For three decades, the translation industry has run on a stable premise: a translation memory retrieves what looks similar, a human translates the rest. Trados, memoQ, and their peers built an entire generation of tooling on that premise — and it worked well enough, for long enough, that most of the industry stopped asking whether it was still the right one.
It isn't. Not anymore.
The tools that speak AI's language are already here
The first generation of CAT tools was built to compare strings. The generation now emerging is built to compare meaning — and it changes what a translation memory can actually do.
memoQ's Adaptive Generative Translation retrieves relevant past translations and feeds them to a large language model as context, so a suggestion reflects how you have already phrased things, not a generic rewrite. Smartling offers a comparable capability, built around the same principle: pull in translation memory, glossaries, and style guidance, and let the model generate from that context rather than from a bare source sentence.
This is retrieval-augmented generation — RAG — and it is a genuinely different way of using a translation memory. A memory searched this way finds a sentence reworded, reordered, or expressed with different terms, as long as the underlying meaning matches. A fuzzy match built on character comparison alone simply cannot do that.
The industry is not choosing whether to adopt this. It already has. The question left standing is how.
The new entry ticket is not the same as the old one
Here is what the shift actually requires, underneath the marketing language: every one of these systems needs, at minimum, a model capable of turning text into vectors, and a database built to search those vectors by similarity, at scale, in milliseconds. Neither of these was part of what a translation agency needed to run a CAT tool five years ago.
That is a different kind of investment than a per-seat licence. It demands infrastructure, and infrastructure demands a competency most language service providers were never built around: the ability to run and maintain machine learning systems, not just linguistic ones.
Large providers can absorb that cost, or acquire the expertise outright — memoQ's recent acquisition of a machine-translation specialist for its own generative offering is one example of an editor buying the capability rather than building it slowly. Most language service providers cannot do either. They are, overwhelmingly, small and mid-sized businesses — and that is not a marginal segment of this industry. It is most of it.
The dilemma this creates
Here is the trap. An LSP that does nothing will watch its productivity fall behind competitors using AI-assisted retrieval, on cost and on turnaround time both. That much is a familiar story — technology moves, laggards lose ground.
But adopting the new generation of tools does not simply solve the problem. It relocates it. To use a RAG-enabled CAT platform today, in practice, means placing your translation memory — the accumulated record of every term your linguists have ever validated, for every client you have ever served — onto the CAT editor's own infrastructure. Not optionally. Structurally: without it, the retrieval half of RAG has nothing to retrieve from.
Look closely at what that means in practice. Vendor documentation for the current generation of AI-assisted CAT tools routinely confirms that the generative component runs on a named third-party cloud provider — the infrastructure choice is stated plainly, sometimes even presented as a selling point on security grounds. Hosting arrangements described as close to on-premise often turn out, on inspection, to mean a dedicated server that the vendor hosts and manages, not one the LSP controls itself. And the pricing for that arrangement is, in more than one case, nowhere published — a conversation to be had directly with the vendor, rather than a number an LSP can plan around.
None of this makes the technology bad. It makes the dependency real. An LSP's most valuable asset — the memory of every sentence its linguists have ever gotten right — becomes, functionally, something it accesses rather than something it owns. Switching platforms later does not mean exporting a file. It means renegotiating access to your own history.
So the dilemma is not "adopt AI or don't." It is: adopt it on someone else's terms, or find a way to adopt it on your own.
A third path: shared infrastructure among LSPs
Here is where the calculation changes. The cost that makes self-hosted retrieval-augmented translation unreachable for one small LSP does not scale linearly with the number of LSPs sharing it. A GPU server, a vector database, the operational expertise to run both — these are largely fixed costs. Spread across ten or twenty independent agencies rather than borne by one, the arithmetic looks very different.
This is not a new idea in adjacent industries. Independent businesses have pooled infrastructure before, when the alternative was dependency on a handful of large suppliers who controlled the means of production. Applied here, it could take more than one shape: a technical cooperative that several LSPs fund and govern jointly, with each retaining full ownership and portability of its own memory within a shared indexing and inference layer; or an infrastructure partner — commercially independent from any CAT tool editor — offering the same architecture as a managed service, but built from the outset around data portability rather than lock-in as the default expectation, not an afterthought bolted on for larger clients who ask.
Either version rests on the same principle: the retrieval engine can be shared. The memory itself never has to be.
What this moment actually calls for
The industry does not have the luxury of pretending this transition isn't happening, and it does not have to accept that the only two available outcomes are irrelevance or dependency. The technology that makes modern translation memory genuinely more useful — searching by meaning instead of by spelling — does not require surrendering the asset that memory represents. It requires infrastructure. Infrastructure can be built once and shared, the way it already is in industries that faced this exact fork before translation did.
The agencies that ask this question now, while the market is still taking shape, will have more room to choose their answer than the ones who wait until the choice has already been made for them.