Act 1: The Snapshot

We begin with our feet firmly planted in the present and reflect on where the language services industry currently stands (spoiler: there’s no way to avoid this thing called GenAI).

But do we understand where we are and where we are headed? Are we asking the right questions, or should we just start with the “Why”? 

There is a sense of urgency that powers the varied market players. We explore new pathways, test new solutions, learn from the mistakes made along the way, course-correct, and start again.

The industry is evolving, and we are in motion. In the following pages, we examine what this movement looks like. 

Erik Vogt

Solutions and Innovation Strategist
at Argos Multilingual

Erik Vogt is a veteran solutions architect, program director, and executive consultant with deep expertise in AI enablement, multilingual content, and global operations at scale. He partners with organizations to unlock business value through strategic innovation, operational alignment, and measurable outcomes across complex, cross-functional programs.

Even when there’s a strong appetite for innovation, the reality on the ground is far messier. Most language services teams aren’t operating with a clean slate. They’re carrying years — sometimes decades — of infrastructure decisions, integrations, and workflows that were built under a very different technological framework. Although it may surprise some of you, there are still clients today who require the resizing of .rc files or compiling CHM files as part of their localization process. Those are literally 20th century technologies, having been introduced in Windows 1.0 and Windows 95, respectively.

Add to this a long list of proprietary systems built by a cohort of perhaps 10,000–25,000 localization buyers (my estimate includes only companies who spend over $100k on localization) and the many permutations of CMS, LMS, DAMS, CRMs, and TMS systems could easily reach 500,000 — not including other significant differences like software versions and let’s not forget, ahem, languages. And the legacy system issues aren’t just technical; they’re also hardwired into organizational structure, business endpoints, and cultural dimensions. They represent baked-in assumptions about quality, accountability, and even the definition of success.

Add to that the sheer velocity and scale challenges of market leaders, where computing, storage, and transaction costs are the primary constraints.

It‘s easy to say, “AI will make this faster,” and we hear that just about every day. I don’t need to cite a single number to prove this point, but I feel I should anyway. According to a 2024 Gartner report, 79% of CEOs say AI is a top priority for business transformation, yet only 15% of companies have successfully scaled AI beyond pilot programs. AI is talked about everywhere, but a staggering number of AI projects fail to deliver value, by various accounts between 70% and 90% in fact. Of course, OpenAI and other LLMs have started a new arms race over the past couple of years that many expect to improve the odds, but aligning systems, maintaining accountability, defining correct KPIs, and reorganizing systems remain challenging.

Still, even against these odds, the language services industry is starting to deliver real value, in part because of decades of experience with foundational technologies like TMS and NMT systems. Adding an AI-based review can add demonstrable value, not to mention exciting new opportunities like orchestration, knowledge graphs, and soon generating content on demand directly from a new set of core assets. These new capabilities are measurable and compelling, but we still run across systematic challenges in deployment.

Case in point, how do we predict project costs for a procurement system that expects word counts and predefined rate cards, when the NMT and AI output is not determinative? Who is accountable if technology doesn’t translate to lower human effort on the part of translators or reviewers? How does one even measure that when there could be half a dozen systems being deployed across a localization program? How should we now manage CMS systems that need to accommodate multilingual (or even multimodal) feedback loops? Or what do we do with a QA process that flags AI-generated content as inherently suspect, regardless of outcome (or can’t effectively determine what AI-generated content is in the first place)?

The constraint, then, isn’t just what AI can do — it’s what the end-to-end system lets it do. Real change requires mapping the whole environment: the business transaction layers, the data flows, the metrics, the regulatory expectations, and the unspoken norms that shape how things actually get done.

In many cases, we’re asking 2025 questions while living inside a 2010 (or earlier) infrastructure.

Three hidden bottlenecks that block AI innovation

While the promise of AI in language services is real, many organizations are finding that the biggest obstacles aren’t about models or capabilities — they’re structural. The innovation bottlenecks are often hidden in plain sight, buried in processes, policies, and systems that were never designed for rapid, intelligent change.

1. Rigid procurement models

Traditional procurement structures are optimized for predictability and scale, not experimentation. They depend on rate cards, fixed scopes, and RFPs that assume human translation is the atomic unit of value and that human effort has a predictable level of output.

But AI-driven workflows don’t always conform. They thrive on iteration, feedback, and blended outputs. There are several major bottlenecks, including buyer-hosted systems, intermediate technology providers, and language service providers, and their subsidiaries. Business endpoints require deterministic metrics, but procurement teams often can’t flex to accommodate those variables, and as a result, innovation stalls before it even starts.

2. Legacy metrics and incentives

A few years ago, I helped pitch metrics heavily, with a “Measure what Matters” campaign. I still find it deeply useful, but I recognize that we need another level of thinking on how to apply it in today’s world. The premise is that you get what you measure — and in most cases, we’re still measuring inputs (words translated, hours billed) rather than outcomes (customer understanding, speed to market, user engagement). AI changes the shape of the work, but if quality metrics and accountability structures remain static, teams will default to the safest path — even if it’s no longer the most effective.

For example, if a program is highly commoditized, it is organized to standardize and minimize unit price. In most cases, that is per word. Words are an obvious choice. They are easy to count, and the human translation process roughly follows somewhat predictable levels of effort. The time it takes to translate 100 words varies by content and translator, but the business layer imposes a structure that generally normalizes around sustainable means. However, with AI, the system performance (as defined by the product of a hybrid system against any standard metrics) may range from 0% to 500% of the non-AI throughput. While we’re on the topic of standards, does MQM still represent the valid framework for quality? How much does severity and classification matter in an AI-powered workflow?

3. Fragmented ecosystems

Language doesn’t live in isolation. It’s woven through CMSs, product databases, analytics platforms, and user interfaces. AI can automate specific tasks, but if the surrounding systems are brittle or siloed, the cost of integration outweighs the benefit of innovation. Too often, the conversation stops at “Can we use AI here?” rather than “What would it take for our whole system to support this kind of intelligence end-to-end?” This helps us get to a deeper journey of discovery for the goals and means of using language to deliver value to organizations in the first place.

I think a fundamental rethinking of language is at the heart of the discussion about LangOps, and how language is handled across multiple business functions, including value delivery (products, services, and experiences), trust and compliance, market access, brand identity, and operational clarity. For example, I have seen many in-country review efforts clash with central localization initiatives due to a profound mismatch of expectations. The in-country reviewers, such as resellers or in-country sales teams, are often not linguists and can introduce errors and provide erratic feedback. However, they might have a bone to pick with the head office about the inclusion of translated features they can’t sell in their region.

Unpacking the chaos

I think many in the language services industry will relate to the challenges I’ve skated through and shrug, “But that’s just our business.” I believe there is a more strategic way to tackle this complex world, and it starts with rethinking some basic assumptions.

One of the first to go is the premise that translation is a fixed process with fixed parameters (such as word count). Describing global strategy through word count is about the same as measuring a play by its word count. Would anyone say they got ripped off because the price to see “A Comedy of Errors” is probably the same as “Hamlet”, when the former only delivers around 14,701 words and the latter roughly twice that at 29,551? Of course not — “the play’s the thing,” not the average cost per spoken word.

To address this, return to the original purpose of the translation, and even whether or not it needs to be translated at all. Estimate the consequences of failure, including the consequences of untranslated, mistranslated, culturally mismatched, inconsistent, high latency, and unclear content. Consider whether some translation spend is underdelivering in its core purpose. Is the in-market review by non-linguists adding more time and cost than it’s worth, especially if contrasted by underspending in terminology control, or in ensuring the right content is being translated? How about when contrasting with deploying a RAG (Retrieval-Augmented Generation) workflow, AI-driven content tiering, or using AI as a Governance Layer? Despite all these amazing new capabilities, does adding them offset the structural inefficiencies of legacy workflows, incompatible business systems, or misaligned data systems?

Or perhaps a more important question is not whether you can deploy the latest AI capabilities, but whether one can define the value it could bring. As Simon Sinek said perfectly in 2011: “Start with Why.”

Read the full 132-page Global Ambitions: (R)Evolution in Motion publication featuring vital perspectives from 31 industry leaders on the ongoing AI-spurred (r)evolution.

Get your digital copy today