Human expertise at the core, AI in the loop: the future of software localization

In this Globally Speaking episode hosted by RWS, Luis Lopes sits down with Milen Epik, President of Epik Global Solutions, and Adam Asnes, CEO and founder of Lingoport, to explore what software localization looks like now that AI coding agents are part of everyday development.

Key topics covered:

Why you can't just point an AI agent at your code and expect internationalization to work (spoiler: it's a good way to break things)
How context, from Figma designs to system prompts, shapes the quality of AI translation
Why cultural intelligence goes beyond words, and what it can teach us about changing business logic, not just text
The governance and guardrails needed when distributed teams all use AI on the same codebase
Why the future is "human expertise at the core, AI in the loop"

Whether you manage localization, build software, or are just curious about what happens between the code and the screen, this episode will change how you think about shipping products in every language.

watch here

Quality First

Internationalization (i18n) and localization should not break your code, jeopardize security, or create bad translations that push the users away.

Expert Technical and Cultural Care

What you need is an experienced team and partners that bring the necessary technical, cultural and linguistic skills and passion to every project, no matter the size.

Understanding to Business Results

What is the intent of your software and the localization effort? What markets and what business goals do you have? Understanding these at the start is what drives project delivery. Result: proven outcomes that succeed in global markets and meaningful gain for your business.

The Role of Context in Modern Localization and AI Translations

CONTEXT IS KING, BUT NOT ALL KINGS ARE EQUAL!

When it comes to localization, providing context to translators and reviewers has been important for a long time. As the industry turns to AI-based translation, we see that the same is true for LLM-based translation. Just like humans, LLMs perform better when they have context.

Context can be provided in multiple ways, but not all approaches are equal. Some approaches to providing contextual guidance date back several decades. For example, terminology and style information provide linguistic and brand context. Similarly, the legacy concepts of an exact match or perfect match utilize the context of the page or surrounding text to improve the accuracy of translation memory segment matches. But there are other, more modern approaches that provide context at a general or segment level.

The table (left) compares these approaches across their strengths, limitations, and typical applications. We have not included multimedia content, as the methods for providing context are highly specific. They can draw on a variety of sources, including source materials (transcripts, captions, scripts, and on-screen text), visual content, audio context, speaker information, timing, and other constraints.

See the post and discussion on LinkedIn here: Post | LinkedIn

Get in Touch

About the Analysis

The analysis in the table shows that there is no single form of translation context that works equally well for every content type.

Each approach has its strengths and limitations, and the most effective form of context depends on the type of content and the information available.

The best use of each method can be summarized as follows:

Terminology/style guides → best for linguistic and brand context
Metadata → best for string-specific context
Reference material → best for broader knowledge and reference context

Textual context → best for surrounding semantic context
Visual context → best for UI, location, and functional context
Live in-context integration → best for current runtime context
Cross-lingual context → best for information derived from existing translations

Accountability and Judgement: Why we still need humans in globalization?

If AI can create content, transform it, perform QA, and even correct its own output, why do we still need humans in globalization?

While some content types and industries may be able to localize without any human involvement, regulated sectors—such as healthcare, legal, and finance—still require human oversight and judgement at critical points of risk. Rather than having human linguists review everything, LLMs can handle much of the work, while governance frameworks determine where human expertise is required. The level of human review depends on the level of risk, and an audit trail ensures that the human reviewer, product manager, or legal counsel who signed off on the work can be identified.

In high-risk environments, localization still requires a person to review, approve, take ownership of, make judgement decisions and ultimately be accountable for the final result.

The same principle applies to software development and software internationalization. In software, the code is the content. While AI can generate and improve code, organizations still need expert professionals in system design, software architecture, and internationalization (i18n) to assess risk, validate AI-generated output, and intervene when necessary. Ultimately, humans remain responsible for the quality of the final product. AI can assist, but responsibility still belongs to people.

A contact in the software industry recently said something in a casual conversation that stuck with me:

"We will never replace all software developers with AI because, in the end, every organization wants someone who is accountable."

Thoughts?

Add to discussion on LinkedIn here: Post | LinkedIn

LLM RESEARCH FINDINGS

LEARN ABOUT MILEN EPIK'S RESEARCH AND DISCUSSION WITH AI EXPERT

Dec 2025

For years, teams had to choose: do you want translations that are fast and cheap, or fast and high-quality? Getting all three at the same time was almost impossible. Now, there is hope that LLMs can finally change this.

But UI strings are still a challenge because they are short, often unclear, and missing context. Many localization specialists are wondering: Can LLMs actually handle UI strings better?

To help answer this question, our team conducted fresh research comparing multiple approaches to software UI string translation:

Machine Translation + human editing
Straight-to-LLM translation
LLMs boosted with extra context (visual cues, better prompts, etc.)

And the results were… surprising.

In the last webinar of the year, Marina Pantcheva (Director of Linguistic AI Services, RWS) will join Milen Epik (VP of Globalization Solutions and Services, Lingoport) to discuss the research: what worked, what didn’t, and why short UI strings remain one of the toughest challenges in localization.

If you’re curious about what LLMs can (and can’t) do for software localization today, this session will give you a clean, honest look at the data.

watch webinar