I run a small localization agency. About 1.2 million words a month across maybe 15 languages. Started using AI tools about two years ago, not because I wanted to but because clients started asking and the agencies that said yes were undercutting us on price.
So we adopted AI. Reluctantly. And the results were not what either the cheerleaders or the skeptics predicted. More boring than that, and more complicated.
Technical content
User manuals. Software strings. Knowledge base articles behind a login. AI does this well. We took a SaaS client's 40,000 UI strings, English into eight languages, and the AI output with post-editing was indistinguishable from human-only in blind quality review. About 35% lower cost. Way faster turnaround. No user complaints.
We now handle basically all technical content this way. Though it took about three months of glossary and style guide tuning before quality reached this level. The first month was rough.
Verdict: 35-40% of volume. Genuinely better than human-only for this content type.
Marketing content
Where everything gets complicated. Fashion brand, seasonal campaign, email sequences and social ads into German, French, Japanese. English source was well written — conversational but not sloppy. The AI translations were grammatically fine across all three. But the tone was identical across all three languages. Which is exactly wrong.
German fashion marketing has a different register than French. Japanese has a completely different sensibility. The AI was imposing a universal "translated marketing" tone. Our human translators restructured paragraphs, replaced analogies, rewrote subject lines entirely. The AI version would have passed quality review. It wouldn't have performed.
Verdict: ~25% of volume. AI should not touch this.
Legal and financial content
Template-based contracts: AI mostly handles fine. But anything involving judgment — contract negotiations, regulatory commentary, financial analysis where the author is making an argument — the AI flattens the rhetorical structure.
We translated quarterly earnings commentaries into Mandarin. The AI preserved all the numbers but killed the argument. Our financial translator said the AI version "told you what happened but didn't make you think the CFO was smart for doing it."
Verdict: Mixed. Templates yes, argumentative content no.
What we actually do now
Before any project, a senior linguist sorts content into three buckets:
1. "AI can handle this" — 35-40% of volume. AI with glossary enforcement, light post-editing.
2. "AI should not touch this" — ~25%. Taglines, creative copy, anything where emotional impact is the point. Straight to human translators.
3. "AI drafts, human really edits" — 35-40%. AI first draft, human does actual rewriting. Less than from-scratch, more than proofreading.
The sorting step is probably the most valuable thing we do. It's not complicated. It's just someone with experience knowing what AI mangles and what it handles.
The honest version
The vendors want you to believe AI is almost as good as humans. The anti-AI crowd wants you to believe it's useless. The truth: AI is good at some things, bad at others, and telling which is which requires human judgment. The translators who adapted well were already good translators. AI didn't make bad translators good. It made good translators faster on the mechanical parts.
At Artlangs Translation, we've been doing this since before it was called human-in-the-loop. The workflow evolved but the principle didn't: the right content goes to the right process, and a qualified human always makes the final call on anything a reader is going to have an opinion about. Across 230+ languages.
