Junior Hebrew Linguist

AI Post-Editing at scale: What a 71,262-Segment study reveals about the future of translation. What is the most effective way to apply large language models to professional translation? September 15, 2026 As organizations continue integrating AI into enterprise translation workflows, one question remains at the forefront of localization strategy: What is the most effective way to apply large language models (LLMs) to professional translation? Some are experimenting with using LLMs to translate content directly from the source language, while others are incorporating them into existing workflows to refine machine translation output. Both approaches promise improvements in quality and efficiency, but until recently there has been little production-scale evidence comparing how they perform in real-world localization environments. Those questions are explored in "AI Post-Editing in Production: A 71,262-Segment Evaluation Across Five Domains, Ten Languages and Five Systems," presented at the 2026 European Association for Machine Translation (EAMT) Conference by Mara Nunziatini and Mercedes Speroni of Welocalize. The research, which was published in Volume 2 of the proceedings, evaluates AI post-editing (AIPE) across more than 71,000 production translation segments spanning ten target languages and five content domains. Rather than relying on benchmark datasets or laboratory conditions, the study measures how different AI translation approaches perform in the complex environments that enterprise localization teams manage every day. The findings suggest that the greatest value does not necessarily come from replacing traditional translation workflows with LLMs. Instead, the strongest results emerged when AI was used to enhance existing translation processes with the help of enterprise knowledge, terminology, and linguistic guidance. Evaluating AI under real production conditions. Many published studies evaluate AI translation using relatively small datasets or a limited number of language pairs. Enterprise localization presents a far more complicated challenge. Content spans multiple domains, translation memories of varying quality, specialized terminology, and diverse business requirements. To reflect those realities, the researchers evaluated 71,262 production translation segments, representing more than 553,000 words translated from English into ten target languages. The dataset covered five distinct content domains, including Product and Service, Marketing, Support, and Medical Devices. In addition to automatic evaluation metrics, the researchers conducted a human assessment using 60 professional translators who reviewed more than 6,600 translation segments using the DQF-MQM quality framework and blind preference rankings. The evaluation compared five different systems: Production AIPE, Generic MT combined with AIPE, direct LLM translation, Google Translate, and DeepL. Generic MT + AIPE. Generic MT translates the text, then an LLM (gpt-4o) post-edits it using style guides and approved past translations. Production AIPE. Mirrors a real production pipeline: TMs are applied first, down to 75% fuzzy matches, then MT (generic or customized) fills the rest. The LLM post-edits this combined pre-translation using the same style guides and approved past translations. Llmt. AI Direct Translator. The same LLM translates from scratch, with no MT injection or TM leverage. It still gets the same style guides and past translations, at about 60% lower cost per segment. Google Translate & DeepL. Off-the-shelf generic MT, with no client-specific resources. This approach allowed the researchers to compare complete translation workflows rather than isolated models, providing a more realistic view of how organizations deploy AI in production. AI Post-Editing produced the strongest results. Across every automatic evaluation metric, the two AI post-editing workflows consistently outperformed the other systems. Production AIPE achieved the highest chrF, BLEU, and COMET scores while also requiring the fewest edits compared with professionally reviewed translations. Generic machine translation followed by AIPE performed nearly as well, while direct LLM translation produced only modest improvements over generic machine translation. These results highlight an important distinction. Rather than asking an LLM to generate every translation from scratch, AIPE starts with an existing translation and refines it using additional context. The workflow incorporates retrieval-augmented generation (RAG), language-specific style guides, human-reviewed bilingual examples, and enterprise terminology to improve accuracy, consistency, and brand voice before the content reaches final review. The findings suggest that, in production environments, the combination of machine translation and targeted AI refinement currently delivers stronger results than relying solely on direct LLM translation. Human evaluation reinforced the findings. Automatic evaluation metrics provide an important measure of translation quality, but professional linguists remain the ultimate benchmark for production content. To validate the automated findings, professional translators independently ranked the outputs without knowing which system had produced them. Their assessments closely matched the automatic metrics. Generic machine translation followed by AIPE received the highest percentage of first-place rankings, followed by Production AIPE. Google Translate and direct LLM translation occupied the middle tier, while DeepL ranked lowest overall. The consistency between automated measurements and expert human evaluation provides additional confidence that the observed improvements represent meaningful quality gains rather than statistical artifacts. For localization leaders evaluating AI investments, this agreement between human reviewers and automated scoring strengthens the case for incorporating AI post-editing into production workflows. Better inputs lead to better outputs. One of the more interesting observations from the study involves the relationship between the quality of the initial translation and the effectiveness of AI post-editing. Although both AIPE workflows relied on the same post-editing engine, Production AIPE generated a higher proportion of major and critical errors than Generic MT combined with AIPE. The researchers hypothesize that this difference may be due to the nature of the pre-translation inputs. Production AIPE processes a combination of machine translation and translation memory matches, including fuzzy matches. While fuzzy translation memory segments represented only 19 percent of the translated content, they accounted for 33 percent of all severe errors identified during evaluation. This suggests that inconsistent or lower-quality source material may limit the extent to which AI can improve the final translation. The finding serves as a reminder that AI is only one component of a larger localization ecosystem. The quality of translation memories, terminology resources, and other enterprise assets continues to influence overall translation performance. Content and language influence performance. The research also demonstrates that AI performance is not uniform across every type of content. Medical device translations consistently showed a higher concentration of major and critical errors than other domains, reflecting the precision required for regulated content. Marketing, product, and support content generally produced fewer severe errors across the evaluated systems. Similarly, although the overall ranking of systems remained consistent across all ten target languages, the severity and distribution of errors varied by locale. Some languages produced relatively few total errors but a greater concentration of major issues, reinforcing the importance of evaluating translation quality beyond simple error counts. Organizations deploying AI across multiple markets should consider both content type and language when determining where automation can be introduced most effectively. The future of AI translation is workflow-oriented. Perhaps the most important takeaway from the research is that successful AI translation depends on more than selecting the newest or most capable language model. The highest-performing systems integrated multiple technologies into a coordinated workflow. Machine translation, retrieval-augmented generation, translation memories, bilingual examples, style guides, and AI post-editing each contributed to the final result. Rather than replacing existing localization infrastructure, the LLM served as an intelligent refinement layer that leveraged enterprise knowledge to improve quality and consistency. As organizations continue adopting AI for multilingual content, workflow design may become an even greater competitive advantage than individual model selection. The research by Nunziatini and Speroni suggests that the future of enterprise localization lies not in choosing between machine translation and large language models, but in orchestrating them effectively within a production-ready workflow. By combining established translation technologies with AI-driven refinement, organizations can achieve higher quality while maintaining the scalability and governance that enterprise localization demands. Research sources. AI Post-Editing in Production: A 71,262-Segment Evaluation Across Five Domains, Ten Languages and Five Systems, by Mara Nunziatini and Mercedes Speroni, presented at the 2026 European Association for Machine Translation (EAMT) Conference.

How global brands are scaling content without losing brand voice. Enterprise AI translation has entered a new phase. May 28, 2026 Enterprise AI translation has entered a new phase. The conversation has shifted from whether AI can translate content quickly enough to how organizations can scale multilingual content production without damaging brand voice, increasing operational risk, or overwhelming internal review teams. That challenge continues to grow as generative AI accelerates enterprise content creation. Marketing teams are producing more campaigns, product pages, support documentation, and regional content variations than ever before. At the same time, global organizations are expected to deliver those experiences consistently across dozens of languages and markets. Traditional localization workflows and current AI solutions were not designed for that level of volume. Welocalize's Opal platform, which recently received the AI Excellence Award from Business Intelligence Group, reflects how the language industry is adapting through AI-enabled operational workflows designed specifically for enterprise-scale multilingual content. What is AI translation? AI translation has evolved far beyond traditional machine translation systems. Earlier generations of machine translation focused primarily on converting text accurately from one language to another. While those systems improved efficiency, they often struggled with nuance, tone, context, and brand alignment. Modern AI translation systems are increasingly built around orchestration rather than standalone translation output. Today's enterprise workflows may combine neural machine translation, generative AI post-editing, quality estimation models, terminology management, reinforcement learning, and human linguistic review within a single operational pipeline. This evolution reflects a broader industry realization: translation quality alone does not solve the enterprise challenge. Organizations also need governance, scalability, workflow intelligence, and brand consistency. Why AI translation has become a critical enterprise infrastructure layer. One of the biggest shifts happening in enterprise AI is the explosion of content volume. Generative AI tools now allow organizations to create content faster than traditional review processes can realistically support. Global marketing operations are scaling rapidly, yet multilingual governance often remainsfragmented across teams, vendors, and disconnected tools. As content production accelerates, enterprises are discovering that the operational bottleneck has moved. The challenge is not generating content. The challenge is validating, adapting, routing, reviewing, and maintaining quality across large multilingual ecosystems. This is where AI translation platforms are beginning to function less like translation software and more like enterprise infrastructure. Why brand voice matters in AI translation. One of the most persistent limitations of generic AI translation systems is brand consistency. Modern models can produce fluent output, but fluency alone does not create effective multilingual customer experiences. Enterprise brands invest heavily in tone of voice, messaging strategy, terminology alignment, and positioning. Those elements often become diluted when content passes through generalized AI systems without brand-specific optimization. For global organizations, this creates significant risk. A luxury hospitality brand, healthcare company, or technology platform cannot afford messaging that feels generic, inconsistent, or culturally disconnected across markets. This is why many enterprise AI translation strategies now rely on layered systems that incorporate brand-trained AI models alongside human linguistic expertise. Opal combines neural machine translation with generative AI post-editing trained on brand terminology and tone. Automated quality estimation then evaluates the output before human reviewers become involved, allowing organizations to prioritize expertise where it matters most. How quality estimation is changing localization workflows. Quality estimation has become one of the most important developments in AI translation operations. Historically, localization workflows evaluated quality after translation and human review had already taken place. That process increased turnaround times and applied similar review intensity across all content types, regardless of business impact. AI-driven quality estimation changes that sequence. Instead of waiting until the end of the workflow, quality signals can now be generated before human intervention occurs. This allows enterprises to make more intelligent routing decisions based on risk, complexity, and content importance. Lower-risk support content may move through workflows with limited human involvement, while high-visibility marketing campaigns or regulated content can receive deeper review from specialized linguists and subject matter experts. This approach helps organizations scale multilingual operations more efficiently while maintaining stronger governance standards. It also reflects a broader trend in enterprise AI adoption: organizations are redesigning workflows around AI capabilities rather than layering AI onto existing processes. What role do human linguists play in AI translation? Despite rapid advances in AI translation, human expertise remains central to enterprise localization. The most successful AI-enabled workflows are collaborative environments where AI handles scalability and repetition while humans provide judgment, cultural intelligence, and brand stewardship. Human linguists continue to play critical roles in cultural adaptation, regulatory interpretation, linguistic quality assurance, terminology alignment, annotation workflows, and reinforcement learning feedback loops. As AI systems improve, the value of human expertise becomes more specialized rather than less important. Many organizations are shifting language professionals toward higher-value activities focused on oversight, optimization, and domain expertise instead of repetitive execution work. How AI translation supports global content operations. AI translation is increasingly becoming part of a larger multilingual content ecosystem. Rather than operating as isolated translation workflows, enterprise localization systems are integrating with broader marketing, product, legal, and customer experience operations. This includes continuous multilingual publishing, AI-assisted content adaptation, regional performance optimization, multilingual SEO strategies, centralized terminology governance, and automated workflow orchestration. The long-term vision for many organizations is not simply faster translation. It is the ability to create adaptive global content systems capable of evolving dynamically across markets while remaining aligned to brand standards. According to the original article, this future may include AI systems that automatically optimize multilingual content performance regionally while preserving brand consistency and governance controls. What is the future of AI translation? The future of AI translation will likely be defined by orchestration, risk intelligence, and human collaboration. Organizations are moving away from one-size-fits-all automation strategies and toward risk-based operational models that combine AI scalability with targeted human oversight. This next generation of AI translation infrastructure is expected to focus on brand-trained AI systems, intelligent quality estimation, multilingual workflow automation, human-in-the-loop optimization, adaptive content generation, and enterprise-wide governance frameworks. For global brands, the goal is not maximum automation. The goal is scalable multilingual communication that still feels authentic, trustworthy, and aligned with the brand itself. As enterprise AI adoption continues to expand, that balance between efficiency and quality may become one of the most important competitive differentiators in global content operations.