A growth-stage manufacturer lands distribution partners in five European markets in the same quarter. It is the kind of milestone that gets celebrated in a Monday meeting. Then the contract lands on the operations team’s desk, and it has to exist in French, German, Spanish, Italian, and Polish before anyone signs. More than that, it has to say exactly the same thing in all five. Same obligations, same payment terms, same liability caps, same termination clause.
This is the moment market expansion quietly turns into a language problem. And it is the moment most teams reach for whatever AI tool is already open in a browser tab, paste in the contract, and hope.
We want to walk through what we did instead, step by step, because the workflow matters more than the tool, and because the failure mode here is specific and expensive.
Why one AI model is a business risk, not a technical detail
Here is the part that does not show up in a product demo. Today’s top language models fabricate or distort content somewhere between 10% and 18% of the time on language-heavy work, according to data synthesized from Intento’s State of Translation Automation 2025 and the WMT24 benchmarks. That is not a bug waiting for a patch. It is a structural property of how a single model generates text.
For most content, a 12% error rate is an annoyance you clean up later. In a distribution contract, it is a liability you sign your name to. One mistranslated clause does not read as “slightly off.” It reads as a different obligation. The cost is not theoretical: a single misplaced comma in a contract once cost the aerospace manufacturer Lockheed roughly $70 million, a story Lokalise has documented in its own research on AI translation quality.
This is no longer a fringe concern, because the tooling is already everywhere. Roughly 70% of language workflows are now machine-assisted, and in the finance sector alone, AI translation adoption rose 700% between 2023 and 2024. The question facing any company expanding across borders is not whether to use AI for this. The same wave of AI tools is already reshaping operations everywhere else in the business. The real question, and the one that belongs in a serious market-entry strategy, is how to use AI without inheriting its error rate on the one document you cannot afford to get wrong.
The five-step workflow we actually ran
We stopped treating “translate the contract” as a single action. We treated it as a verification process with five steps.
- Map the risk surface before touching the language. We marked the clauses where a wrong word changes a legal obligation: liability, indemnity, payment timing, termination, governing law. Everything else is lower-stakes. This tells you where to spend your verification budget.
- Stop relying on one model’s opinion. Instead of running the contract through a single tool, we ran each segment through com, an AI translator which compares the outputs of 22 AI models and selects the translation that most of them agree on. The logic is simple. Hallucinations are idiosyncratic. A given model invents a given error. It is unlikely that a majority of 22 independent models invent the same error in the same place. So the majority rendering is, by construction, the safer one.
- Read the disagreement, not just the output. This is the step teams skip, and it is the most useful one. When models split on a segment, that split is a flag. It is the system telling you exactly which sentences to look at.
- Take the majority rendering, isolate the outliers. For the bulk of the document, where the models converged, the majority output was production-ready and we moved on. For the handful of segments where they did not converge, we held them back.
- Send the few high-stakes outliers to a human. The clauses we flagged in step one, plus the segments the models disagreed on in step three, went to a professional linguist for final sign-off inside the same platform. That is a few dozen sentences getting human attention, not a few thousand.
What the model disagreement told us
The disagreement was not noise. It was a map of where single-model translation would have quietly failed.
In our internal testing on complex multilingual legal contracts, the individual models broke in different, predictable ways. One model showed a 12% error rate handling honorifics in Asian languages. Another hallucinated numerical dates in Romance languages, the kind of error that turns a March deadline into a different month. A third failed to hold the formal register that German corporate filings require. Any one of those, shipped alone, is a clause that means something other than what the original said.
Run through the cross-checking workflow, the effective error rate on that same dataset dropped to near zero. Across our benchmarks, the consensus approach reduces critical errors to under 2% and cuts overall error risk by roughly 90% compared with trusting a single model. It also holds terminology steady where single models drift: consistent terminology and register at a rate above 96% across multi-document work, against an industry baseline near 78%. For European languages specifically, where top single models plateau around 84% to 87% on French, German, and Spanish and fall to about 76% on Polish, the cross-checked approach held 93% to 95% across Western and Southern Europe and lifted Polish to 88%.
“The mistake companies make is asking which AI model is the smartest,” says Ofer Tirosh, CEO of Tomedes. “The smartest single model still produces the same error rate as any other single model of its size. What you can actually trust is the rendering most of them agree on, with a human on the few that carry real risk. That is a different question, and it is the one that protects you when you sign.”
What any leader can take from this
You do not need to localize a five-country contract to use this. The principle generalizes to any high-stakes document leaving your business in another language: a vendor agreement, a compliance filing, a customer-facing policy, an investor update.
The lesson is that the trustworthy output is rarely the one a single tool hands you with confidence. It is the rendering that survives being checked against 21 other attempts, with human judgment reserved for the clauses where being wrong is expensive. Disagreement between models is not a flaw to be hidden. It is the most honest quality signal you have, and it tells you precisely where to look before you commit. Expansion will always turn into a language problem at some point. The teams that handle it well are the ones that stop hoping a single answer is right and start verifying it.
