On April 3, 2025, representatives from 14 organizations gathered at the beautiful FIFA Museum in Zurich for a series of roundtable presentations and discussions organized by Slator on behalf of DeepL. Titled “The Boom of ‘Bring Your Own Language AI’ – Opportunities and Risks,” the event brought together diverse stakeholders from the insurance, banking, government, life sciences, telecom, transportation, and international organization sectors.
Mordstein’s presentation focused on the rise of custom language AI within organizations, highlighting the language access challenges many face in global business. He noted that although English remains the dominant language of international commerce, only 20% of the world’s population speaks it fluently. This lack of widespread fluency creates costly communication issues, with some estimates suggesting that companies lose up to USD 55,000 per employee annually as a result.
Language-related hurdles affect about a third of businesses attempting to grow internationally, added Mordstein. AI is rapidly becoming the answer to such hurdles, with a substantial 60% of employees worldwide already using AI for work, including tasks like text generation (50%) and translation (20%). However, the use of unauthorized AI tools is prevalent, with 55% of employees admitting to using them.
Given these challenges, there is a critical need for secure, enterprise-grade AI solutions that offer reliability, personalization, data protection, and seamless integration with existing systems. DeepL’s specialized AI is a decidedly superior alternative to general AI in solving those needs, commented Mordstein. “Unlike general AI, which relies on public data and carries ethical and legal risks, specialized AI is trained on company-owned, secure, and legally compliant data. This approach ensures higher quality results, reduces errors, and is tailored to specific business workflows,” added Mordstein.
Specialized AI solutions can help overcome many challenges while realizing significant cost savings, increased efficiency, improved communication, and enhanced data security. Furthermore, DeepL’s technology outperforms competitors like ChatGPT-4 and Google Translate, offering features like DeepL Voice and DeepL Write specially tailored to the needs of thousands of businesses. Mordstein shared that the ultimate goal is to facilitate international operations and deliver significant economic benefits, including potential savings of up to EUR 3m (USD 3.4m) and a 90% reduction in translation time.
Faes’s presentation highlighted a trajectory of machine translation (MT) capabilities over the past 20 years, moving from translating single phrases to entire company communications to today’s advancements in sophistication, quality, and scale. This evolution has consistently shaped user perceptions and expectations of the technology, from unusable to present-time near-perfect AI-enabled output.
Comprehensive multilingual communication management is also becoming increasingly feasible through AI, as is the integration of translation capabilities into various applications, platforms, and workflows (Translation-as-a-Feature, TaaF). “Rather than existing as standalone tools, users can embed translation services within their existing digital environments,” explained Faes.
Faes also highlighted the distinction between Business-to-Consumer (B2C) and Business-to-Business (B2B) contexts for language AI. DeepL’s solutions, for example, consider the specific needs and requirements of these two sectors when developing and deploying its AI translation solutions.
A consensus developed among attendees about the relevance of efficient translation for day-to-day business. Some were long-time users of DeepL, while others were curious about DeepL and Slator research. AI is increasingly being used to address cost, time-to-market, system integration challenges, and other issues, and organizations require an AI-enabled translation solution that is also secure and reliable and can be customized for their specific needs.
Specialized AI translation trained on proprietary data offers a significant advantage over general AI in terms of accuracy, security, and relevance for multiple use cases. DeepL emerges as a complete solution, addressing cost and operational concerns as AI translation reaches ever-higher quality levels, requiring minimal or no human intervention. As language AI today encompasses multiple modalities, use, interest, and demand continue to increase exponentially.
Next on the horizon is a more integrated and versatile approach where technology is expected to handle increasingly diverse content formats and complex real-time multilingual interactions across various media, all with high accuracy. To find out more about a custom language AI solution from DeepL for your organization, start a conversation today.
Meanwhile, in the world of artificial intelligence, Meta Platforms (NASDAQ:META) recently unveiled a suite of new AI models that push the boundaries of machine perception and language understanding. This suite includes the Perception Encoder, Perception Language Model (PLM), Meta Locate 3D, Dynamic Byte Latent Transformer, and Collaborative Reasoner, each designed to tackle complex challenges in their respective fields.
The Perception Encoder stands out for its ability to interpret visual information from images and videos, surpassing existing models in zero-shot classification and retrieval tasks. It has demonstrated proficiency in difficult tasks, such as identifying animals in their natural habitats, and has shown significant improvements in language tasks after integration with a large language model.
Meta’s PLM is an open-source vision-language model trained on a combination of human-labeled and synthetic data, designed to handle challenging visual recognition tasks and comes in variants with up to 8 billion parameters. The PLM-VideoBench, a new benchmark released alongside the PLM, focuses on fine-grained activity understanding and spatiotemporally grounded reasoning.
In robotics, Meta Locate 3D represents an innovation in object localization, enabling robots to understand and interact with the 3D world using natural language prompts. This model can accurately localize objects within 3D environments, a crucial step towards more autonomous and intelligent robotic systems. Meta has also released a dataset to support the development of this technology, which includes 130,000 language annotations.
The Dynamic Byte Latent Transformer is another groundbreaking model from Meta, designed to enhance efficiency and robustness in language processing. This byte-level language model architecture matches the performance of traditional tokenization-based models and is now available for community use following its research publication in late 2024.
Finally, the Collaborative Reasoner framework aims to develop social AI agents capable of collaborating with humans or other AI agents. It includes a suite of goal-oriented tasks that require multi-step reasoning and multi-turn conversation. Meta’s evaluation shows that current models can benefit from collaborative reasoning, and the company has open-sourced its data generation and modeling pipeline to encourage further research.
As Meta integrates these advanced AI models into new applications, the potential for more capable AI systems across various domains is set to expand, marking significant progress in artificial intelligence research and development.
