A European insurance company needed to translate 2.3 million words of policy documents, claims forms, and customer communication templates into 14 languages within 90 days. The pure human translation quote came back at €1.4 million with a 7-month timeline. The pure machine translation option would have cost €28,000 and taken 48 hours, but the quality was unusable for regulatory-facing documents — terminology was inconsistent, policy conditions were mistranslated, and the formal register required by insurance regulators in each target country was completely absent. They went with MTPE. Final cost: €380,000. Timeline: 82 days. Quality: passed regulatory review in all 14 jurisdictions.
That’s the MTPE value proposition in one paragraph, and also its limitation in one paragraph. You get roughly 70% cost savings versus human translation, roughly 3-4x speed improvement, and quality that’s adequate for most business purposes but not equivalent to human translation at its best. The question isn’t whether MTPE works. The question is understanding what you’re trading off and whether those tradeoffs are acceptable for the content you’re translating.
What is MTPE, exactly, because the term gets thrown around loosely. Machine Translation Post Editing means taking the output of a machine translation engine — these days almost always a neural machine translation (NMT) system like Google Neural, DeepL, or a custom-trained engine — and having a human editor review and correct the machine’s output. The editor doesn’t retranslate from scratch. They start from the machine’s draft and fix errors, which is faster than translating from scratch because the machine gets the general structure right most of the time, and the editor is correcting rather than composing.
There are two levels of post editing, and the distinction matters because they produce different quality at different cost. Light post editing (sometimes called “rapid post editing”) focuses on fixing only errors that affect comprehension: wrong word choices that change the meaning, missing negation, obviously garbled syntax. The editor doesn’t improve style, doesn’t smooth awkward phrasing, doesn’t optimize for fluency. The output is understandable but not polished. It’s appropriate for internal documents, reference materials, and content where the reader will accept rough language as long as the meaning is correct.
Full post editing aims for quality comparable to human translation. The editor corrects all errors, improves style and fluency, ensures terminology consistency, and polishes the output to read naturally. Full post editing takes more time than light post editing — typically 60-70% of the time it would take to translate from scratch, versus 30-40% for light post editing. The cost savings are correspondingly smaller: maybe 30-40% versus human translation for full post editing, versus 50-60% for light post editing. But the quality difference is significant, and for external-facing content — marketing, legal, regulatory — full post editing is usually the minimum acceptable level.
The part that nobody in the MTPE industry likes to talk about is that the quality of the machine translation output varies enormously depending on the language pair, the domain, and the training data. Neural machine translation between English and French for general business content is quite good. The raw output might need 15-20% correction for full post editing quality. Neural machine translation between English and Japanese for pharmaceutical regulatory content is much worse. The raw output might need 50-60% correction, which erodes the time and cost savings significantly. In some cases, the post editor is retranslating so much of the source that MTPE is slower than translating from scratch, because the editor has to read both the source and the flawed machine output, decide what to keep and what to discard, and then rewrite the discarded parts. That’s cognitive overhead that doesn’t exist in direct translation.
I’ve seen MTPE projects where the post editors reported that the machine output was so bad for their language pair and domain that they were essentially ignoring it and translating from scratch, but the project was still being billed and reported as MTPE. The client thought they were getting a cost-optimized workflow. They were paying for human translation plus the overhead of processing bad machine output. The savings were negative. This happens more often than MTPE providers like to admit, and it’s the reason that any serious MTPE engagement needs to start with a quality assessment of the raw machine output for the specific language pairs and content domains involved, before committing to a workflow and a pricing model.
The engine selection question is where enterprise MTPE gets strategic. Off-the-shelf engines like Google Translate and DeepL are trained on massive general-purpose datasets. They perform well on common language pairs and general content, but they don’t know your company’s terminology, your industry’s jargon, or your regulatory context. A custom-trained engine — one that’s been fine-tuned on your translation memories, your terminology databases, and your previously translated content — produces significantly better raw output, which means less post editing effort, which means more actual cost savings. The upfront investment in engine customization typically runs $20,000-$80,000 depending on the language pairs and the volume of training data available. For enterprises translating more than 5 million words per year, the custom engine pays for itself within 6-12 months through reduced post editing costs.
The insurance company I mentioned earlier went with a custom-trained engine. They had 15 years of translation memories from previous human translation projects. The engine was trained on those memories plus their terminology database and regulatory glossaries. The raw machine output for their domain was good enough that full post editing took an average of 55% of the time required for from-scratch translation, which was better than the 60-70% industry average for full post editing and reflected the quality advantage of the custom engine. If they had used an off-the-shelf engine, the post editing time would have been closer to 70-75%, and the cost savings would have been marginal enough that the MTPE approach might not have been worth the complexity.
There’s a content triage decision that precedes the MTPE workflow, and it’s the most important strategic decision in any enterprise translation program. Not all content deserves the same level of translation investment. Internal operational documents might be fine with light post editing or even raw machine translation with no post editing. Customer-facing content needs full post editing at minimum. Legal and regulatory content might need full post editing plus a second human review. Brand-critical marketing content probably shouldn’t go through MTPE at all, because the stylistic compromises inherent in post-editing machine output are visible to readers who are attuned to brand voice.
The triage framework I’ve seen work best assigns each content type to one of four tiers: raw MT (internal reference only, no quality guarantee), light MTPE (understandable, not polished), full MTPE (comparable to human translation), and pure human translation (brand-critical, literary, or legally sensitive). The cost per word drops by roughly 50% for each tier down from pure human translation, so even shifting 30% of your content volume from pure human to full MTPE generates meaningful savings across a large translation program.
Artlangs Translation provides enterprise MTPE services across 230+ language pairs: custom engine training on your translation memories and terminology, light and full post editing by domain-specialist linguists, quality assessment of raw machine output before workflow commitment, and content triage consulting to match translation investment to business value. Because the question isn’t MTPE or human translation. The question is which content gets which treatment, and whether you’re making that decision intentionally or by default.
