Twenty Languages, One Company: Where Machine Translation Holds and Where It Breaks
What machine translation handles well at scale, what still needs a native speaker, and how to decide which content gets which treatment.
Translating a 1,200-word page into twenty languages used to be a purchase order, a per-word rate and a two-week turnaround. Now it is a script that runs overnight. That change is real, and most companies have not adjusted their thinking to match it. They still treat translation as a budget line to be rationed, when the constraint has moved somewhere else entirely.
What the machines get right
Three things, reliably.
Volume, obviously. A model will produce a competent Japanese version of your service page in the time it takes to read this paragraph, and the twentieth language costs the same as the second.
Consistency, if you give it something to be consistent with. Feed a model a glossary — your product names, the terms you translate and the terms you never translate — and it will hold that vocabulary across hundreds of pages better than a rotating pool of freelancers ever did.
And, least appreciated, currency. The quiet failure in most multilingual companies is not a bad translation, it is an old one. The English page gets updated four times a year; the German page still describes last year's pricing because nobody wanted to pay to redo it. When retranslation is nearly free, you stop patching and simply regenerate from the source. Your twenty languages say the same thing because they were all produced from the same sentence yesterday.
Where a human is still not optional
Register is the first place it breaks. Formal or informal address is a business decision, not a linguistic one, and a model will make it for you inconsistently across languages and sometimes within a single page. Decide it once per market, write it down, and check it.
Then anything with legal or financial weight: contracts, warranties, safety instructions, regulatory claims, pricing terms. Not because the translation is likely wrong, but because being wrong is expensive and the review is cheap by comparison.
Then anything that becomes an identifier — a URL, a file name, a language code in a config file. These are not translation problems, they are engineering problems wearing a translation costume, and they produce the ugliest failures.
Finally, the copy that carries the brand: your homepage headline, your tagline, the two sentences a customer will remember. A model produces a defensible rendering. A native speaker produces one that sounds like a person wrote it on purpose.
The failures are not the ones you expect
Nobody in practice is undone by a clumsy sentence in Dutch. They are undone by the page that never shipped. A language code that drifts between two configuration files, and one locale silently produces no file at all — while the index page keeps linking to it. That is a live 404 in a market you cannot read, and it can sit there for months.
Related: models fail unevenly. One language will fail its translation pass every single run while nineteen succeed, and unless something checks for missing output, nobody notices, because nobody on your team reads that language. Build the check. Make the pipeline loud when a language comes back empty.
Tier the content, not the languages
The useful decision is not which languages deserve humans — it is which content does. A workable split:
- Human-reviewed, always: contracts, pricing, safety and regulatory text, the homepage, the top three commercial pages per market.
- Native-speaker skim: service descriptions, campaign copy, anything customer-facing that carries the brand voice.
- Publish machine output, correct on report: blog archives, support articles, documentation, changelogs. The alternative is not a better translation, it is no translation at all.
That third tier is where the economics actually pay. Four hundred support articles in twenty languages was never going to be approved as a project. Now it is a weekend.
The bill nobody budgets for
Publish in twenty languages and people will write to you in twenty languages. Inbound is the cost that surprises companies, and it lands on sales and support, not on marketing. Before you expand the language list, answer one question: when a customer replies in Korean, who reads it, and how fast. If the answer is nobody, you have not opened a market. You have built a very well-translated wall.
We run this site in twenty-one languages with exactly this arrangement — machine translation carrying the load, humans placed where being wrong costs something. It works. It works because the boundary between the two is written down, not improvised each time.
Do the same and translation stops being a budget question. It becomes a question of where you spend attention — which is the only scarce thing you had to begin with.