## Overview Smartling is an enterprise-scale AI-first agentic translation delivery platform founded in 2009 that has evolved from statistical and rule-based machine translation approaches to sophisticated LLM-powered systems. The platform serves Fortune 50 and Fortune 100 companies including Disney and IBM, handling massive volumes of diverse content types ranging from UI strings to legal documents, marketing materials, user-generated content, and multimedia. The company's VP of AI, Olga, brings extensive experience in computational linguistics and has been with Smartling for over three years, initially hired as VP of Machine Translation before the role evolved to VP of AI to reflect the broader technological shift in the industry. The platform addresses five core enterprise challenges: automation of manual translation processes, centralization of linguistic assets and workflows across departments and country offices, compliance and data governance (especially preventing exposure of corporate content to public LLMs), brand voice consistency across locales, and the ability to parse and handle diverse content types from different sources. Smartling positions itself as a "buy versus build" solution, noting that many enterprises initially attempt to build in-house translation systems but find them financially unviable once they account for the necessary data scientists, computational linguists, engineers, and in-country linguists required for proper multilingual operations. ## Agentic Workflow Architecture The core technical innovation in Smartling's approach is their fully agentic workflow architecture. Rather than relying on a single LLM pass, the platform implements multi-step processes where different functions or instances of models check and validate the outputs of previous steps. This architecture addresses one of the fundamental challenges with out-of-the-box LLM translation: inconsistency, hallucinations, and unexpected behaviors like adding emojis or inappropriate tone adjustments that make raw LLM outputs unsuitable for enterprise production use. A particularly important component is their automated post-editing (APE) step, which validates the initial translation output. This is where hyper-localization occurs—adapting content not just to broad language groups (like "Latin American Spanish") but to specific demographics in countries like Chile or Peru. The APE step also performs hallucination checking and adapts formal/informal tone appropriately for target languages. For instance, German and Japanese require careful differentiation between formal and informal registers that may not have direct English equivalents, and the system needs to avoid both overly casual ("drunk sailor") and overly formal ("writing to the King of England") extremes. ## Model Selection and Hybrid Approaches Smartling takes a pragmatic multi-model approach rather than committing to a single vendor or technology. Their pipeline incorporates models from OpenAI (GPT-4, GPT-5), Anthropic (Claude), Google (Vertex AI Model Garden), Mistral, IBM Watson, and AWS, selecting models based on fit-for-purpose criteria for specific languages and content types. This flexibility is crucial because their benchmarking reveals that different models excel at different tasks. Importantly, Smartling has not abandoned neural machine translation (NMT) in favor of pure LLM approaches. Their benchmarking across eight automated metrics (BLEU, COMET, METEOR, METRICX, among others) plus human judgment correlation shows that NMT still has significant value, particularly for lower-resource languages and high-risk content where deterministic, predictable outputs are prioritized over fluency. The platform intelligently routes content through either NMT or LLM pipelines based on these considerations, representing a measured hybrid approach rather than wholesale replacement. The platform also accommodates customer-built models, allowing enterprises that have invested in fine-tuning their own models to plug them into Smartling's infrastructure, which then compensates for missing capabilities in the overall workflow. This hybrid "build and buy" model acknowledges that some enterprises have valuable custom models while still lacking the complete infrastructure for production-scale translation operations. ## Data Management and Training Data quality and management are foundational to Smartling's approach. The platform emphasizes that "garbage in, garbage out" remains a fundamental truth even with advanced LLMs. For custom model training, they've identified that a healthy starting point is approximately 10,000 translation segments (translation units averaging 10 words each, so roughly 100,000 words total) to see meaningful impact from customer-specific corpus when fine-tuning generic foundational models or NMT engines toward customer-specific domains and terminology. The platform addresses the challenge that enterprise data is often messy, even at large scale. User-generated content, knowledge bases authored by multiple engineers without technical writing standards, and content from disparate sources all create training data challenges. Smartling's value proposition includes centralizing and cleaning this data, performing source analysis, hallucination mitigation, and source pre-editing and optimization before translation begins. The system leverages RAG (Retrieval-Augmented Generation) extensively to pull from validated local-specific sources, customer glossaries, legacy content, and brand style guides. This grounding mechanism helps mitigate hallucinations and ensures consistency with established brand voice and terminology. The more validated sources the system can fetch from, the higher the likelihood of producing relevant, accurate content that adheres to enterprise standards. For lower-resource languages where insufficient training data exists, Smartling employs synthetic data generation and can hire linguists to build up corpus to minimum viable thresholds for model training. This addresses the reality that foundational models are heavily English-centric due to training data availability, and languages like Urdu have dramatically less representation than languages like Latin American Spanish. ## Prompt Engineering and Complex Prompting Smartling emphasizes the critical role of sophisticated prompt engineering in achieving enterprise-quality translation. The platform employs computational linguists and data scientists who invest significant effort into complex prompting strategies that incorporate multiple considerations: target language grammatical complexity, formality registers, brand voice guidelines, terminology constraints, and content type-specific requirements. The prompting approach is context-aware and adaptive. For example, the system can reference stop lists (like China's government-mandated restrictions on superlatives in marketing content), inject customer-specific style guides, and adapt to different content types from UI strings to legal documents. This level of sophistication is presented as the key differentiator from enterprises that attempt simple prompt-plus-LLM approaches and find them inadequate for production use. There's also an acknowledgment of the "faster, better, cheaper" triangle that still exists in LLM translation. The R&D team negotiates trade-offs: faster translation requires pruning prompts to fewer tokens and potentially cutting agentic steps, while better translation demands 3,000+ token prompts with more comprehensive agentic validation. While the triangle has narrowed compared to three years ago, and the platform aims to continue shrinking these trade-offs, they remain a practical consideration in production systems. ## Quality Evaluation and Benchmarking Smartling employs rigorous quality evaluation methodologies combining multiple automated metrics with human judgment. They benchmark across eight different automated translation quality metrics including BLEU, COMET, METEOR, and METRICX, while also measuring correlation with human judgment. This multi-metric approach provides a more robust quality assessment than relying on single metrics that might miss important aspects of translation quality. The platform tracks what they call "edit distance"—the measure of how much human post-editing is required to bring translations to publication quality. For more structured content types and easier-to-translate languages (like Latin American Spanish), edit distance is shrinking and can reach 70% human parity. This represents a significant achievement, though the case study appropriately notes that results vary substantially by language pair and content type, with Southeast Asian languages proving more challenging than Romance languages due to grammatical complexity and training data scarcity. An interesting finding from their benchmarking is that the leap from GPT-4 to GPT-5 showed minimal improvement for translation tasks—only a couple of BLEU points and COMET points—suggesting that current transformer-based models have hit a plateau in translation quality. This observation informs Smartling's view that the next generation of improvements will come more from reasoning models and better application layer capabilities (prompting, RAG, fine-tuning, agentic workflows) than from raw model improvements. ## Storage and Reuse Strategy The platform implements a sophisticated strategy for storing and reusing translation outputs, which has important cost and latency implications. Traditionally, only human-vetted translations were stored in translation memory systems for reuse. Smartling has evolved this to store AI-generated translations that have passed through their agentic validation pipeline and been vetted to human quality standards. This approach saves costs by eliminating the need to re-run content through the entire pipeline and re-spend tokens on previously translated segments. However, the system discriminates between AI outputs that have been fully validated versus those that still require human-in-the-loop review. Only the former are stored for automatic reuse, while content that needed human intervention stores the human-corrected version. This balanced approach maximizes efficiency while maintaining quality guarantees. The storage strategy also supports both just-in-time translation (generating translations on-demand as users request content) and pre-translation (generating and storing translations in advance). The choice depends on customer requirements around latency tolerance and content update frequency. ## Multimodal Capabilities While Smartling's core business has historically focused on text-based translation, they are expanding into multimodal capabilities to handle the full range of enterprise content. This includes automated speech recognition (ASR) for transcription, text-to-speech for generating multilingual voice content, and computer vision for handling complex formats like PDFs (with OCR), images, and Photoshop files requiring desktop publishing adaptations. The multimodal expansion addresses traditionally expensive and labor-intensive processes. Desktop publishing for multilingual images previously required extensive human work, but AI-generated multilingual images significantly reduce this burden. Similarly, parsing complex PDFs with mixed text and visual elements becomes more tractable with multimodal AI capabilities. The platform is working to consolidate these multimodal capabilities within unified models rather than maintaining separate acoustic models and language models as in traditional approaches. This consolidation reduces the number of processing steps and can improve latency, though there remain trade-offs between the speed of instant translation and the quality achievable with more comprehensive processing. ## Enterprise Scale Requirements Operating at true enterprise scale imposes requirements beyond raw translation quality. Smartling emphasizes repeatability—ensuring that what works in a lab setting consistently works in production across millions of translation segments. Scalability requires managing latency across massive corpus volumes and high request volumes. Compliance and data governance are critical, particularly preventing enterprise content from being exposed to public LLMs where it might become training data or violate NDAs. The platform serves customers with hundreds of thousands of employees (not merely thousands), handling diverse content types from UI strings to user assistance documentation, legal content, marketing materials, user-generated content, and multimedia. Each content type requires different handling strategies, and a true enterprise platform must accommodate all of them within a unified infrastructure. Smartling's approach to localization maturity helps them diagnose customer sophistication levels. Companies lower on the maturity curve typically have spreadsheets, Google Docs, and manual processes scattered across departments and country offices. More mature customers have centralized processes and understand their requirements, but may still struggle with optimization. The platform positions itself as a trusted advisor that helps customers understand their actual problems, which may not be immediately apparent when they first engage. ## Ownership and Organizational Integration An interesting aspect of the case study is the discussion of how translation ownership has evolved within enterprises. Historically, translation was treated as a cost center line item on P&L statements, with ownership scattered across product, engineering, marketing, or local sales depending on who was willing to absorb the expense. Translation was often an afterthought, with content "thrown over the fence" after development completion. Modern enterprises increasingly place translation and localization within product or engineering organizations for two key reasons: translation processes are now more deeply integrated into development environments and workflows, and "translation as a feature" has become a recognized product capability. AI accessibility has accelerated this shift, making translation more of a core product concern than a separate operational function. This organizational shift reflects the concept of "translation first" development, where multilingual support is considered from the beginning rather than bolted on afterward. Developers now understand concepts like pseudo-localization (using machine translation to preview how content will appear in various languages with different text expansion/contraction properties) and increasingly trust subject matter experts on language-specific complexities like right-to-left rendering, code-switching between languages, and grammatical constraints they may not be aware of. ## Business Outcomes and KPI Evolution The case study notes that translation KPIs have evolved dramatically from purely linguistic quality metrics toward business outcomes. Modern KPIs map directly to metrics like sales conversion rates in specific regions, measuring how translation quality impacts business results rather than just linguistic accuracy. This shift reflects the growing recognition that translation is a business driver, not just a cost center. The platform demonstrates concrete improvements across the traditional "faster, better, cheaper" triangle. Translation throughput increases from the human translator average of 2,000 words per day to 6,000-7,000 words per day with AI-first approaches. Costs drop by 4-10x compared to human translation. Quality, while still language-dependent, approaches human parity for structured content in easier language pairs, with around 70% edit distance for optimal cases. Interestingly, the case study acknowledges that the "faster, better, cheaper" triangle still exists and involves trade-offs, though the triangle has narrowed significantly compared to three years ago. The platform's roadmap for H2 2025 focuses on continuing to shrink these trade-offs. ## Critical Perspectives and Limitations The case study provides some balanced perspectives on limitations and ongoing challenges. Pure LLM approaches are acknowledged as non-deterministic, which creates issues for high-risk content requiring predictability. Neural machine translation maintains advantages in these scenarios despite lower fluency. The gap from GPT-4 to GPT-5 in translation quality suggests diminishing returns from raw model improvements, with future advances more likely to come from better application layer techniques. Lower-resource languages remain significantly more challenging than high-resource languages due to training data scarcity and grammatical complexity. Southeast Asian languages, for example, produce substantially different results compared to Romance languages. The platform acknowledges that achieving consistent quality across all language pairs remains an ongoing challenge. The case study also notes the broader industry context where approximately 95% of AI initiatives reportedly fail, often because organizations take a "silver bullet" approach rather than measured implementation with clear KPIs. Smartling positions itself against this trend by advocating for purpose-built vertical-specific models rather than expecting generic foundational AGI to solve all problems. ## Technology Evolution Context The interview provides valuable historical context on translation technology evolution. Starting from Levenshtein distance-based translation memories in the early 2000s, through rule-based machine translation (like Alta Vista Babelfish), then statistical machine translation, and neural machine translation in the 2010s, to current transformer-based LLMs. This trajectory mirrors broader NLP evolution and helps contextualize why certain technical decisions are made. The observation that current transformer models may have "hit a wall" in terms of pure quality improvements for translation—having consumed essentially all available training data and reaching parameter limits—suggests the next wave of improvements will come from reasoning models and better application layer capabilities rather than simply scaling existing architectures. This perspective informs Smartling's focus on agentic workflows, RAG, sophisticated prompting, and hybrid model approaches rather than betting everything on the next generation of foundational models.