Products don't break in new markets because of language. They break because the assumptions don't travel.
Soy milk becoming gelato in São Paulo was a case of assumptions not fitting the market — so the shape had to change. The same thing happens in software localization. But most companies are still trying to fix it with better translation.
The bottleneck isn't translation quality
NMT, LLM-based translation, post-editing workflows — the tooling keeps improving. The bottleneck doesn't move.
Because the bottleneck isn't translation quality. It's the process design.
When you translate, the source text becomes untouchable. Every decision gets evaluated against the original. That's why you need a human to post-edit — not to catch errors, but to reconcile two different assumptions about how the world works.
The alternative
Don't translate. Write.
Give the LLM the original intent, not the original text. Add the right guardrails — cultural context, tone, market assumptions. Let it write from scratch in each language.
Human involvement shifts from reconciling translations to setting the cultural frame upfront — or doing a final check. The loop gets shorter. The assumptions travel better.
The idea is older than the tooling
Localization has always been redesign, not translation. The tooling is finally catching up to that idea.