Ben Faes, CEO of RWS – Interview Series

Ben Faes is Chief Executive Officer of RWS Group. He brings more than 25 years of experience in leading digital transformation and scaling technology-driven businesses, with a strong track record of driving profitable growth, building innovative go-to-market models, and developing high-performing international teams.

Before joining RWS, Ben held senior leadership roles across the technology and business services sectors. At AOL he rose to Managing Director for France, before moving to Alphabet in 2008 where he pioneered YouTube’s monetization in Europe and later led multiple Google businesses across the EMEA region, culminating as Managing Director of Google Cloud for Southern Europe and Emerging Markets. In 2021, he became UK CEO of Webhelp, and following its acquisition by Concentrix, led global transformation and technology initiatives at Concentrix Catalyst.

RWS is a global AI solutions and content services company that helps organizations create, manage, translate, and protect information at scale, combining advanced technology with human expertise. The company specializes in areas such as language translation and localization, intellectual property services, and AI-powered content and data solutions, enabling businesses to communicate effectively across markets and deploy enterprise AI systems with cultural and contextual accuracy. With a large global workforce and decades of experience, RWS supports major enterprises in transforming complex data into clear, trusted content while accelerating innovation and ensuring ideas are understood worldwide.

You have led major transformations at companies like Google Cloud and now at RWS. How has your experience scaling enterprise AI platforms shaped your approach to building Language Weaver Pro as a mission critical translation system rather than a general purpose tool? 

What I’ve learned from scaling enterprise AI is this: general-purpose tools hit a ceiling fast in high-stakes environments.

Enterprises don’t just need something that works most of the time. They need something they can rely on – something they’re comfortable embedding across critical workflows, from HR to localization to customer support.

With Language Weaver Pro, we didn’t start by asking, “How do we build a better translation model?” We asked, “What does a translation system need to look like for an enterprise to trust it at scale?”

That shifts the focus. It’s not just about model quality. It’s about consistency, control, scalability and security – the things that make a system usable in the real world.

One of the things I’m most proud of is that Language Weaver Pro didn’t have to be adapted for enterprise later. It was designed for it from day one.

While others are now moving toward specialization, that’s something Language Weaver has built into its foundation for years. Our partnership with Cohere gave us world-class AI capabilities – and we layered in the enterprise controls and linguistic expertise needed to make them truly usable at scale.

Language Weaver Pro is built on a 100 plus billion parameter model. What architectural or training decisions were critical to achieving strong performance at the paragraph and document level, where many translation systems typically struggle?

Most translation systems are optimised for the sentence as the unit of translation – which reflects how the localisation industry has historically worked. The problem is that meaning doesn’t always live at the sentence level. Ambiguity, terminology consistency, and tone all operate across longer spans of text. Getting strong paragraph-level performance required deliberate decisions, including opting for a much larger context window than typical translation systems.

One problem that most systems see here is that handling larger context typically comes with increased processing times. This is where the Mixture of Experts architecture comes into play. MoE allows Language Weaver Pro to be simultaneously large enough to handle complex, long-form content and efficient enough to deploy it at enterprise scale.

The platform ranked first in 31 out of 32 languages in human led benchmarks. How does human evaluation differ from automated metrics, and why is that distinction important for enterprise use cases? 

Automated metrics are essential for scale. To consistently benchmark quality, you can’t continually run human evaluations across 32 languages and thousands of segments without them. But there are limitations too. Traditional metrics like BLEU measure surface-level word overlap, which means a translation can score well while missing nuance, or score poorly while being perfectly accurate. Even more sophisticated neural metrics are proxies for human judgement, not substitutes for it. A translation that is technically accurate but tonally wrong in a cultural context can cause real damage.

This is why we ran both. The blind human evaluations we ran with our own professional translators gave us the domain and cultural sensitivity that our enterprise customers and their end users really care about.

Many AI translation systems struggle with maintaining consistency across long documents. How does Language Weaver Pro preserve meaning and context across paragraphs, especially in legal or regulatory content?

Consistency is one of the toughest challenges in translation at scale – and it’s not something general-purpose models can solve on their own.

It’s also an area where we’ve had a real head start. Solutions like Trados and Language Weaver have long built in strong terminology controls, so you can enforce consistent language across even the largest, most complex content sets.

That matters more than ever in regulated environments. In legal or compliance content, consistency isn’t a nice-to-have – it’s critical. A single variation in how a defined term is translated can invalidate a contract, distort a regulatory filing or introduce real risk.

Terminology control removes that uncertainty. It ensures the final output uses exactly the language that legal and compliance teams have approved – consistently, at scale.

You have described this as a shift from translation to Language Intelligence. What does that mean in practice, and how does it change how enterprises interact with multilingual content?

For decades, the industry has viewed translation as a mechanical process which converts words from one language to another. And while specialized human translation processes such as transcreation have existed for a while, this has all hinged on a relatively slow process that focused on word counts. “Language Intelligence” fundamentally changes this paradigm. It’s a shift from converting words to truly understanding meaning; it is about grasping context, the nuance, the brand voice and specific intent behind the content, regardless of the original language. We’re moving from a passive, reactive task, to a proactive, strategic capability.

In practice, when our AI tools are able to make translations more fluent and context aware, the Cultural Intelligence human layer is able to do more than simply post-edit content, they’re able to focus their efforts on relevancy and impact.

The platform emphasizes governance, terminology control, and data security. What specific systems are in place to ensure translations meet compliance requirements in industries like finance and healthcare?

Trust is the cornerstone of our platform, especially for regulated industries. First, we offer deployment flexibility that public tools cannot. Clients can run Language Weaver on-premises or in their own dedicated private cloud. This ensures sensitive data never leaves their secure environment, which is non-negotiable for meeting data residency requirements.

Second, we provide granular control over the outputs that can change with an industry’s requirements. Clients embed their own approved glossaries directly into the model. This guarantees that a specific legal disclaimer in finance or an approved medical term in healthcare is translated correctly and consistently every single time, eliminating the risk of non-compliant language.

Finally, governance is built-in. The platform integrates with existing enterprise security frameworks, providing full audit trails and access controls, and customers can always route the content to experts humans in the loop for certificates of accuracy or for mandated review.

Unlike general purpose AI models, Language Weaver Pro is purpose built for translation. What are the trade offs between specialization and flexibility, and why is specialization becoming more important for enterprise AI?

A general-purpose model is a bit like a Swiss Army knife. It’s versatile – it can suggest a recipe from what’s in your fridge, sketch out a landscape plan, write a poem or even attempt some code.

But when it comes to mission-critical enterprise work, “good enough” isn’t good enough. You wouldn’t use a Swiss Army knife for surgery unless you had no choice. You’d want a precision instrument.

That’s the difference with a purpose-built model like Language Weaver Pro. We’re not trying to do everything. We focus on doing one thing exceptionally well – delivering accurate, reliable translation at scale.

That level of specialization is becoming essential. The stakes are simply too high for generalist tools. Enterprises need translations they can trust – content that’s clear, actionable and accurate every time.

The collaboration with Cohere brings together language expertise and secure AI infrastructure. How is the system designed to balance performance, privacy, and deployment flexibility across cloud, on premise, and hybrid environments?

That collaboration is fundamental to our strategy because it was architected around a core principle that world-class AI must adapt to the client’s security posture, not the other way around. We refuse to force enterprises into a one-size-fits-all, public cloud model.

Our design achieves this balance by separating the AI model from the infrastructure it runs on. Cohere provides the state-of-the-art base model while being a world-class partner of ours, and we provide the linguistic expertise and the deployment flexibility that we’re already known for. This allows us to deploy these performant, fluent models wherever our clients’ data resides. For a bank or intelligence agency, that means running it completely air-gapped in their on-premises machines. For a healthcare provider, it can be in their dedicated private cloud to meet HIPAA regulations. And for a global tech company, a hybrid model can provide performance and flexibility across regions, while maintaining data security where they need it.

The key is that the client is always in control. They get the full performance of the industry’s largest, most advanced translation model without ever having to compromise on their data privacy or compliance mandates. We deliver the power to their perimeter.

One of the biggest risks for enterprises is translation that appears correct but contains subtle errors. How does your platform identify and reduce these risks, especially in high stakes scenarios?

This is a fundamental flaw of general-purpose AI, and as a company that specializes in Language Intelligence, we understand that this is an unacceptable risk. Plausible-sounding nonsense is a liability, not an asset.

We seek to mitigate this risk in several ways. First, our models are not trained on the unfiltered public internet. They are built on a foundation of four decades of high-quality, human-translated, and domain-specific linguistic data. This teaches the model to be right, not just fluent. It learns the correct way to translate complex legal clauses and precise medical instructions because it has learned from clean data, not internet chatter.

We also give our clients direct control over the output. Through terminology management and feedback, they can lock in the correct translation for their most critical and ambiguous terms. It ensures that a specific patent claim or a drug interaction phrase is rendered with the correct translations, each time.

Finally, we close the loop. Our systems are designed to integrate human expertise for review and refinement. This constant feedback from expert linguists and the client’s own reviewers creates a cycle of continuous improvement, catching the very nuances and subtle errors that generalist models are blind to.

With strong competition from tools like DeepL and Google’s Gemini, do you see the future of AI translation shifting toward domain specific models, and what does that mean for the broader AI ecosystem? 

Let me answer this with an anecdote. We have a customer, a major name in computing, who works with multiple translation providers. We work with the majority of their portfolio, and one of the key reasons they choose us is that when something isn’t right, we can fix it. Korean technical terminology not quite landing? We have a team for that. German numerals causing problems? We have people who can help.

That’s the broader point. We expect future progress in AI to come less from pure model advances, and more from solutions that can leverage relevant context – linguistic, cultural, and domain-specific. RWS, through its extensive product portfolio, decades of proprietary data, and deep customer relationships, has access to exactly that kind of context. General-purpose providers can match a model benchmark; but they cannot replicate that. The future belongs to providers who combine frontier AI with the humans and institutional expertise to make it work in the real world – and that is precisely where we sit.

Thank you for the great interview, anyone wishing to learn more should visit RWS.