Most teams discover the cost of a bad translation after a customer reads it. A mistranslated refund policy, a support macro that promises something the English version never did, a status message that means the opposite of what was intended. By the time it surfaces, it has already shaped how a caller or chat user feels about the business.
We recently localized a 1,200-entry customer support knowledge base into five languages, and the project taught us more about translation reliability than any benchmark could. This is the full breakdown: the document we started with, the workflow we used, the errors we caught, and a checklist any support team can reuse. The goal is not to sell you on a method. It is to show you, concretely, where multilingual support content breaks and what catching those breaks actually requires.
The multilingual gap CX teams underestimate
When a company decides to support customers in more than one language, the conversation usually starts with staffing and channels. It rarely starts with the source material. Yet the knowledge base software, macros, IVR prompts, and canned responses are the backbone of every interaction, automated or human. If they are wrong in Spanish, every Spanish-speaking customer inherits that error.
The business case for getting this right is well established. CSA Research found that 75% of consumers are more likely to buy from a brand again when customer care is offered in their own language, a finding detailed in its Can’t Read, Won’t Buy study. The reverse is also true. Care that is technically translated but subtly wrong erodes trust faster than no translation at all, because the customer assumes the company meant what it said.
Our knowledge base covered billing, account access, shipping, returns, and escalation paths. Those are exactly the topics where a small wording shift changes meaning. “We may issue a refund” and “we will issue a refund” are one word apart and worlds apart in a complaint.
Why a single AI model is the wrong tool for a support knowledge base
The instinct is to pick the best available model and run the whole document through it. The problem is that there is no single best model. Performance varies by language pair, by domain, and by sentence. A model that handles French marketing copy beautifully may stumble on a German compliance clause or a Polish honorific.
This is not a fringe concern. Intento’s State of Translation Automation 2025 reported that baseline translation systems averaged 10 to 15 errors per text in testing, and that structured, multi-step approaches were what cut that figure to near zero. Synthesizing Intento and WMT24 findings, individual top-tier large language models fabricate or distort content somewhere between 10% and 18% of the time on translation tasks. For a 1,200-entry knowledge base, a 10% error rate is not an abstraction. It is roughly 120 entries shipping with a defect, and you will not know which 120.
There is a second, quieter failure mode: inconsistency. The same source sentence can produce meaningfully different output across sessions with a single model, which is where context handling tends to break down at scale. Translate “account holder” three different ways across a knowledge base and your customers, and your own agents, lose the thread.
The document, and what “done” had to mean
We started with a flat export: 1,200 entries, mixed length, mixed register. Some were two-sentence FAQ answers. Some were multi-paragraph policy explanations with formatting and inline variables that could not be allowed to break. The target languages were Spanish, French, German, Brazilian Portuguese, and Polish, chosen because they spanned both high-resource and morphologically complex cases.
“Done” could not mean “translated.” It had to mean three things at once: faithful meaning on high-stakes entries, consistent terminology across all 1,200, and preserved structure so nothing had to be manually rebuilt afterward. We used MachineTranslation.com, an AI translator, to run the workflow, because its SMART system is built around the exact problem we were trying to solve: instead of trusting one model, it compares the outputs of 22 AI models and selects the translation that most of them agree on.
Step by step: how we ran 1,200 entries through model consensus
Step 1. Segment before translating. We broke each entry into sentence-level segments rather than translating whole blocks. Segmenting keeps variables and formatting intact and makes the later review pass far more precise, because you are checking a sentence, not hunting through a paragraph.
Step 2. Run every segment through consensus, not a single model. Each segment went through the 22-model comparison. Rather than accepting one model’s guess, the system discards outlier renderings and keeps what the majority of models agree on. Across our internal benchmarks, this consensus approach reduces critical errors and hallucinations by roughly 90%, and brings the residual error rate to under 2%. That is the difference between proofing 120 suspect entries and proofing a handful.
Step 3. Read the quality score before accepting anything. Every segment carried a quality score, so the review queue sorted itself. High-confidence segments needed a glance. Low-confidence ones got flagged for a human. We did not have to read 1,200 entries with equal suspicion, which is what makes a project this size finishable.
Step 4. Lock terminology once, apply it everywhere. “Account,” “refund,” “escalation,” and a dozen other terms had to read identically across every entry and every language. We selected the preferred rendering for each key term and held it constant. In multi-document workflows, this kind of consensus-plus-terminology control held consistent terminology and register above 96%, against an industry baseline of roughly 78% for single-model output at the same volume.
Step 5. Escalate the genuinely ambiguous to human verification. A small set of entries, mostly legal-adjacent policy language, were too consequential to leave to any automated pass. Those went to in-platform human verification for a 100% accuracy guarantee. The point of consensus is not to remove humans. It is to make sure humans spend their time only where it actually matters.
Step 6. Reimport with structure intact. Because we segmented up front and preserved layout throughout, the localized entries dropped back into the knowledge base without a reformatting pass.
The errors a single model would have shipped
The most instructive part of any project like this is what gets caught. Polish was the clearest case. Single large language models plateau around 76% accuracy on Polish because of its morphological complexity, and several individual models produced grammatically confident but subtly wrong renderings of our account-access instructions. The consensus pass rejected those outliers and lifted Polish accuracy to 88% before any human touched it.
German surfaced a different problem. One model rendered a conditional refund clause as a definite promise, exactly the “may” versus “will” trap that turns a policy into a liability. It was not flagged by that model as low-confidence, because the model was not wrong about the grammar. It was wrong about the commitment. Only the disagreement between models exposed it. Across French, German, Italian, Spanish, and Portuguese, top single models cluster around 84% to 87% accuracy on this kind of content, while the consensus approach held 93% to 95%. Those percentage points are where customer-facing mistakes live.
A repeatable checklist for any multilingual support team
Whatever tooling you use, this is the workflow that survived contact with a real document:
- Segment at the sentence level before you translate anything.
- Never trust a single model on customer-facing content. Compare multiple outputs and look for agreement.
- Sort your review queue by confidence, not by entry order.
- Lock high-frequency terminology once and enforce it everywhere.
- Reserve human review for the genuinely high-stakes segments instead of spreading it thin.
- Preserve structure from the start so localization does not create a second formatting project.
What this means for multilingual phone and CX operations
A knowledge base is not the end of the chain. It is the source that feeds your agents, your scripts, and increasingly your automated systems. When you are automating voice operations or evaluating the voice AI tools built for multilingual customer conversations, the multilingual content underneath them determines the ceiling on quality. A voice agent answering in fluent Spanish is only as accurate as the policy text it was trained and grounded on.
The lesson from 1,200 entries is simple. Reliability in customer-facing translation does not come from finding one smarter model. It comes from refusing to let any single model have the last word. Catch the mistake before it reaches the customer, not after.
