Language Industry 2026: The TMS Pivot

Introduction

At the end of 2025, I decided to take up the pen and write a series of articles breaking down the current state of the language industry. I realize it's easy to call out certain practices, tools, or workflows, but I wish to emphasize that my goal isn't criticism for criticism's sake. Instead, I aim to identify and analyze the reasons behind the race to the bottom that is putting a lot of pressure on our beautiful industry, and brainstorm publicly about the actions we can or should take to secure a future where technology will undoubtedly play an even more prominent role, but where human experts reclaim ownership over how multilingual content is architected and delivered — and get proper recognition for it.

Meanwhile, the ongoing boxing match between AI optimists and naysayers continues to dominate the conversation, which to me feels like a battle we shouldn't even be fighting. I believe that this debate is as artificial as the cover image of this article, and that it blinds us to what's really happening. Slowly but steadily, the language industry has ended up in an uncomfortable yet predictable economic situation where the supply of human translation has exceeded demand, resulting in historically low prices for the service. At the same time, the demand for efficient and responsible AI-assisted localization has exceeded the available supply, creating an imbalance that is producing a number of unpleasant side effects: while localization budgets are being reduced or allocated elsewhere, buyers are experimenting directly with AI, exploring substitute solutions, or engaging with new types of competitors.

I strongly believe that the first step towards solving a problem is understanding its causes. In my previous articles*, I already wrote extensively about what I believe are the root causes:

• The quality debate, where we fail to separate linguistic perfection from broader business outcomes. We obsess over dots and commas instead of measuring impact and intent, leading to misalignment between localization teams and upper management. Furthermore, quality is in the eyes of the beholder, which implies emotion and preference — a dangerous cocktail in business.

• The neural machine translation shock as the origin of our industry's pain. Neural MT was our industry's first experience with a disruptive AI breakthrough, and we weren't ready for it in terms of customization as well as implementation and monetization. It even took a couple of very promising companies down (Lengoo, Unbabel).

• The overvaluation of the translation memory which, despite being one of our most celebrated innovations, initiated the gradual degradation and commoditization of human expertise by reducing translation to units and promoting per-word bargains. When workflows are optimized for TM leverage and fuzzy match discounts, it's no surprise that price becomes the sole differentiator.

I still haven't said everything I wish to say about the matter, so I'm going to address the elephant in the translation room one more time. For now, this will be my final article on this topic. One last look back, then full speed ahead.

Legacy 101

In today's economic and technological climate, modern enterprise localization is far more complex than it once was. Localization teams must deal with several challenging requirements including, but not limited to:

• High-velocity or even real-time content that needs to be shipped at the speed of light.

• Multimodality; audio, images, and videos are now an equally prominent part of the localization scope in addition to text and documents.

• Dozens of content repositories and file formats that must be integrated. A GitHub connector is nice, but only starts to make sense and generate ROI when Figma, Jira, Notion, Slack, etc. are also thrown in the mix.

With localization managers under constant pressure to cut costs at all costs, they are simultaneously expected to do a lot more with far less, i.e., deliver more content at a faster pace, with only a fraction of last year's budget and, of course, without compromising quality.

So what is the professional language services industry offering in response? Processes that are driven by copy-paste, documents, emails, manual triggers, and minimum charges. The days when LSPs could wait for 50,000-word documents into 20 languages to come in on a daily basis are over, but we're still applying the exact same processes we developed for them, running localization like it's 2010.

The unavoidable consequence of this legacy problem isn't necessarily the undervaluation of translation quality; I believe it's the notorious perception of localization being a burden and a cost center. When translation feels like a clunky, expensive, and time-consuming practice, it's natural for buy-side stakeholders to look for cheaper and quicker solutions. Buyers will be enticed to experiment with AI themselves, and feel positively surprised about the good-looking quality and the speed at which their documents get translated by ChatGPT or DeepL. Furthermore, if they take care of the AI work they would expect their supplier to do for them, why not ask for a discount since the translation is only a quick read away from being perfect?

According to me, the most representative example of the legacy problem is the MLV > SLV > freelancer pipeline. In the typical translation workflow, source content moves from the enterprise TMS to the MLV and their TMS, then the SLV and their TMS, and finally the freelancers (and their CAT tool). It travels through inboxes, spreadsheets, and Plunet or XTRF instances, is copy-pasted into different file formats, and gets engineered multiple times to be compatible with the tech stack of the different operators in the supply chain. By the time a freelance linguist receives the actual work, it's completely detached from its context and the business purpose it's supposed to serve. Machine translation is typically proposed to create some sense of automation, but in the end, the poor freelancers are paid peanuts to do work they consider an insult to their profession.

And then we're not even mentioning the security gaps in this pipeline.

Zero Trust in Translation by Crowdin

TMS is dead. Long live TMS.

For more than 30 years, the TMS has been the beating heart of all professional translation activity. For every problem or requirement — from connectors and MT models to proxies, reports, and string capabilities — the spontaneous reflex was to turn to the TMS market in search of a solution. It became so central to our operations that it grew into the exclusive starting point and end point for language operations, and we took this for granted (including myself).

The TMS monopoly created a deep dependency where we mirrored innovation to TMS compatibility. For certain advancements, this worked well. Neural MT, for example, was bolted on right after its breakthrough as it could easily be sprinkled over XLIFF units and exposed as an additional translation result alongside TM matches. Nevertheless, every step beyond the safety zone of translation memories, term bases and CAT grids gradually increased the complexity of new development.

Consider connectivity as an example. Connectors have always been a lucrative business for TMS vendors, but despite many attempts involving "headless TMS" and connector boxes, the TMS market never managed to solve connectivity for the non-technical stakeholder. Early initiatives resulted in either acquisitions (Clay Tablet, Cloudwords) or rebrandings (Crosslang, iLangl). The rise of low-code integromats (Make.com, n8n, Zapier), the launch of Blackbird, and the pool of integrations at Crowdin reaching 700 have made it irreversibly clear that the typical TMS connector is limited to point-to-point data exchange with little to no customization, heavy maintenance, and serious lock-in danger.

Another example is multimedia localization. Basically every TMS provider, from memoQ and Trados to Smartcat and XTM, added subtitling through live video previews, but none dared to move into speech or address fundamental problems such as the incompatibility between subtitles and translation memories (safety zone, remember). AI dubbing became the prerogative of specialized products — losing count of products with "dub" and ".ai" in the name — that claimed a significant share of the TMS market.

If the advent of LLMs has taught us one thing, it's that sprinkling peacockish GPT features over segmented text, fuzzy matches, and term base hits isn't sufficient to make LLMs the go-to engines for localization programs. The fact that Crowdin and Phrase launched full-fledged AI dubbing studios, memoQ teamed up with Voiseed, and XTM acquired TXTOmedia might indicate that taking back market share from multimedia solution providers is a higher priority than bolting on GenAI. The TMS market experienced its very own trough of disillusionment, and is still climbing out of it.

Conclusion

After three articles on the state of the language industry, the conclusion still hasn't changed: we're stuck in a debate that won't produce a winner. On one side, the AI optimists who believe that AI is a kind of magic that will make everything better, faster, and cheaper. On the other side, the skeptics who dismiss it as a pattern-replicating tool. I believe the truth lies somewhere in between.

AI is dangerously (pun intended) powerful, and I always ask myself what would happen if, instead of giving it only one meaningless shot to produce a solid result from some context-poor input, we gave it access to the universe of resources in the localization pipeline, including the opportunity to communicate with coordinators and to correct itself.

At the same time, it won't take over our entire operation at the push of a button. Implementing AI isn't a matter of plugging in a model and expecting it to produce miracles. The road toward efficient and responsible AI is a long one and requires certain conditions to be fulfilled. According to me, those conditions are:

1. Deep customization: Off-the-shelf language models are under no circumstances good enough to carry the weight of an enterprise localization program.

2. Multi-system integration: Practicing AI can't be limited to enabling one model in a single system.

3. Support for multiple use cases: Narrowing AI down to machine translation means diluting its strength.

4. Human in control: For business-critical operations, human involvement continues to be essential — no matter how big or small.

The good news is that we have everything we need for a much-needed reset of professional language services. We have the models, the APIs, the workflow builders, and plenty of human expertise. The ultimate goal is to build sophisticated end-to-end workflows within a larger global content ecosystem where localization is triggered at content creation and carried through to publication. In this ecosystem, AI serves as an agnostic and dynamic layer, doing a lot of the heavy lifting that human operators typically experience as cumbersome, repetitive, and counterproductive.

To make this happen, there are pressing challenges to address. While the whole planet has had a taste of ChatGPT, AI agents, and vibe coding, the language industry still relies on processes rooted in the DTP era, driven by a static, context-poor database that was invented 30 years ago and hasn't fundamentally changed since. No language tech category illustrates this stagnation better than TMS, which used to be the center of everything in localization. This model is starting to break.

In the AI era, the logic of designing processes based on tools must be flipped: start with the desired workflow, then build or configure technology around it. This shift is subtle but powerful, moving localization away from static systems and toward agile processes that keep pace with modern content demands rather than imposing rigid structures. In practice, this means openness, robust connectivity, and orchestration layers that let platforms handle real-world complexity. The winner in this space will be the one that makes it easiest to build — orchestrating rather than monopolizing.

Long live TMS.