Every platform now claims to translate. Almost none can tell you what happens when it goes wrong.

Don’t be an account ho….lder

An agent at a streaming company typed a few words to a customer: “Just to confirm, am I talking to account ho.” The agent was cut off mid-word. “Account holder” never finished. The translation engine did not flag the typo. It rendered it, faithfully and fluently, as a German word for prostitute.

The customer screenshotted the exchange and posted it to an audience of 250,000. Management heard about it within hours. Nothing malfunctioned. The engine did exactly what it was built to do. That is the problem.

The claim everyone makes and no one defines

Walk any CX trade show floor this year and the pitch is identical. Every platform supports 100+ languages. Look closely and most are a language model or a standard translation engine with a glossary attached. Ask what happens after the translation leaves the system and you get a feature list, not an answer.

Supporting a language means text goes in and text comes out. It says nothing about whether the customer understood, whether the message sounded like your brand, or whether it was respectful in a culture where tone is part of the content.

Why the pilot passed

Default tools win pilots for a simple reason. The pilot is small, the content is generic, and the people running it work in customer success, not localization. Nobody pounds on the messages that break: the misspelling, the slang, the acronym, the half-typed sentence.

Then volume arrives. Language IO customers push millions of chats through translation every week, and real support conversations are full of exactly the input a pilot never tests. At that scale, the one-off error is a weekly event.

Failure mode one: the engine that translates everything literally

Most platforms ship a neural machine translation (NMT) engine as their default. It is fast, cheap and good at ordinary sentences. It also has one defining trait: it translates the words, not the situation. A cut-off “account holder” gets rendered as exactly what it says.

That is how the streaming incident happened. It is also how one dropped letter turns the Spanish word for “loading” into a word no brand wants in a status message. In another case, an enterprise software customer escalated after a routine question about “child org units” came back from one engine with the Spanish word for orgy. The glossary entry for that term was correct. The error came anyway.

A glossary is a list. It is not a quality system.

Failure mode two: the engine that answers instead of translating

The obvious upgrade is a large language model. Give it the whole conversation and it would likely recognize a cut-off word for what it is. That is a real advantage.

But language models fail differently. Ask one to translate “Are you a robot?” into Spanish and you may get back “Sí, soy un robot.” It did not translate the question. It answered it.

Your customer just received a statement your brand never approved, from a system nobody was watching. In our testing, larger models were sometimes more prone to this, not less. The more capable the model, the more it wants to help. We saw it early, so our prompts carry an explicit rule against it. Ask any vendor whether theirs does.

The stakes are legal as well as reputational. In February 2024, a tribunal held a major international airline liable for what its chatbot told a customer. A company owns what its AI says in its name. Translated text is included.

Right words, wrong respect

A Japanese celebrity paid roughly 700,000 yen for a business class seat and was placed in economy. When he reached out to support, the agent ended the exchange with a flippant “see ya.” He posted it. The story spread. The company that began by offering about 35,000 yen in points ended up refunding roughly 1 million.

The root failure was service. The accelerant was tone. In Japanese, honorifics and register carry weight that English never asks of a sign-off. A system with no instructions on how to address your customers will choose a register for you. It will not tell you which one it chose.

The wrong answer that looks right

A global travel brand ran its first German quality review after launch and found “transit time” coming back as “flight time” on some segments of an itinerary and not others. Same phrase, same trip, different results. No alert fired. A traveler read layover as time in the air and planned around it.

Errors like this never announce themselves. They show up as repeat contacts and customers who stop believing what you tell them.

What catches it

We built Toxicity Shield because the streaming incident was not rare. It also protects agent from abusive customers, but the main reason it exists is to protect customers from offensive AI translations, whichever engine sits underneath.

It is one layer of several. Behind the engine sits a routing framework that picks a model based on the transaction: a short chat or a long email, clean text or heavy HTML. A second model judges every output in real time. If a translation falls short, it fails over to another model. That is written into our SLAs with enterprise customers.

Tone, honorifics and terminology are set by instruction, not left to chance. And because prompts, glossaries and brand calibration are not locked into any single model, a model change does not break them. Prompts drift. They need retesting per language pair with every new model release, and that work is heavier than most teams expect. A model that is good in general is not calibrated for your brand, your languages or your customers. The engine is one layer. The controls around it are the product.

The question no pilot asks

Picture this. Your Korean customers start taking offense at the tone of your translated support messages. Your glossary needs recalibrating for English to Korean. Who do you call?

A CRM vendor is not in the computational linguistics business. Default tools generally do not let you tune prompts per language pair and per direction, and nobody on the vendor’s side is paid to find out why your German customers feel talked down to. The savings are real on day one. The support is what you discover you did not buy.

The math nobody puts in the business case

Every example above looks like a rounding error when it happens once. Enterprise support does not happen once.

At 99% accuracy, a company translating a million messages ships 10,000 that are wrong. The question that decides your quarter is not how many. It is which ones. A stiff phrase is a footnote. A slur in front of 250,000 followers is a news cycle.

The cost lands in places a per-word price never shows:

• Agent time spent clarifying, apologizing and escalating
• Repeat contacts from customers who did not understand the first answer
• Churn that arrives with no ticket attached
• Brand damage that arrives as a screenshot
• Liability for everything your AI says on your behalf

When a CFO sees translation bundled into a platform the company already pays for, the savings
are visible on day one. The exposure is not visible until something goes wrong. By then, the
savings are gone.

Six questions to ask before you sign

If a vendor cannot answer all six, they are not selling you quality. They are selling you coverage. Language IO was built for the gap. We hold 99% accuracy across languages and build the controls around the engine that catch these failures before a customer does. It runs inside the tools your teams already use, so nothing gets ripped out.

See the difference. Book a demo.