Intento has released its 9th annual industry report, The State of Translation Automation 2025 (formerly State of Machine Translation).
The report analyzes the latest advances in translation automation and shows how AI improves translations to meet specific business and technical requirements. It gives global enterprises practical guidance to lift end-user satisfaction, drive adoption of multilingual systems, and align translation strategy and tooling with their specific language requirements.
Intento analyzed 46 machine translation engines and LLMs (large language models) across 11 language pairs, conducting full-text evaluations against five enterprise requirements: general translation quality, terminology, tone of voice, formatting (tag handling), and full-text consistency.
The report compares:
- Off-the-shelf models—both NMT systems (e.g., Amazon, Baidu, DeepL, Google NMT, Microsoft, Oracle, Tencent) and LLMs (e.g., Anthropic, Cohere, Gemini, LLaMA, OpenAI),
- The same models adjusted to meet specific requirements using available customization options (excluding fine-tuning),
- A multi-agent workflow (Translator → Reviewer → Post-Editor),
- Human translation.
What the data shows
- Requirements-based translation outperforms generic engines: Solutions customized to specific requirements outperformed off-the-shelf models. Human evaluators often couldn’t distinguish AI-generated from human translations—and in some cases rated the AI higher.
- Multi-agent workflows deliver the best results: A multi-agent solution (with separate agents to verify requirements and test outputs, designed to avoid compounding hallucinations) produced the highest average performance, earning “best” ratings in 9 of 11 language pairs. These workflows combine agents for terminology integration, tone adjustment, and post-editing to consistently raise quality.
- Clear requirements, fewer errors: In testing, baseline systems averaged 10–15 errors per text. Requirements-based solutions cut that to 0–2—at least 80% fewer errors, and in some cases eliminating them entirely. This gap makes requirements-based customization essential for professional translation.
“Automatic translation is no longer about choosing the ‘engine’—it’s about building solutions that meet specific translation requirements,” said Konstantin Savenkov, CEO & Co-founder of Intento. “The most telling indicator emerged from our human evaluation: reviewers often couldn’t distinguish AI from human translation—and sometimes rated AI translations higher. The multi-agent approach, utilizing multiple AI agents to verify and test requirements, delivered the best average performance, as expected. However, these agents require language-specific customization, and standardizing this process is critical to prevent excessive custom engineering and debugging overhead that currently limits adoption to only large-scale applications.”