Earley AI Podcast - Episode 93: AI Translation, Brand Voice, and Global Content with Olga Beregovaya

Why the Gap Between an AI Translation Demo and Enterprise Production Is Wider Than Most Organizations Realize Guest: Olga Beregovaya, VP of AI at Smartling Host: Seth Earley, CEO at Earley Information Science Published on: June 17, 2026 In this episode, Seth Earley speaks with Olga Beregovaya, VP of AI at Smartling, who brings 25 years of experience across every major evolution in natural language processing - from rules-based systems through statistical models, neural translation, and now LLMs. They explore why plugging into a commercial model at token-level pricing is not a translation strategy, how brand voice fractures at 300,000 employees, why information architecture is just as essential for language pipelines as it is for retrieval, and what it actually takes to deliver consistent, on-brand, multilingual content at enterprise scale. Olga shares candid and specific insights on language complexity, the human-in-the-loop imperative, and why the organizations that are finally succeeding with AI have stopped treating it as art for art's sake. Key Takeaways: The price of a commercial model's tokens is not the cost of enterprise AI translation - data integrity, pipeline architecture, linguistic assets, and human review are the real cost drivers. Brand voice fractures the moment every employee can generate content autonomously - a Fortune 10 company discovered it had 300,000 voices overnight after deploying a co-pilot tool. Information architecture is equally essential for language pipelines as for retrieval - nested HTML tags, tokenization failures, and unstructured content break translation before the model ever sees the text. LLMs unlocked context that neural machine translation never had - resolving pronouns, disambiguating terminology, and working at document level instead of sentence by sentence. The assumption that AI translation works equally across all languages is one of the most dangerous misconceptions in the space - morphological complexity, writing systems, and training data representation vary enormously. Human review is not optional even in fully automated pipelines - it is how models learn, how ground truth is established, and how brand consistency is maintained over time. The organizations now succeeding with AI translation have moved from implement-and-fail to measured deployment - defining use cases, respecting prerequisites, and matching tooling to actual requirements. Insightful Quotes: "Yes, you can totally consume your million tokens at a super low price point, but what exactly are you buying for this money? Everybody can totally produce a translation or generate copy, but is it going to represent your brand? That's a different question." - Olga Beregovaya "He installed a co-pilot tool and said, it's great, except my company has 300,000 employees and now my company has 300,000 voices. That's not necessarily what I was prepared for in different countries." - Olga Beregovaya "If you want your models to evolve, and if you want your models to learn, you obviously need somewhere for these models to learn from - and this is where human review comes in. It is always twofold: guaranteeing the quality to your customers, and helping your models evolve." - Olga Beregovaya Tune in to discover why AI translation at enterprise scale requires far more than a model and an API key - and what the organizations getting it right have built that their competitors have not. Links LinkedIn:   / olga-beregovaya-04b5   Website: https://www.smartling.com