When 95% of Localization Is AI Verification—Except When It Isn't

The 95% Figure Everyone Is Quoting

Florian Faes, Managing Director at Slator, made a claim recently that landed hard across the language industry: 95% of localization workflows involving humans are now verification workflows. Not translation. Not creative adaptation. Just checking what AI produced.

He is not wrong — for most language pairs. Machine translation has matured fast enough that the human role in Spanish, French, German, Japanese, and dozens of other languages has genuinely shifted toward post-editing. The bread-and-butter work that defined this industry for decades is shrinking, and top-line revenue for core translation activity is declining as a result.

But that 95% figure contains a hidden assumption: that AI output exists in the first place.

For a significant slice of the communities that US healthcare systems, school districts, and government agencies are legally obligated to serve, it does not.

Where the AI Pipeline Breaks Down Completely

Chuukese and Pohnpeian are the two dominant languages of Micronesia. Together they represent the primary languages of tens of thousands of people living in US territories, Hawaii, Guam, and Pacific Islander communities across the mainland — many of them low-income, underinsured, and dependent on publicly funded health and education services.

There is no production-grade machine translation engine for either language. There is no large language model that reliably generates Chuukese or Pohnpeian output worth a professional's time to review. The verification workflow does not apply here, because there is nothing to verify.

What that means in practice:

• A hospital in Hawaii cannot run a Chuukese-language consent form through an MT engine and hand it to a post-editor. It has to start from scratch with a qualified human translator.
• A school district with Pohnpeian-speaking families cannot use AI-assisted tools to produce translated parent communications at scale. Every document requires a fluent subject-matter expert.
• A language service company that subcontracts Pacific Island language pairs to a generalist vendor is almost certainly getting output from a bilingual individual with no formal training, no QA process, and no accountability structure — because that is the only way a generalist can fill the gap.

This is not a niche problem. The Federated States of Micronesia has a Compact of Free Association with the United States, which means Micronesian citizens can live and work in the US without a visa. That population is concentrated in places with active obligations under Title VI of the Civil Rights Act and the Affordable Care Act's language access provisions. When those agencies use AI-assisted workflows as a blanket solution, they create compliance exposure they may not even know exists.

The Two-Tier Reality of Language Access in 2026

The Slator data points to a structural shift: large language companies are becoming what Slator now calls Language Solutions Integrators — firms whose value comes not from translation itself but from maintaining expert human networks inside AI-driven pipelines. That framing is accurate for high-resource languages.

But it creates a two-tier system.

Tier one languages get faster, cheaper, and more scalable every quarter. Tier two languages — the ones serving the most vulnerable populations — get nothing from that progress. The gap between them widens.

Generalist agencies selling AI-forward workflows to healthcare and government buyers are not lying when they describe their process. They are just describing a process that does not exist for the communities that often need language access the most.

What Rare-Language Expertise Actually Requires Right Now

Because there is no MT to post-edit, working in Chuukese or Pohnpeian demands something the current industry conversation is systematically undervaluing: a qualified bilingual professional who can produce accurate, culturally appropriate output from a cold start, in a subject area — medicine, law, education — where errors carry real consequences.

That means vetting translators for subject-matter competency, not just fluency. It means building glossaries from scratch because no standardized terminology databases exist for these languages in clinical or legal contexts. It means quality review by a second qualified speaker, because there is no automated QA tool that catches errors in a language it cannot parse.

This is slower and more expensive than MTPE workflows. It is also the only approach that produces output a healthcare provider can stand behind.

For language service companies that subcontract Pacific Island language pairs, the question to ask your vendor is direct: do you have named, credentialed translators for this language, or are you relying on bilingual individuals found through community networks with no professional accountability? The answer tells you everything about the risk you are accepting.

The Takeaway

The 95% verification figure is a real and important data point about where the language industry is heading. It should reshape how generalist agencies staff, price, and position themselves. But it should not reshape how healthcare systems, school districts, and government agencies think about their obligations to Pacific Islander communities — because AI has not changed anything for those communities at all.

If you work in language access procurement and you serve populations that include Chuukese or Pohnpeian speakers, the safest move is to confirm — before you sign any contract — that your vendor's workflow for those pairs is human-led from start to finish, with named linguists and a documented QA process.

If you want to talk through what that looks like in practice, we are happy to walk you through our process.