The quality problem that gets harder as AI translation gets faster

AI translation has solved the speed problem. Enterprise teams can now translate millions of words in hours rather than weeks. But monitoring translation quality at that volume is a different challenge entirely, and one that manual processes were never designed to handle. The answer is automated quality scoring built directly into the translation workflow: AI agents that evaluate output by error type and severity using Multidimensional Quality Metrics (MQM) scoring, operating at the same speed as the translation itself rather than as a review step that follows it. Smartling's recognition as the winner of the Machine Translation Innovation Award in the 9th annual AI Breakthrough Awards program reflects the company's work on exactly this problem. The award, selected from more than 5,000 nominations across more than 20 countries, recognizes the infrastructure Smartling has built to make AI translation trustworthy at enterprise scale, not just fast.

Why translation quality monitoring breaks at volume

Traditional linguistic quality assurance (LQA) works by having trained reviewers evaluate a sample of translated content, record errors against a defined taxonomy, and generate a quality score. It is a rigorous process, but it was designed for programs translating tens of thousands of words per month, not tens of millions.

At enterprise scale, three things happen that traditional LQA cannot address:

• Coverage becomes statistically insufficient. Sampling 5 to 10 percent of content might catch systematic errors, but it misses the string-level issues that most affect customer experience. A hallucinated product claim in a single high-visibility page can cause real damage regardless of how well the rest of the program performs.
• Turnaround time creates a publication lag. If human LQA review adds days to a workflow, the content cadence that makes AI translation valuable is undermined. Speed and quality become a tradeoff rather than a combination.
• Feedback does not reach the AI in time. When quality issues are identified after content is published, the AI that produced the problematic output has already moved on. The correction exists in a review document, not in the system that will handle the next translation job.

The solution is not more human reviewers. It is quality monitoring that operates at the same speed and scale as AI translation itself.

AI agents that monitor quality at every step

Smartling's approach uses a suite of AI agents that each handle a specific quality function within the translation workflow. Each agent makes autonomous decisions at its step but loops in humans for high-stakes reviews.

LQA Agent: instant MQM scoring without human involvement

The Linguistic Quality Assurance (LQA) Agent provides instant quality review on translated content using MQM scoring that matches human judgment, without requiring a human reviewer to work through a sample. The LQA Agent evaluates translations continuously and flags content that falls below defined quality thresholds, providing quality signal that informs routing, escalation, and program reporting at a pace that keeps up with AI translation output. For context on what that MQM score means in practice: Smartling's AI-Powered Human Translation (AIHT) consistently achieves MQM scores of 98 or above, exceeding the 95 to 97 industry benchmark for traditional human translation from most language service providers.

AI Post Editing Agent: catching errors before they reach customers

The AI Post Editing Agent operates earlier in the workflow, fixing grammatical errors, adjusting formality, and detecting hallucinations before content advances to publication or human review. This agent addresses the quality issues most common in large language model (LLM)-generated translation: fluency errors, tonal inconsistencies, and fabricated-but-plausible output that makes LLM hallucinations particularly difficult to catch through reading alone.

TM Optimization Agent: quality through consistency

Translation quality is not only about catching errors in new translations. It is also about ensuring that approved translations are reused consistently across the program. The TM Optimization Agent analyzes translation memory and automatically approves high-quality past translations for reuse, increasing consistency and reducing the terminology drift that occurs when the same product feature or regulatory term appears differently across markets.

Auto Select: the right engine for every string

Quality monitoring after translation is easier when the first-pass output is as strong as possible. Smartling's Auto Select routes each string to the highest-performing engine from a pool of more than 20 LLMs and machine translation engines including Amazon Bedrock, Microsoft Azure, Google Vertex AI, OpenAI, and DeepL, selecting by language pair and content type automatically. Starting with the best available output for each string reduces the volume and severity of quality issues that downstream agents need to address.

What AI Breakthrough recognized

The AI Breakthrough Award for Machine Translation Innovation recognizes Smartling's work on building translation infrastructure that is trustworthy at enterprise scale. Steve Johansson, managing director at AI Breakthrough, described the distinction directly:

"Smartling turns translation into true infrastructure rather than a standalone service. Trained NMT cannot keep up with global teams' need to translate volumes of multilingual content at enterprise scale. Traditional fragmented processes lead to lost revenue, inconsistent quality, and limitations in automating translations across tools and markets. With Smartling, there's never been an easier or smarter way to translate. Smartling's AI translation software combines automation, quality control, and security in an end-to-end workflow with optional human validation for high-quality global content delivery."

AI Breakthrough is part of Tech Breakthrough, a market intelligence and recognition platform for global technology innovation. The program has previously recognized companies including Nvidia, Glean, Databricks, and EY.

Embedding localization in AI workflows: the Smartling MCP Server

Alongside its AI agent capabilities, Smartling has launched an MCP (Model Context Protocol) server for the translation industry. The MCP server lets enterprises embed localization directly into their AI applications and developer workflows, giving engineering teams access to Smartling's translation capabilities, linguistic assets, and quality checks from within tools like Claude Code or custom AI assistants, without context switching to a separate platform.

This extends the quality monitoring story beyond the Smartling platform itself. Organizations building AI-powered products can incorporate Smartling's translation and quality infrastructure directly into their development pipelines, treating localization as part of how their product is built rather than a downstream process that follows it.

What this means for localization programs

For localization leaders evaluating how to scale translation quality monitoring as AI output volumes grow, the practical question is whether the quality infrastructure keeps pace with the translation infrastructure. Speed and volume are largely solved. The programs that maintain quality at that speed are the ones where monitoring is continuous rather than periodic, automated rather than sampled, and embedded in the workflow rather than applied after the fact.

The combination of the LQA Agent, AI Post Editing Agent, TM Optimization Agent, and Auto Select in a single integrated workflow is designed to deliver exactly that. Enterprise teams using Smartling's AIHT have seen meaningful results: IBM cut localization time in half and improved translation quality by 40 percent across more than 170 countries, while a Fortune 500 software company saved $3.4M in translation costs in a single year while maintaining quality across more than 20 million words annually.

See how Smartling's AI translation platform works at smartling.com/demo.