Most machine translation (MT) engines perform the same on day one as on day one thousand. They do not learn from your content, adapt to your terminology, or get better the more you use them. Google Translate and DeepL are trained on broad multilingual data. They produce competent general-purpose translations, but they have no mechanism to improve based on your specific preferences. 

How Opal Learns From Your Content 

When a client processes content through Welocalize’s Agentic AI platform, Opal, the system captures the relationship between source segments and their final, human-approved translations. These approved decisions accumulate over time, building a body of quality signals specific to that client’s content, terminology, and style. 

On subsequent jobs, Opal draws on these accumulated signals during processing. If a similar source segment comes through, the system uses the learned patterns to inform its output. The result is translation that reflects the client’s established preferences rather than a generic engine’s best guess. 

This goes deeper than traditional translation memory (TM). A TM match retrieves an old translation. Opal’s feedback loop changes how new translations are generated, not just which old ones get retrieved. The learning is embedded in the processing itself. 

What the Benchmark Found

In the Benchmark (five MT systems, 10 locales, seven client accounts), this feedback-driven learning mechanism was the one hypothesis the evaluation fully accepted. It consistently lifted quality across every system it was paired with, every locale, and every content type. No other variable produced such uniform results. 

“Fully accepted” is a specific statistical claim. The quality improvement was significant across all tested configurations, with no exceptions. Other variables in the benchmark showed mixed results depending on language pair or content type. The feedback mechanism had no edge cases. The lift was present in Germanic languages, Romance languages, Asian languages, technical content, and marketing content. 

Why this is Different From Custom MT Training

Custom-trained MT engines address part of the terminology problem. You can train a DeepL or Google model on your data. But training is expensive, requires engineering resources, and needs to be refreshed as content evolves. It is a point-in-time optimization. Opal’s feedback loop is continuous: every job processed adds to the system’s understanding of your content. Quality improves without any manual training step. 

The Compounding Effect 

A client who has processed 500,000 words through Opal has a meaningfully richer feedback history than one at 50,000 words. The system’s output for the higher-volume client will be more consistent and terminologically accurate. This is also, honestly, a switching cost: the accumulated quality advantage is specific to Opal. Whether you view that as a feature or a lock-in risk depends on your perspective. The data says the accumulation works. 

Generic MT Closes the Gap 

The benchmark also showed that generic (off-the-shelf) MT engines combined with Opal matched or outperformed expensive custom-trained MT engines combined with Opal. Once Opal’s feedback loop and processing layers are in the pipeline, the quality difference between generic and trained MT disappears. If you are spending money on custom MT training, this finding has direct cost implications. 

What this Means for Your Program

Ask yourself: is your current MT configuration getting better as you use it, or is it the same today as it was six months ago? Most MT setups are static. Opal, through its feedback-driven learning, is not. 

The benchmark confirmed this across a large, diverse dataset. Feedback-driven learning is the one finding in the entire evaluation that held without exception across every tested variable. In a field where MT quality comparisons tend to be hedged with caveats, that kind of clean result is unusual. 

The full benchmark results are available in the Opal in Production report here.