What GRC 2026 taught us about AI translation quality

When it comes to translation, speed has been solved: artificial intelligence (AI) gets localized content to market faster than any team ever has. Yet speed without a quality foundation creates its own problems, from terminology drift and tone that shifts between markets to brand terms that mean something different in Japanese than they do in English — and often, nobody catches these problems until something ships wrong.

The more difficult problem is trust. Can you stake your brand on what AI produces? Can you put your help center in 21 languages and know it's right? Can you ship a product update to global users and feel confident that it’s accurate?

That's the conversation Global Ready Conference 2026 was built around. Localization leaders, marketing practitioners, product teams, and learning and development (L&D) professionals spent the day asking these questions from every angle — and the answer that emerged, across every session, is that trust isn't a review step. It has to be built into the process, and that means it has to be built into the platform.

Smartling launched the largest AI innovation release in company history at this year’s Global Ready Conference. Every capability in it is built around the same principle: quality should be automatic, continuous, and measurable. Not something teams scramble to verify after the fact.

The launch: quality built into every layer

Smartling VP of AI Olga Beregovaya and VP of Product Andrew Saxe opened the product roadmap session discussing how AI translation volume is growing faster than any team's capacity to oversee it. The gap between what gets produced and what gets reviewed is the place where quality often breaks down.

Solving that challenge is the centerpiece of Smartling's launch — LQA Agent, an AI-powered quality evaluation that applies industry-standard Multidimensional Quality Metrics (MQM) to score translations automatically and at scale, providing continuous quality monitoring that surfaces where additional attention is actually needed and tracks trends over time.

Beyond LQA Agent, the release also includes:

Øjeblikkelig AI-oversættelse inside connected tools including Figma, Adobe Creative Suite, and AEM, so your translation memory, glossaries, and Style Rules for AI apply to every translation.
Auto Select LLM, which handles model selection, prompt engineering, and continuous benchmarking behind the scenes for higher-quality output with no workflow changes required. Built-in RAG ensures every translation pulls from translation memory, glossaries, and AI Style Rules for higher-quality output than traditional MT, with zero setup and no model lock-in.
Stilregler for AI, which applies punctuation, date formats, regional language standards, and optional brand voice or domain requirements across more than 30 locales to keep every output on-brand.
AI-billedoversættelse, which extracts text from images, translates it, and rebuilds the images with the translated text in place — extending the same quality standards you expect for text-based assets to images like product photos, banners, and social assets.
Sproglig tilpasning to close regional variant gaps, like US English to UK English, without the cost of full retranslation.

Every piece is built on the same foundation. A translation platform where quality is the default — embedded in the engine, the rules, the output, and the evaluation.

What the day's sessions taught us about quality

What goes wrong when speed outruns structure? Olga and Alex Yanishevsky, Senior Director of AI and MT Solutions, ran the breakout that tackled this directly.

Ultimately, they shared that speed without governance creates its own problems, including inconsistent terminology, no audit trail, and the potential for hallucinations that surface at exactly the wrong moment. The session laid out what AI governance actually means in practice: reliable data and robust quality checks, detection and monitoring, transparency and auditability, and risk mitigation.

In Smartling's implementation, that looks like hallucination detection that surfaces AI-introduced errors, quality scoring that routes flagged content to human reviewers, and LQA Agent providing continuous monitoring so teams can track quality trends over time.

The AI localization framework that came from hard experience

Inside SumUp's operation, Jasmin Barth, Senior Localization Manager, and Maja Milosavljevic, Global Localization Manager, shared the content matrix their team built to route translation decisions at scale, with quality as the criterion at every branch.

The insight behind the framework is that content type determines quality risk. High-visibility, regulated, and brand-critical strings need a different path than internal documentation or low-stakes notifications. Quality programs succeed when each content type gets the workflow it actually needs.

Their session didn't soften the failures that shaped what they built. They shared their journey of uncovering the workflows that broke in production, the assumptions that didn't hold, and the moments where a monitoring signal would have caught a problem before it shipped. What came through clearly is that doing the hard work early is what their program's strength is built on.

When a consulting firm recommended a generic LLM

Michelle Kerr, Director of Product Transformation at IHG Hotels & Resorts, shared her team's experience evaluating whether a generic large language model could handle localization across 21 languages and every owned channel.

The result? It couldn’t. Generic AI produces a translation, but it doesn't apply approved terminology, enforce style across markets, route content to the right quality tier based on visibility level, or provide an audit trail when something goes wrong. Every one of those requirements needs infrastructure, like a platform that brings together the linguistic assets, governance controls, and quality framework in one place.

IHG rebuilt its localization architecture to move translation upstream into the content management system (CMS). Content now routes to the right translation method at scale. Quality checks happen before content reaches guests, not after it ships. The moment IHG's team declined the generic LLM proposal was more than a technology decision: it was a decision about what quality actually requires.

What marketing teams don't see until quality has already drifted

Dervilla O'Reilly, Lead Web Manager of Localization and International Operations at Docusign, made the case that ungoverned AI translation is the quality risk marketing teams are most exposed to (and least prepared to catch).

The problem isn't that AI produces obviously bad translations. It's subtler, because without brand-trained workflows, glossary enforcement, and a feedback loop back into the platform, there's no signal when quality drifts away from the company’s guidelines: a brand term translated inconsistently in German; a tone that shifts in French; a phrase that lands differently in a regulated market. None of that shows up in a delivery report or a word count.

Her team's work involves two things most marketing localization programs skip: educating stakeholders on what ungoverned AI actually produces in practice, and adapting content before AI ever touches it. This way, quality in marketing localization stops being a downstream checkpoint and instead becomes a decision made before translation starts.

From 200 to 2,000 hours: where automation stops and quality starts

Bruno Gonçalves, Global Program Strategist and Learning Experience Head at IBM SkillsBuild, brought a number that reframed how L&D teams think about scale: IBM went from 200 to 2,000 hours of translated learning content with a smaller team than before.

As the team approached scaling to this volume, full automation wasn't the answer: the content was technical, some learners were as young as 14, and some languages required right-to-left support the tools weren't designed for. What worked instead was a human-in-the-loop system with explicit decisions about where AI translation could be trusted and where it couldn't. They built translation memory specifically for L&D content types, solved SCORM and Rise (common e-learning file formats) compatibility challenges that had no established playbook, and set clear rules for exactly when human validation was non-negotiable.

The quality lesson from IBM SkillsBuild is that scale doesn't mean removing humans from the workflow. It means putting them precisely where AI falls short, and building the infrastructure to know the difference.

The 95% nobody warns you about

Chris Dell, Senior Advisor at Next Chapter Advisory + Coaching, has been on both sides of the build-versus-buy decision at enterprise scale. His framing was that while the API call is the easy 5% of building a translation system, the other 95% — translation memory integration, glossary enforcement, compliance controls, connector maintenance, and retuning the prompt every time a new model ships — stays with your engineering team, permanently.

Every one of those components is a quality variable. Homegrown stacks have no translation memory, no continuous quality monitoring, and no audit trail. They have whatever your team has the time and capacity to build and maintain long-term, alongside everything else they're responsible for. At the end of the day, the real cost of building in-house isn't the initial development, it's the ongoing quality debt.

The Global Ready Awards

The conference closed with recognition for the programs doing localization the right way. This year's Global Ready Award winners: Atlassian (Excellence), Busuu (Accessibility), Docusign (Collaboration), Fortra (Acceleration), Kyndryl (Globalization), Netskope (Partnership), and the Commonwealth of Virginia / VITA (Trailblazer) — honored for onboarding 67 state agencies on the Globalt leveringsnetværk to make critical government content accessible to residents in their own languages.

What localization leaders are building toward

Anna Schlegel, Chief Executive Officer of Universalization.ai, opened Global Ready Conference by naming what most localization leaders already know but haven't fully reckoned with: the AI era hasn't made this work easier, it's just raised the bar. The leaders who thrive are the ones who own the trust question, because trust in what AI produces is now one of the hardest problems in global content. The people who can solve it are the ones with the deepest expertise in how language actually works across markets.

Anna’s core ideas emerged in every session. Every practitioner on stage had a version of the same story: AI made translations fast, and the new challenge became making sure the output was something you could stake your brand on. That's the ultimate value of Smartling's quality innovations — translation you can trust, at the speed AI demands, with quality monitored automatically at scale.

All eleven Global Ready Conference sessions are available on demand. Watch the launch, the practitioner sessions, and the full conversation at smartling.com/conference.