8 Jobs an LLM Can Do in Translation, and the One It Can’t

BEFORE TRANSLATION

Setting the model up to succeed

Garbage in, garbage out is the whole game in translation. A model can only work with the text it is handed, and in customer support that text is a mess by default. Agents type on phones between chats. “THX” stands in for thanks, “u” for you, “acct” for account. Typos slip through, product jargon runs next to shorthand that no dictionary lists, and the grammar bends whichever way gets the message out fastest.

Feed that raw into a translation engine and it breaks. “THX” comes back untranslated, or turns into a word no reader recognizes. One garbled sentence becomes the same garbled sentence across all 155 languages you support.

Companies with expertise like Language IO manage this before translation starts. Translation Optimization reads the shorthand, fixes the typos, expands the abbreviations, and settles the non-standard grammar into text a model can translate cleanly. Fix the source once and the error stops multiplying downstream.

Terminology: One term, one translation

Terminology is where translation quietly goes wrong, and in regulated industries it goes wrong with consequences. A medical device company had instructions that read “inject this into your body.” In one target language, “body” came back in the sense of a corpse. The sentence was fluent. It was also telling a patient to inject a dead body, and a reader skimming the translation would never have flagged the source.

That is the failure terminology control exists to catch. A word with more than one sense picks the wrong one, or the same concept gets two names across a document, and the output reads clean while meaning something the writer never said.

An LLM enforces a glossary at a scale no manual review reaches. Give it your approved terms and their target-language equivalents and it holds the line on every segment, so “body” carries the sense you intended and the term you settled on in review is the term that ships every time.

DURING TRANSLATION

Cleaning the source gets you a sentence the model can translate. It does not tell the model how the sentence should sound. That is the next decision, and it is where two companies with identical source text should end up with different translations.

Industry and brand voice: sounding like the company that wrote it

A translation can be accurate and still wrong. A medical device manufacturer and a mobile game studio might send the exact same sentence, “please confirm you want to continue,” and a model translating both the same way has failed one of them. The device maker needs the formal register a regulated audience expects. The game studio wants the casual voice its players would recognize. Same words in, different translations out, and the difference is brand and industry, not accuracy.

An LLM handles this through prompting tuned to a company’s voice and its field. Feed the model the terminology, the register, and the tone rules for a given brand, and the translation comes out sounding like that brand wrote it in the target language, rather than like a machine that happened to get the words right.

Tone and register: a choice the source never made

Most languages force a decision English leaves open. Address one person or a formal audience and English says “you” either way. French makes you pick tu or vous. German splits du and Sie. Spanish, Japanese, Korean, and dozens more encode the relationship between speaker and listener directly in the grammar, and there is no way to translate the sentence without choosing.

Get the choice wrong and the translation insults the reader while every word stays technically correct. A support reply that addresses a customer with the familiar form can read as presumptuous in one market and friendly in another. The formal form reads as respectful in one context and cold in the next. The right call depends on the brand, the audience, and the norms of the specific locale, and it can flip between two markets that speak the same language.

An LLM holds this consistently once it knows the rule. Tell the model that a brand addresses customers formally in German and warmly in Brazilian Portuguese, and it applies that decision on every segment instead of guessing per sentence. The register stops drifting, and a reply that opens formally does not slide into casual grammar three lines down

What is model routing in AI translation?

Here is the assumption worth killing: that you pick the best translation model and use it for everything. There is no best model. There is a best model for German legal text, another for casual Brazilian Portuguese, another for Japanese honorifics, and they are rarely the same model. Lock yourself to one and you are accepting its weak language pairs to keep its strong ones.

Quality holds up when each job goes to the model that performs best on it. That means testing models against real content in each language pair and routing the work accordingly, rather than betting the whole operation on one vendor’s average performance. A model that trails on one pair may lead on the next, so the routing is a standing decision that shifts as models improve and new ones arrive.

This is the part most teams cannot build on their own. Evaluating models across 155 languages, keeping the comparisons current as the field moves, and routing each segment to the right one is infrastructure, not a setting you toggle. It is also the reason a model-agnostic approach beats a single-model one on quality, not just on principle.

AFTER TRANSLATION

The model has produced a translation. Now the question is whether it is any good, and at the volume of customer support, nobody can read every segment to find out. This is where an LLM stops being the translator and becomes the reviewer.

LLM as a judge: grading the output at scale

Total Quality Assessment used to mean a linguist scoring translations by hand, rating adequacy and fluency segment by segment. Accurate, and impossible to do across thousands of tickets a day. A model trained to run that assessment does it at the speed and volume the work actually moves at, scoring every segment instead of a sample.

One rule is not optional. The judge cannot be the model that did the translation. Ask a model to grade its own work and it rates itself high, because the same judgment that produced the translation is the judgment being asked to check it. It approves its own reasoning by definition. A separate model, trained specifically to evaluate rather than to translate, catches what the first one was blind to. Language IO trains its judge models for exactly this, kept apart from the models doing the translation.

Automated post-editing: fixing the output before a human sees it

A raw translation is a first draft. It carries the meaning but reads a little stiff, or picks a phrasing a native speaker would not, or applies the style guide unevenly. Historically a human editor smoothed all of that, one segment at a time, which is where post-editing cost piles up.

A second model can take the first pass. It refines fluency, aligns the text to the style rules, and resolves the small awkwardnesses that separate a translated sentence from one that sounds native, before the content reaches a person. The human who does step in is now approving a near-final draft instead of rebuilding a rough one, which is a faster and cheaper place to start.

Confidence scoring: pointing humans at the segments that need them

The last job an LLM does here is decide what it is unsure about. A model can score its own confidence segment by segment and flag the low-confidence ones, so review stops meaning “read everything” and starts meaning “read the parts that are actually at risk.”

That changes the economics of quality. A linguist reviewing every segment does not scale past a certain volume. A linguist reviewing the flagged 5 percent does, and they spend their time where the risk lives instead of confirming translations that were fine. The machine handles the volume. The human handles the doubt.

The expertise that you can’t automate

Read back over the workflow and count the jobs an LLM did. It cleaned up the source. It held terminology in line. It set the register, routed each segment to the right model, judged the output, post-edited the rough spots, and flagged its own doubts. That is a lot of a translation workflow running on models, and it is real. You can get close to publishable quality with LLMs alone.

Close is the operative word. Every one of those steps involves the model deciding what something probably means, and now and then it decides wrong in a way that reads perfectly. That is the failure mode worth fearing, because a clumsy error announces itself and a fluent one does not. “Inject this into your body” came back grammatically flawless and told a patient to inject a corpse. No confidence score catches that. No second model reliably catches it either, because the same blind spot that produced it can pass it. A person who knows the language and the subject catches it.

So the human linguist does not disappear as the models improve. Their job changes. They stop translating segment by segment and start doing the two things a model cannot do for itself: calibrate the standard, deciding what good means for this brand in this language, and fact-check the confident output, reading the flagged and high-stakes segments to catch the fluent mistakes that survive every automated check.

But the linguist is one role inside a larger one. Look at everything the workflow required: rubrics for the judge model, glossaries the enforcement model applies, prompts tuned per brand and locale, and a routing layer that knows which model wins on German legal text and which wins on Japanese honorifics. That is not the work of a translator. It is the work of a provider fluent in both linguistics and AI, because each half is useless without the other. Linguists without AI engineering cannot build the system. AI engineers without linguistic expertise cannot tell a good translation from a fluent wrong one. The rare thing, the thing worth partnering for, is both in one place.

Which is why one model is never the answer, and neither is one discipline. The best translator is not the best judge, and the judge cannot be the translator. The best model for German is not the best for Japanese. A workflow this good depends on matching each job to the model built for it, building the standards the models enforce, and putting human judgment where the models are blind. Language IO runs the multi-model architecture and the linguistic expertise as one system, across 155 languages, so the models do what they are good at and the people catch what the models cannot see.