AI in Localization: Where It Works, Where It Does Not, and What Buyers Should Ask

Every language vendor now says they use AI. That much is settled, and it is not the useful question. The useful question for any buyer is narrower and harder: where does AI actually help, where does it quietly fail, and what should you ask a vendor before you trust it with content that matters. After more than thirty years in this industry, my short answer is that AI has earned a real place in the work, and that the place has edges. Knowing where those edges are is most of the job.

Where it works

Machine translation has become genuinely good at a first pass. For high-volume, lower-risk content it moves faster and costs less than starting from a blank page, and paired with a well-maintained translation memory it keeps wording consistent across thousands of pages in ways a team of humans would struggle to match by hand. For internal documents, product catalogs, support content, and similar material where speed and scale matter more than nuance, AI is not a gimmick. It is the right tool, and a vendor who refuses to use it is leaving your money on the table.

Where it breaks

The trouble is that AI does not fail the way people expect. Everyone watches for the obvious error, the garbled sentence any reader would catch. That is not the dangerous one. The dangerous one is the sentence that reads perfectly and means the wrong thing. It is fluent, confident, and wrong, and because it looks finished, no one questions it. It sails through and lands in a consent form, a benefits notice, a label, or a contract, where the meaning is the whole point.

You have seen the version of this story outside our industry. Lawyers have filed briefs full of citations that an AI tool produced cleanly and convincingly, and that turned out not to exist. The lesson is not that the tool is useless. The lesson is that fluent and correct are not the same thing, and that the gap between them is invisible until someone qualified goes looking for it.

The failure is quiet

This is the part I would underline for any buyer. AI does not fail loudly. It fails quietly, in the one sentence nobody re-read, and the cost shows up later and somewhere else: a regulator’s question, a patient who misunderstood, a recall, a relationship that cools because the work was almost right. The quiet failure mode is exactly why the human step is not a nicety. It is the control that catches the error the machine is most likely to make.

What buyers should ask

So when a vendor tells you they use AI, the follow-up questions are simple, and they separate the serious from the rest. Ask who reviews the output before it reaches you, and whether that reviewer is qualified for your subject matter. Ask whether that review is a documented step or a line on a sales slide. Ask how they decide when to use AI and when not to, because the answers always and never are both wrong. Ask how they govern the machine-plus-human process, and whether it is certified. The international standard for that process, ISO 18587 (Machine Translation Post-Editing), exists precisely because the post-editing step is where quality is won or lost. A vendor who can point to it is telling you the human step is a system, not a hope.

Where we sit

For our part, Dynamic Language is not in the business of selling you a proprietary AI. We are in the business of helping you use it well. We advise clients on when AI fits a given content type and when it does not, we run a human-plus-AI workflow with qualified reviewers, and we hold ISO 18587 along with our other certifications so the governance is documented rather than described. That is the honest position for a company our size: a clear-eyed guide to the technology, not a vendor overclaiming a lab.

A healthcare client translating health benefit plans across several markets came to us with a fixed budget. Rather than run the whole set through one process, we engineered the files once and triaged the content inside them. The material carrying real liability, the parts a member could be harmed by getting wrong, went through full human translation with two linguists and a separate quality review. The lower-risk material, such as lists of products covered under a given plan, went through machine translation with human post-editing. The engineering investment happens once, and the savings repeat across every language in the program. Those savings covered two languages outside the original budget, Vietnamese and Korean, so the client’s members ended up with access in more languages than the budget was scoped for.

What it comes down to

AI is a tool that extends a good process. It does not replace one, and it punishes anyone who treats it as a shortcut around judgment. The companies that will earn trust with AI are not the ones with the most impressive demo. They are the ones who can show you, plainly, where the human still stands in the work.

Frequently asked questions