Multilingual AI and translation
The barrier to selling in six countries is rarely the product. It is that nobody in the building can answer a question in Italian on a Tuesday afternoon, so the enquiry waits, and by Thursday it has gone somewhere else.
Three different problems that get called “translation”
1. Conversation
Answering a customer in their language, right now, with your facts and your tone. This is where machine translation is at its best and where speed matters more than polish. A reply in ninety seconds in decent Italian beats a perfect one in two days, every time.
2. Content
Your site, your product descriptions, your documentation. Here quality compounds, because the text is read thousands of times and search engines judge it. This work runs in batches through models we host, then passes a review step where a human checks the pages that matter commercially. We have run this at the scale of thousands of pages across eleven languages.
3. Terminology
Every industry has words that must not be translated freely — a legal term of art, a medical procedure name, a product family, a certification. We build a glossary with you and enforce it mechanically, so the same term is rendered the same way in every language, every time.
The mistake we see most often
A site adds a language switcher that swaps text in the browser without changing the URL. Visitors see a flag and assume the site is multilingual. Search engines see one English page, index it once, and none of the translated content exists as far as they are concerned.
Done properly, each language gets its own URL, its own lang attribute, and reciprocal
hreflang links between versions. It is not glamorous work and it is the difference between
having six markets and having six flags.
We have made this mistake ourselves. An earlier version of this very site had a seven-language switcher that produced exactly zero indexable pages. It is a fast way to feel international and a slow way to find out you are not.
Where a human stays in the loop
Machine output is good enough to publish for informational pages. It is not good enough, unreviewed, for anything that carries legal weight, medical meaning or a price. We mark those categories at the start and route them to a reviewer. The point of automation here is to make human attention affordable by aiming it at the ten per cent that needs it.
Cost
Because the bulk work runs on models we host rather than a per-word API, the economics change: a large catalogue is a fixed compute cost rather than a bill that scales with your ambition. This is generally the moment a stalled localisation project becomes possible again.
How many markets are you leaving on the table?
Tell us which languages your customers write in and which ones you can answer. The gap is usually the cheapest growth available to you.