The conventional wisdom about AI and translation runs something like this: machines will eventually replace human translators, but we're not there yet because nuance, culture, and context remain stubbornly human provinces. This framing misses what has already happened. The disruption arrived years ago, and it didn't come for the translators—it came for the translation agencies.
For decades, the global translation industry operated on a remarkably inefficient model. A pharmaceutical company in Munich needing its clinical trial documents rendered into Japanese would contact a language services provider, who would farm the work to a freelance translator in Osaka, taking a cut of forty to sixty percent. The agency's value proposition was coordination: finding qualified translators, managing deadlines, ensuring consistency across large projects. It was, in essence, a matching problem dressed up as expertise.
Neural machine translation demolished that value proposition. Not because the machines translate better than humans—in most specialized domains, they don't—but because they translate well enough that the coordination layer became optional. A competent human translator can now use machine output as a first draft, tripling their throughput. More significantly, clients can interact directly with translators, using machine translation to bridge the communication gap that once made agencies indispensable.
The hollowing of the middle
The numbers tell a stark story. The largest translation agencies have seen their per-word rates compressed by roughly half since neural machine translation went mainstream. Meanwhile, elite translators—those handling literary work, high-stakes legal documents, or culturally sensitive marketing—have maintained or increased their rates. The middle tier, the journeyman translators who handled routine business correspondence and technical manuals, found themselves competing against machines for the first time.
What emerged is a barbell economy. At one end, machine translation handles the vast bulk of the world's translation needs: tourists navigating foreign menus, businesses parsing international emails, researchers skimming papers in unfamiliar languages. At the other end, human translators command premium rates for work where error carries real cost or where the text must not merely communicate but persuade. The middle—competent but undifferentiated human translation—has largely vanished.
The pattern repeats
This dynamic offers a template for understanding AI disruption more broadly. The technology rarely eliminates a skilled profession outright; instead, it collapses the economic scaffolding around that profession. Travel agents, stockbrokers, and now translation coordinators share a common fate: their expertise was always partly genuine and partly a function of information asymmetry. When AI erodes the asymmetry, the genuine expertise must stand on its own.
The translators who thrived adapted in predictable ways. Some became "post-editors," refining machine output rather than creating from scratch—a less romantic but more sustainable business model. Others retreated upmarket, emphasizing the irreducibly human elements of their craft: the ability to make a German joke land in Portuguese, to preserve the rhythm of a poem, to catch the legal landmine buried in an ambiguous clause. The agencies that survived did so by becoming technology companies, building proprietary machine translation engines and selling integrated human-machine workflows.
Our take
The translation industry's experience should humble anyone making confident predictions about AI and employment. The prophets of automation consistently overestimate the technology's ability to replace human judgment and underestimate its ability to restructure markets. Translators still exist; many are doing quite well. But the industry that employed them has been remade from the inside out, and the workers who thrived were those who understood that the machine was coming not for their skills but for their economic moat. That distinction matters more than most AI discourse acknowledges.
