Most AI models are trained primarily on English text. This isn’t a secret — it’s a consequence of the internet being roughly 60% English by content volume. When you use a cloud AI service in any other language, you’re using a model that learned your language as a second language. It works. But it works the way a fluent non-native speaker works: competently, sometimes awkwardly, occasionally missing the register entirely.

For casual use, this is fine. For business communication, it’s a problem.

The register problem

Every language has layers of formality that carry meaning. The difference between a formal greeting and a casual one isn’t just politeness — it signals the nature of the relationship, the seriousness of the communication, and the professionalism of the sender. An English-first AI model routinely misjudges these layers because it learned the rules of your language from translated text, not from native business correspondence.

This shows up in ways that are hard to catch if you’re not looking for them. The AI drafts a client email that’s technically correct but uses phrasing that feels slightly off — too casual for a legal notice, too stiff for a follow-up, too generic for an industry where specific terminology signals credibility. You fix it manually, every time, and after a while you stop noticing how much time you spend editing AI output that was supposed to save you time.

The problem compounds with regional variation. Business writing conventions differ not just between languages but between countries that share a language. A professional letter written for the German market doesn’t read the same as one written for the Swiss market or the Austrian market. Spanish in Mexico follows different conventions than Spanish in Spain. Portuguese in Brazil has a different register than Portuguese in Portugal. An English-first model treats these as the same language. Your clients don’t.

The multilingual generation gap

The newer generation of open-source models — particularly those trained with substantial multilingual corpora — handle this significantly better than their predecessors. Models like Qwen 2.5 were trained on diverse multilingual data from the ground up, not fine-tuned onto an English base as an afterthought. The difference is noticeable: more natural sentence structure, better grasp of formality conventions, fewer anglicisms leaking into professional text.

This matters because the quality gap between English and non-English AI output has been one of the biggest unspoken barriers to adoption outside the anglophone world. Businesses tried AI, found the output mediocre in their language, and concluded that AI wasn’t ready. It wasn’t that AI wasn’t ready — it was that the English-first models weren’t ready for their language. The new multilingual models are a different product entirely.

But model quality is only half the equation. The other half is customisation — and that’s where local deployment changes the game.

Your language, your voice

A cloud AI service gives you the same model as every other customer. You can tweak the system prompt, adjust the temperature, maybe save some custom instructions. But the underlying model is shared. Its sense of “professional German” or “formal Japanese” or “business Portuguese” is an average across millions of training examples. It writes generically because it was trained generically.

A local system can be configured differently. Your standard templates become reference material. Your preferred terminology gets reinforced. Your formatting conventions — how you structure an offer, how you close a letter, how you address different types of clients — become part of the system’s context. Over time, the AI doesn’t just write in your language. It writes in your voice. The kind your clients recognise. The kind that sounds like it came from your office, not from an algorithm doing its best impression.

This isn’t about perfection. It’s about the difference between output you can send directly and output you have to rewrite. Every minute spent editing AI-generated text is a minute the AI was supposed to save you. If the language quality isn’t there, the efficiency promise collapses.

Language isn’t a feature. It’s the entire interface between your AI and your clients. If that interface feels foreign, everything the AI produces feels foreign. Getting the language right isn’t polish — it’s the foundation.