Serving visitors in several languages is one of the most concrete advantages an AI robot offers over a scripted machine and over a small staff. It is also more work to set up correctly than the marketing material suggests.
Where multilingual capability genuinely pays
Tourism and hospitality. Hotels, attractions and transport hubs receive visitors of many nationalities, and staff who speak a given language are not always on shift. This is the clearest case.
Healthcare. Where a language barrier affects care, even basic navigational and procedural help in a visitor's language has real value.
Public services. Foreign residents dealing with unfamiliar procedures.
Trade and exhibitions. International visitors at events.
Education. International students and language practice.
Where it matters less. Local retail and services with an overwhelmingly local customer base. Adding six languages there consumes setup effort for very little return.
The practical filter: look at your actual visitor mix before deciding how many languages to support. Most deployments need two or three well rather than eight badly.
How language selection actually works
Three approaches with different trade-offs.
Manual selection. The visitor picks a flag or a language name on screen. Reliable, obvious, requires one extra interaction. Still the most robust option.
Automatic detection from speech. The robot identifies the language from the first utterance. Convenient when it works; unreliable in noise, with short utterances, and with mixed-language speech.
Default plus switch. Start in the local language, with an obvious way to change. A sensible compromise for most deployments.
What causes problems. Mixed-language speech, which is very common — a sentence in one language containing product names or place names in another. Detection systems handle this poorly.
Practical recommendation. Manual selection displayed prominently, with automatic detection as a convenience rather than the primary mechanism. And make switching mid-conversation possible, because people discover the option after starting.
Content in multiple languages
The part that determines quality and the part most often underestimated.
Machine translation of your content works, imperfectly. It handles general information adequately and struggles with brand terms, product names, local place names and anything culturally specific.
What needs human review. Anything customers act on — prices, procedures, opening hours, safety information. Machine-translated procedural text can be subtly wrong in ways that matter.
Brand and product names should not be translated. Specify them as fixed terms so they survive translation intact.
Tone differs across languages. A formality level appropriate in one language may read as cold or overly familiar in another. Worth having a native speaker check the greeting and closing at minimum.
Maintenance multiplies. Every content update must propagate to every language. This is the recurring cost that makes eight languages expensive and three languages manageable.
Practical approach. Full human review for the thirty most common questions in each language; machine translation with spot checks for the rest.
Quality varies substantially between languages
A fact rarely stated in product material.
Recognition quality differs. Systems perform best on languages with the most training data, and the gap between the best-supported and least-supported languages a product claims is often large.
Speech synthesis quality differs. Some languages sound natural; others sound noticeably synthetic, which affects the impression considerably.
Understanding quality differs. Handling of colloquial phrasing, regional variation and mixed-language input is uneven.
How to check. Test each language you intend to support, with a native speaker, before committing. Do not assume that supported means well supported.
What to do about weak languages. Rely more on the screen for those languages, keep the content simpler, and set expectations accordingly rather than offering a poor experience.
And ask the supplier directly. Which languages are strongest, and what the difference is. A supplier who answers honestly is telling you something useful; one who claims uniform quality across a long list is not.
Setting it up well
A practical sequence.
One: check your visitor mix. Actual data if available, staff observation if not. Choose two or three languages beyond the local one.
Two: test each chosen language with a native speaker before finalising the product choice.
Three: prepare the core content properly in each. The thirty most common questions, human-reviewed.
Four: fix brand and product terms so they are not translated.
Five: make language selection prominent and switching easy.
Six: check the greeting and closing with a native speaker. These are the lines every visitor hears.
Seven: build the update process for all languages at once. Otherwise the secondary languages drift out of date within months, which is worse than not offering them.
That last point is where multilingual deployments usually fail — not at launch but six months later, when the primary language is current and the others are stale.
Frequently asked questions
How many languages should a deployment support?
Usually two or three beyond the local one, chosen from actual visitor mix. Most deployments benefit more from three languages done well than eight done badly, because maintenance multiplies with each one.
Is automatic language detection reliable?
Not reliably enough to be the primary mechanism. It struggles in noise, with short utterances and with mixed-language speech, which is very common. Prominent manual selection with detection as a convenience works better.
What content needs human review rather than machine translation?
Anything customers act on — prices, procedures, opening hours, safety information — plus brand and product names, which should be fixed as untranslated terms.
Where do multilingual deployments usually fail?
Six months after launch, when the primary language is current and the secondary ones have gone stale. Building an update process covering all languages at once is what prevents it.
More in Local ecosystem and Robots in Vietnamese business.
