W
hen I founded my business in 2000 from the basement apartment of my house in Mississauga, I was a single mother to two daughters. Does it sound like the classic cliché on entrepreneurship? Perhaps, but it’s the truth. Even then, we never shared translation memories (TMs) across clients. Each client had its own TM on our secure on-premises servers. Trados did the work. Our infrastructure held the data. At the time, we didn’t call it “sovereign.” We called it doing the job properly.
Twenty-six years later, “sovereign AI” has become a marketing category. A cool term. Vendors are packaging that same principle, where client data stays with the client, the model runs on the client’s infrastructure as a differentiator. And here is the truth the industry should say out loud: The architecture is not new. The branding is.
Advertisement
The translation industry’s relationship with technology has always been more sophisticated than outsiders assume. From the first machine translation experiments in the 1950s through the computer-assisted translation (CAT) tool era of the 1990s and 2000s to today’s large language model (LLM)-driven pipelines, the fundamental challenge has never changed: How do you handle human language with all its ambiguity, register, and cultural weight at scale, without losing quality or client trust?
Trados and its peers gave us translation memories, terminology databases, and a model built around human expertise augmented by software. That model served language service providers (LSPs) well for three decades. It also created the habits, workflows, and client expectations that now make the AI transition genuinely hard. The co-pilot era doesn’t just add a new tool. It challenges the production model from the ground up.
Today’s platform landscape reflects that tension. Trados GroupShare with Studio Copilot can run local LLMs through Hugging Face. memoQ TMS has offered on-premises deployment since launch. Lilt advertises air-gapped and dedicated sovereign-cloud configurations for regulated and government work. XTM Server does too. Meanwhile, DeepL is moving toward AWS-only infrastructure, and the foundation models — Claude, ChatGPT, Gemini — cannot be truly on-premises because the model itself lives in someone else’s cloud.
The deployment menu is wide. What most LSPs are missing is not the technology. It’s the framework to explain the trade-offs honestly to clients.
I want to be precise here because intellectual honesty matters. Some things in 2026 are genuinely new, and they deserve the attention they are getting.
What is new is the combination: on-premises deployment philosophy layered with bring-your-own-LLM capability, packaged under a sovereignty narrative that is also anchored in specific ownership structures and community values. Companies like LIC (Language Intelligence Corporation), with its Canadian-Indigenous ownership model, are not simply reselling a technical architecture. They are making a value argument about who controls language infrastructure and who benefits from it. That is a different conversation, and an important one.
What is also new is the pace. The window between “experiment” and “client expectation” has collapsed. LSPs that took a wait-and-see posture in 2023 are now explaining gaps to clients who have been running their own AI pilots for two years. The technology did not wait for the industry to form a consensus.
The danger I see most clearly for LSPs right now is not that they will adopt AI too slowly. It is that they will adopt the language of AI sovereignty without doing the harder work underneath it.
Sovereignty is a choice you make in how you build, not a badge you purchase from a vendor. An LSP that cannot walk a client through the full deployment menu and explain the data governance trade-offs for each item is not offering sovereign capability. It is reselling sovereignty marketing.
That distinction will matter in the regulated sectors that represent the highest-value, most defensible work in our industry: legal, medical, government, financial. Those buyers have compliance requirements that will force the conversation. LSPs that have built the practice — not just adopted the vocabulary — will win the work. The ones who bought the badge will lose it.
Advertisement
This is not a prediction. It is what I see working.
Audit your current stack honestly. Which platforms do you actually use? Where does client data sit today: on your infrastructure, a vendor’s shared cloud, or somewhere in between? Most LSPs, when they do this exercise, discover the answer is messier than their contracts suggest.
Build a deployment menu you can explain. Not every client needs on-prem. Most don’t. But you should be able to offer a clear choice, explain the implications of each option, and put it in writing. That capability is a differentiator whether you ever call it “sovereign.”
Train your people, not just your models. The LLMs are available to every competitor. It’s the quality of your post-editors, terminology specialists, and project managers who can catch what the model misses that is hard to replicate at scale. Invest there.
Be the honest voice in the room. Clients are drowning in AI claims. The LSPs that build trust in this moment will be the ones willing to say: “Here is what the technology can do, here is where it falls short, and here is how we protect your data regardless of which tool we use.” That conversation sounds like integrity. It is also a business strategy.
I have been in this industry for more than two decades. I have watched it absorb fax machines, the Internet, globalization, cloud computing, and now LLMs. Every wave produced the same pattern: early panic, vendor overclaiming, and a period of genuine disruption. Finally, a new equilibrium emerges in which the fundamentals — quality, trust, client relationships, and human judgment — reassert themselves at a higher level of complexity.
We are in the disruption phase right now. The LSPs that come out of it stronger will be the ones who refused to confuse the marketing with the practice. Sovereignty is not a product feature. It never was. It is how you choose to build and how you choose to serve.
Ema Dantas is the founder and former CEO of Language Marketplace, which she grew from a basement startup in 2000 into one of Canada’s largest privately owned LSPs before selling it in 2022. She is the author of Mental Mountains; a speaker on entrepreneurship, AI, and mental health; and the founder of Peaks for Change Foundation. Her white paper, “From CAT Tools to Co-Pilots: How AI Is Reshaping Translation, Localization, and the LSPs That Will Lead the Next Decade,” is available by email request only, at ema@emadantas.com.
Advertisement
Related Articles
Interprefy partners with the Europe-Asia Economic Summit 2025 to provide real-time AI-powered speech translation and captions, enabling inclusive, multilingual dialogue across continents.
→ Continue ReadingBy providing additional supported languages for the framework, local communities worldwide will be able to engage in meaningful movement and sustain lifelong participation in physical…
→ Continue ReadingVisual design of the user interface (UI) is an integral part of delivering a great user experience, be it for consumers, enterprise workers, or other…
→ Continue Reading
