually requires
Language detection that does not embarrass you
Detected per message rather than set once per session, because people switch mid-conversation, quote text in another language and write in a mixture. Where confidence is low the assistant asks rather than assumes, which is far better than answering in the wrong language.
Retrieval across a language boundary
If your content exists only in English, a Spanish question must still find the right passage. That is done with multilingual embeddings, query translation or both, and it is measured per language because recall varies considerably between them. The retrieval foundation is described under RAG system development.
Terminology that stays put
Product names, legal terms, scheme names and anything with a regulated meaning belong in a glossary that is enforced rather than left to the model. Translated product names are among the most common and most avoidable failures in multilingual deployments.
Native review before launch
A native speaker who understands the domain reviews a sample of real outputs per language and signs off before that language goes live. This is not a formality: fluent-sounding output can be subtly wrong in ways an automated score does not capture, and your team will not catch it in production.
Register and formality
Languages that distinguish formal and informal address need an explicit decision, applied consistently. Getting this wrong reads as rude or as oddly familiar, and neither is what you intended when the English version sounded perfectly friendly.
The practical details
Right-to-left layout, date, number and currency formats, name ordering, character limits in messaging platforms, and fonts that render every script you support. These are unglamorous and they are what makes an assistant feel local rather than translated.
Report quality per language, never as an average
An average across nine languages hides the two that are failing. We measure and report each language separately, with a defined bar a language must clear before it launches and a defined trigger for taking one back offline.
Real-time translation between people
A related but distinct use is translating live between a customer and an agent who do not share a language. It has its own requirements:
- Low latency, because the conversation is happening now and delay breaks the rhythm.
- Terminology enforcement, so product and policy terms stay consistent in both directions.
- Visible original text alongside the translation, so the agent can see what was actually said.
- A record of both languages for audit, dispute and quality review.
- A clear disclosure to the customer that translation is machine-assisted.
This extends the reach of an existing support team without hiring per language, and it pairs naturally with contact centre AI.
How the engagement runs
Each language earns its launch by clearing a measured bar with a native reviewer's sign-off.
Language strategy and approach
Markets prioritised, approach chosen per language, glossary and formality decisions agreed with local stakeholders.
Retrieval and generation
Cross-language retrieval built and measured per language, generation configured with glossary enforcement.
Evaluation per language
Real questions per language with agreed answers, scored, with native reviewers assessing a sample.
Localisation details and pilot
Formats, layout, character limits and fonts handled, then a pilot per language with transcripts reviewed natively.
Staged launch and handover
Languages released as they clear the bar, with per-language dashboards and the process for adding the next one.
What you receive
Each language launched on evidence, with the means to add the next one properly.
Language strategy
Approach per language with the reasoning, and the bar each must clear before launch.
Cross-language retrieval
Multilingual retrieval measured per language against real questions.
Enforced glossary
Product, legal and scheme terminology fixed per language, applied at generation.
Per-language evaluation
Scored sets with native reviewer assessment, wired in as a release gate.
Localisation handling
Formats, layout, register and platform constraints applied per locale.
Per-language dashboard
Resolution, escalation and satisfaction reported separately, never averaged.
Is this the right engagement?
Worth being direct. Multilingual and Real-Time Translation Bots is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- You serve customers in several languages and support quality is uneven between them.
- Content exists mainly in one language and translating all of it is not realistic.
- Native reviewers can be made available for the languages that matter.
- Terminology consistency matters for legal or product reasons.
- You would rather launch three languages properly than nine approximately.
Choose something else if
- There is no native reviewer for any language, so quality cannot be verified.
- The base-language assistant does not work yet. Fix that first.
- Regulated advice must be given in each language, where machine generation may not be appropriate.
- Volume per language is tiny and human handling is simpler.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Can one chatbot support many languages?
Yes, and the practical design is usually native generation over retrieval in your base language, with fully native handling reserved for priority markets. What matters is that quality is measured per language rather than averaged, because an average hides the languages that are failing.
Do we need to translate all our content?
Usually not. Cross-language retrieval can find the right passage in your base language and generate the answer in the user's, which covers most cases. Translating content properly is worth doing for priority markets where the answer quality and the brand both matter.
How do you stop product names being translated?
With an enforced glossary applied at generation rather than a request in the prompt. Translated product and scheme names are among the most common failures in multilingual deployments and among the easiest to prevent.
How do you know the quality is acceptable in a language nobody internally speaks?
You do not, without a native reviewer, which is why sign-off from one is a precondition for launching a language. Automated scores catch structural failures but not the subtly wrong or oddly impolite output that damages trust.
Can it translate live between a customer and our agent?
Yes, with low latency, enforced terminology, the original text visible to the agent, both languages recorded for audit and a clear disclosure to the customer. It is a practical way to extend an existing team's reach without hiring per language.
Often paired with this
Most clients combine two or three engagements from the Conversational AI pillar. These are the ones that most often run immediately before or after.
AI Chatbot Development
A chatbot grounded in your own content, measured on resolution and satisfaction rather than on messages handled.
Read more →Speech-to-Text and Text-to-Speech
Recognition and synthesis tuned on your own audio, measured on your vocabulary rather than on clean benchmarks.
Read more →Contact Centre AI and Call Analytics
Every conversation analysed: automated QA, live agent assist, wrap-up and the reasons behind your contact volume.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.