#USA #San Francisco #Translation #Automation #Intento

Intento’s 9th Annual Report on Translation Automation 2025

Intento has recently released its ninth annual report, titled The State of Translation Automation 2025. This report, which continues the tradition of analyzing machine translation industry trends, examines the latest developments in translation automation powered by artificial intelligence (AI).
The core of the report delves deep into how AI can enhance translation quality, tailoring it to meet specific business and technical demands. By utilizing this report, global enterprises can gain practical insights that aim to boost user satisfaction, promote the adoption of multilingual systems, and ensure that translation strategies and tools are finely tuned to their individual linguistic requirements.

A significant evaluation within the report involved a comprehensive analysis of 46 machine translation engines and large language models (LLMs) across 11 language pairs. This evaluation assessed these technologies against five critical enterprise requirements: general translation quality, terminology accuracy, tone of voice adherence, formatting integrity (tag handling), and full-text consistency.
The findings reveal a remarkable contrast between different translation approaches:
1. Off-the-shelf Models vs Customized Solutions:
- Off-the-shelf models include various Neural Machine Translation (NMT) systems like Amazon, Google NMT, and Microsoft, alongside several LLMs such as OpenAI and Anthropic.
- In contrast, the report shows that solutions specifically customized to meet defined requirements outperformed these generic engines. Surprisingly, human evaluators found it difficult to distinguish between AI-generated content and human translations, with some instances where AI translations were rated even higher than those produced by humans.
2. Multi-Agent Workflows:
- The report emphasizes that a multi-agent workflow—which delineates functions among different roles such as Translator, Reviewer, and Post-Editor—delivers superior results. These workflows integrate agents for maintaining terminology, adjusting tone, and performing post-edits, achieving the highest average performance ratings in nine out of eleven language pairs.
3. Error Reduction Through Clarity:
- A standout aspect of the findings is the striking reduction in errors. Baseline systems averaged between 10 to 15 errors per text, while solutions employing a requirements-based approach managed to reduce these errors to a mere 0 to 2—translating to at least an 80% reduction in errors, sometimes completely eliminating them.
- This stark contrast highlights the pressing need for customizing translations based on clear requirements—a shift from merely selecting an “engine” to developing solutions that adequately address specific translation needs.
Konstantin Savenkov, CEO and Co-founder of Intento, remarked, “The most telling indicator emerged from our human evaluation; reviewers often could not distinguish AI from human translation and sometimes rated AI translations higher. The multi-agent approach, which uses various AI agents to verify and test requirements, consistently provided the best performance. However, these agents necessitate language-specific customization, making the standardization of this process crucial to eliminating excessive custom engineering and debugging. Currently, this limits adoption to larger-scale applications.”

In summary, Intento’s State of Translation Automation 2025 report presents critical insights for enterprises aiming to improve their multilingual strategies through enhanced AI-driven solutions. Organizations can download the full report to leverage these insights and optimize their translation approaches, thus improving communication and collaboration across global markets.

Overall, the importance of aligning translation strategies with specific business requirements cannot be overstated. This is a key factor in lifting both the quality of translations and overall user satisfaction across diverse commercial applications.