Blog
I ran an experiment last month. A client needed 70 pages of a Drupal site translated into a second language. Usually, this is a slow, expensive process involving export files, external translators, and a lot of manual copy-pasting.
I decided to do it with zero manual work. Just a custom AI agent and the Drupal API.
The translation itself was surprisingly fast. The AI handled the technical terminology and the tone of the site better than many junior human translators I’ve worked with. It didn’t just swap words; it understood the context of the page.
But it wasn’t all smooth sailing.
AI has a specific kind of “blindness” when it comes to interface elements. In one section, it translated the “Home” breadcrumb not as the destination (the homepage) but as the physical concept of a house. In another spot, it got a bit too creative with professional titles, turning a “Senior Lead Engineer” into something that sounded more like a character in a fantasy novel.
These are the kinds of errors that a human catches in a split second but an AI might miss because it is perfectly grammatically correct.
What I learned is that the quality of the translation isn’t about the model you use. It’s about the instructions you give it before it starts.
When I first ran the script, the error rate was about 15%. After I added a specific “glossary” of terms and a set of “forbidden translations” (like names of certain proprietary systems), the error rate dropped to near zero.
We still did a human review. But instead of the human doing the translation, they were just proofreading. It’s much faster to fix one “funny” word on a page than it is to write the whole page from scratch.
At ilf.studio, we use this AI-native approach to accelerate content migration and multilingual setups. It allows us to ship global sites in a fraction of the time.
If you have a large site that needs to go multilingual and you’re dreading the manual work, let’s talk.
ilf.studio — AI-native web studio, Gdansk, Poland.