Language Governance for Enterprise AI: What CTOs Are Missing

E-Illingworth · an hour ago

Somewhere in your AI program, a red team ran the model through the usual gauntlet: jailbreak prompts, refusal-rate checks, sign-off. Almost certainly, all of it happened in English.

That matters more than it should. Research on multilingual jailbreaks from Chinese University of Hong Kong[1] found that low-resource languages are roughly three times more likely to produce unsafe output than high-resource ones, on the same model, under the same safety training (though the size of that gap shifts with the model, the language, and how researchers define 'unsafe' in the first place). The study’s survey of the field's own safety benchmarks found that 78.5% of them are written exclusively in English. The safety work behind your model was tested almost entirely in the one language your deployment isn't limited to.

This isn't a translation gap. Nobody mistranslated a refusal. If no model was never asked the question in Hindi, or Vietnamese, or Polish, nobody actually knows what it would have said.

Governance, not just translation

Regulatory pressure and AI adoption are consistently near the top of what pushes CTOs and AppDev leaders to bring new controls into a platform review. Regulation around AI and language access varies by country, but the safest option for anyone responsible for signing off on an a multilingual AI deployment is that “we tested it in English” simply isn’t enough.

There's a cost angle too, and it's just as concrete. Imagine you have a multilingual deployment of your enterprise software – let’s say English, French, Japanese, and German. You want to enter an Indian market and you need to deploy in one of the 22 languages scheduled in the Constitution of India. Aside from the looming question of which language should be your fifth Platform language integration based on your e-commerce strategy, user base, or other criteria, you also have to figure in the cost of translation and the cost of your ongoing AI token usage.

Tokenizers are trained mostly on English-heavy text. This means a sentence in English can require dramatically more tokens once it's translated into another language. Petrov et al.'s 2023 NeurIPS[2] study measured tokenization-length differences of up to 15x for the same content translated across languages, even on tokenizers built for multilingual support. In Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models (Ahia et al.)[3] extended that finding to commercial APIs and found the price of processing equivalent content varies by an order of magnitude language to language, purely from how the text gets cut into tokens before the model does anything with it.

A very recent 2026 study[4] measuring ten tokenizers against Indian languages found an average 8X tax relative to English, reaching 13X for Malayalam (the official language of the Indian state of Kerala) meaning the same document costs thirteen times the tokens in Malayalam that it costs in English, before a single unit of reasoning happens. If nobody is tracking spend by language, the fifth market doesn't just cost more to build. It costs more to run, indefinitely, in a way that's easy to miss until finance asks why the AI bill doesn't match usage.

Take Now Assist, the ServiceNow genAI suite that powers AI agents, summaries, task flow generation, comment-to-code, chatbot, and more. As part of the core Platform, Now Assist handles multilingual use natively, and our AI is built-in, not bolted on. And because that AI layer lives inside the platform itself, choosing how it handles language isn't necessarily a one-time setup step — it's a runtime decision, and one that ties into a broader globalization and localization strategy with downstream effects on IT, risk, and finance, not to mention employees and end-users. This strategic gap is exactly where Localization Workspace comes in as the orchestration layer for your multilingual implementation.

What does Language Governance mean?

Though it may sound like a LinkedIn buzzword to some ears, language governance is a specific and practical set of policies and tools for standardizing terminology, implementing approval guardrails, managing translations, and tracking spend and compliance across global deployments. In other words: one glossary, one set of rules, no matter who's writing or which market they're writing for.

“Millions of people work in global settings while viewing everything from their own cultural perspectives and assuming that all differences, controversy, and misunderstanding are rooted in personality. This is not due to laziness. Many well-intentioned people don’t educate themselves about cultural differences because they believe that if they focus on individual differences, that will be enough.”

― Erin Meyer, The Culture Map: Breaking Through the Invisible Boundaries of Global Business

Erin Meyer's The Culture Map makes a related but separate point: culture creates gaps across distinct axes — how directly feedback lands, whether trust builds through completed tasks or through relationship — and translating words correctly does nothing on its own to close those gaps. Take the term “half ten.” If you’re someone in the UK or Ireland, this probably means “half past ten,” or ten thirty. If you’re from Germany, however, it means “halfway until 10,” or nine thirty. That's not a one-to-one translation error of the individual words “half” or “ten.” It's a reminder that the term itself, not just the words around it, needs a governed, reviewed definition — which is a governance layer that Localization Workspace is built to hold.

In practice, the language governance layer has five pieces:

Terminology management centralizes the business glossary so Finance and HR and Legal all use the same definition of the same term.
Cost visibility turns translation spend from a line item you discover at the end of the month into something you can see by language, by region, by vendor, before the invoice arrives.
AI and translation control governs how enterprise content moves into translation and AI systems in the first place while a human in the loop still owns cultural and linguistic accuracy.
Workflow approval gives new terms, product language, and regulatory disclosures a clear owner and a sign-off chain.
Data lineage ties every localized term back to the data catalog it came from, giving you a record of who translated what, when, and how so that a regulator or auditor has a trail to follow. These audit trails are especially helpful when doing compliance work for legislation like the EU AI Act or the EAA.

If your language governance process looks like five applications in a trench coat with your glossary in a spreadsheet and your audit trail in an email conversation thread and your final approval bottlenecked by an overworked application admin, Localization Workspace was built with you in mind.

Our single, integrated governance solution gives you:

AI-powered glossary building. A Terminology Agent scans your existing knowledge base and surfaces the terms worth governing, instead of someone building the glossary by hand, term by term. The human in the loop has the final say, but our AI agent does the heavy lift in a fraction of the time.
One place to see translation spend and estimates. A governed glossary resolves a term once and reuses it everywhere, instead of paying to re-translate a slightly different version of the same term every time it appears — closer to a fixed cost instead of a wildly variable one. A glossary reduces inconsistency and duplicated effort, and when you configure the governance center with your translation providers, each piece of content can be tied to your spend estimate up front.
A terminology role, not an admin bottleneck. No longer is approval on the desk of an application administrator competing with security, accessibility, and AI configuration requests. Our dedicated terminology management role ensures a localization SME can approve or reject a term or piece of content directly, which is what actually keeps translation moving.
Built into the platform, not beside it. Localization Workspace runs inside the ServiceNow platform itself — the same instance, the same permissions model, the same audit tooling you already run everything else on. It isn't a satellite system bolted on beside your workflows; it localizes the workflows and content you already have.
Works with what you already run. It plugs into the translation providers and content systems you already have, rather than asking you to rip out a vendor relationship you've already negotiated.
Full visibility from request to delivery. Every translation request is logged from the moment it's requested to the moment it's delivered. This includes who requested it, who approved the content, and which glossary version it used, ensuring no guesswork in your audit trail.
Built to scale. The same consistent user experience whether you're governing five languages or twenty five.

A red team that only tests in English can't tell you what the model jailbreaks in Hindi. A tokenizer trained on English-heavy data can't tell you what a document costs in Malayalam until the invoice lands. A glossary in a spreadsheet might not tell an auditor which term was approved, by whom, or when.

As a CTO or IT decision-maker, we know you don’t need another system to secure, patch or defend. Localization Workspace doesn’t sit alongside your AI program as a separate risk to manage: it is the governance layer inside the same platform instance you already run everything else on. Fewer integrations, fewer systems to reconcile, fewer ways a language governance gap turns into an incident report.

Learn more:

Get Localization Workspace on the app store
Check out our Localization web page
Explore localization documentation
Get up to speed on the latest updates

[1] https://research.cuhk.edu.hk/en/publications/multilingual-jailbreak-challenges-in-large-language-mod...

[2]https://www.researchgate.net/publication/371040637_Language_Model_Tokenizers_Introduce_Unfairness_Be...

[3] https://aclanthology.org/2023.emnlp-main.614.pdf

[4]https://www.researchgate.net/publication/410888242_The_Tokenizer_Tax_Quantifying_and_Explaining_the_...