The economy has been governed by the ebb and flow of supply and demand for the last 300-odd years. The demand for goods and services impacts the cost of employment, profits and retirement funds impact the capital markets, and market growth drives capital investments towards new goods and services. While the importance of these individual factors changes over time, and governments take different approaches to regulating the markets, the key factors and considerations remain the same.

In the 20th and 21st centuries, marketing did everything possible to work against the perfect information theory postulated by theoretical economists to describe a well-working capitalist society. With perfect information in a market, all consumers and producers have complete and instantaneous knowledge of all market prices, their own utility, and own cost functions. This is one of the prerequisites for the perfect market situation, which traditional models best describe. 

Enter AI. Or rather, Artificial General Intelligence (AGI), the next development in line to disrupt the status quo. Forget about the narrow niche of localization for a moment. AGI could be the invisible hand guiding you while it crunches everyone else’s possible decisions. It could make more informed decisions than employees — good old humans — who often decide for their own good and take the path of least resistance rather than the one in the best interest of the organization they work for.

It is a heady question, what will happen to the linguistic jobs? An even bigger question is what will happen to jobs in general? Let’s assume AGI can make better purchasing, hiring, and investment decisions than people, and computers will not only deliver products and services, but also make buying decisions. If computers read our texts, they will not need impeccable quality. Do we really need to worry about linguist jobs more than we worry about the world’s welfare? 

Then there is the TMS. The question isn’t what will happen to the TMS, but more what will happen to money, capital, and the feeling of security. Will tech enable such a societal leap and render our current problems obsolete? 

Data and structured data

If computers are to take over the organization of our information, we will probably see complex structuring of data. As a side note, this is not what transformer models are good at today, but developments like knowledge graphs are the first steps in this direction. As a buyer, you perform text-based discovery. However, when comparing options and presenting to others, you make comparison tables, which is already a structuring of information. 

Structures change in complexity over time because of external factors. Take the example of a Request for Proposal (RFP) answer, one type of structured information. Before 2023, no client-side translation management system RFP contained any LLM- or AI-related questions. On the vendor side, before 2023, no TMS vendor compared themselves to others on what LLMs they support, whether they support only single-pass or multi-pass prompts, whether they store information about segments where a translation was made or corrected by an LLM, and if so, which prompt was executed at what time by which engine, etc. These pieces of information are essential to compare today’s LLM results with the results you get a year or two from now. They do not change the usefulness of a TMS now. They are future-proofing them.

Engineering, as a principle, is about understanding a problem in all its details and solving it, if possible, in one go. Software technologies today often become outdated because the underlying database does not support the changes required by the passage of time. Databases can be updated, and new databases can be built and optimized. However, this requires a willingness and understanding on the vendor side. 

 The actual need is often overshadowed by what the vendors want you to perceive. For example, I spent weeks trying to understand how an attractive technology like HubSpot could be used to manage sales in a translation company. HubSpot’s data structure is missing two important concepts used in our industry: 

1. There is no straightforward way of applying fuzzy matches. 
2. Either/or products are not an option: you cannot offer your customer two quotes, one on MT post-editing and one on human translation, in one pipeline. 

Either HubSpot implements these requirements and becomes more feature-rich to support the current translation industry, or the translation industry changes and/or finds a way to use what it supports. In this particular example, guess which outcome is likelier than the other.

Will AI-driven technologies revolutionize data structuring and software refactoring? Will we, as an industry, retain the same requirements or change them to accommodate what needs to be stored? Our industry is currently not driving the change. Therefore, we need to examine the question from two perspectives: Why do we need a TMS now, and how are the LLM systems evolving?

Why do we need a TMS?

Let’s focus exclusively on the enterprise side because in this B2B industry, if enterprises do not need human translation services, language service providers and translators will be partially or fully out of business.

What does a TMS do for an enterprise?

• It allows the integration of data flows. You can already do this with any LLM.
• It offers customer portals. What is simpler than a chat-based interface?
• It handles complex file formats. Two main complexities are DTP and multilingual and cascading file formats (e.g., HTML in Excel). The client requirement is: “Only translate the highlighted text.”
• It offers to add steps. For example, if the format is Framemaker, there could be a preparation step and a post-processing step performed by a human or a computer.
• It offers a visual interface for understanding what was done to each document e.g., a project management interface.
• It offers a visual interface for editing the text. This is the linguist interface.
• It allows filling some of those segments with automation results. Think about machine translation engines, LLMs, etc. – the right routing according to user preference.
• It allows quality assurance tools to be run on these segments (some are ridiculously simple and unintelligent).
• It offers underlying databases that can affect the editing experience, such as translation memories, term bases, stop word lists, and corpora.
• It allows tasks to be outsourced to people and companies. You have multiple assignment methods and job-taking methods.
• It allows people’s work to be monitored (e.g., who delivered on time, etc.).
• It offers a business model called fuzzy analysis. While there are other business models, those have not been successfully applied in human translation (e.g., post-translation analysis).
• It stores transactional information, e.g., who did what and when.

What are the assumptions behind TMSes?

• Understanding and enforcing a consistent set of steps improves quality.
• The ability to identify who is at fault can improve the process.
• The process is engineered toward output quality: what got into the translation memory can be considered perfect. This is the assumption that the reproducibility of a translation is more reliable than getting the same thing translated twice differently.
• Translation quality has to be uniform because this feeds the translation memories.

What does an LLM currently do? When ChatGPT first came out, it could not handle file formats. Now it pretty much can, and its PDF filter is surprisingly accurate. What is more, it started to generate target-language files (not available for OCR yet, but there are some other cloud OCRs available):

I also tried to use ChatGPT to translate sentences with other machine translation engines and set up projects in TMSes via APIs. When I tried it, requests timed out, so I could not do it, but I received Python scripts. The capabilities of these technologies are scary in the long run.

So far, I do not see any capability for dashboarding, visual representations, or structuring the interactions. 

The human’s place in the loop

The question with AI is whether a human in the loop will be needed. Here, I see two possibilities:

1. Language quality will not reach human parity despite the inclusion of context — everything that a human can perceive and receive. In this case, a human in the loop will be needed, and a future TMS will be needed if you need a visual management interface for translations.
2. Automated translation requires the same method of improvement as human translation: the shift-left approach, having more extensive briefs and clarifying all questions at the moment of input. Here, there are two sub-situations:
   1. Humans will care about and pay money for preferential edits 10 years from now. This comes from the human psychology of ownership and feelings about the impact of text. This will necessitate a TMS, as you have to feed changes back into the system so it learns from them.
   2. Humans won’t be able to spend money for preferential edits. In this case, a TMS won’t be likely.

Can we already start working on such a TMS?

I would argue that the right point to start considering this will come when one feels comfortable understanding the process aspects of “LLMs.” I put LLMs in quotation marks because I tested the file format parsers and the execution capabilities of ChatGPT, but I have no experience with the other LLM executor platforms. This is a capability of the platform rather than of the model. Most of the investigations in the industry focused on language capabilities, but this is a very different aspect. If someone started to work on such a platform, I would consider investing some money, but not too much: LLM providers may easily surprise us there, eliminating the value of the investment.

Consider this: right now, it seems as if companies like OpenAI are operating with the mentality of “us against the world.” Uncompromising, all-encompassing, these tools are marketed with a lot of money to create the message: this is the future. Their product management is solid, and before we know it, an LLM becomes an AI management system, like how the 1990s Trados became the enterprise TMS of today. Reproducibility, human oversight, and human-in-the-loop are not aspects exclusive to translation — every industry operates on this basis due to the existing human structures at organizations. Therefore, the investment question becomes: Do we want to learn from other industries and design systems, whether we end up labeling them a TMS or not, that solve more general problems — efficiently coordinating humans and machines for a start — than specific tasks, such as content creation and translation?