Machine translation quality crossed a meaningful threshold somewhere between 2021 and 2023. The combination of neural MT improvements and large language models produced output that, for a wide range of content types, is usable after light post-editing rather than full retranslation. That change is real and it has affected localization economics at the enterprise level.
But the workflow bottleneck in enterprise localization was never primarily translation quality. It was handoff latency. And that bottleneck, for the most part, remains.
What Changed
The cost per word for machine-translated post-edited content dropped significantly. For content categories where MT quality is high -- technical documentation, UI strings with controlled vocabulary, product descriptions following a template, support knowledge base articles -- the post-edit rate (percentage of MT output that needs human revision) is now low enough that MT-first is the default for many localization operations, not an experiment.
Translation memory hit rates also improved. Modern MT systems, when properly integrated with TM, can apply TM matches for near-duplicate strings at a higher threshold than traditional fuzzy matching alone, effectively extending TM coverage.
For content at volume -- annual reports, large documentation libraries, marketing localization for established brands -- these improvements translated to real cost reduction. An enterprise that was spending $0.25/word for human translation can now achieve acceptable quality for many content types at $0.04-0.08/word for MT post-edit.
What Did Not Change
The workflow that delivers strings to translators (or to the MT system) and returns them to the product has not fundamentally changed in most enterprise localization operations. The path is still:
- Source content changes in the product or documentation system.
- A localization request is filed (manually, in most organizations).
- Strings are extracted and uploaded to the TMS.
- TM and MT are applied, human review happens for flagged segments.
- Translated content is downloaded and integrated back into the product.
Steps 1 and 5 are still largely manual in many organizations. Step 2 is still a human-initiated step in most. The latency between "source content changes" and "translated content is integrated back into the product" is measured in days or weeks, not hours -- and MT has not changed that, because MT operates inside the TMS, not at the handoff layer.
Where AI Actually Moved the Needle
Beyond translation quality, AI has meaningfully changed three specific localization economics:
Cultural Adaptation at Scale
Large language models are demonstrably better at cultural adaptation than earlier MT systems. For marketing copy, UI strings that require idiomatic localization rather than literal translation, and content where brand voice needs to carry across languages, LLM-based translation with appropriate prompting produces output that requires less substantive post-editing than phrase-based or earlier neural MT. For enterprise brands with strong voice requirements, this reduces the time-intensive review cycles that drove localization costs higher than per-word rates suggested.
Quality Estimation
AI-based quality estimation (QE) models have matured to the point where they can reliably predict which MT segments need human review and which do not. Rather than routing all MT output through human post-edit, quality estimation routes only the high-uncertainty segments. For a large documentation corpus, this can reduce the volume requiring human attention by 40-60% without a proportional quality hit.
Glossary and Terminology Consistency
Maintaining brand terminology consistency across a large localization operation -- ensuring that "workspace" is always translated as the approved equivalent in each locale, not as a generic synonym -- was previously enforced through terminology databases and TM constraints with imperfect coverage. LLM-based translation, when given the brand glossary as context, maintains terminology consistency at a higher rate than constrained MT approaches without the glossary needing to be exhaustive.
The Bottleneck That Remains
Fast, accurate translation -- whether human or MT -- sitting behind a slow handoff workflow produces the same user-facing result as slow translation: strings that do not reach the product until after the source version has already shipped.
For software localization specifically, the economics that matter most to product teams are not per-word translation costs -- those are TMS line items visible to localization operations. The economics that matter to product teams are time-to-ship and rework cost. MT has compressed the translation step within the TMS. It has not compressed the handoff latency before and after the TMS step.
A product team that was experiencing a two-sprint localization lag before AI improvements to MT quality is still experiencing a two-sprint localization lag if the extract-upload-download cycle is still manual and the localization request still depends on a human noticing that strings have changed. MT running faster inside the TMS does not help when the TMS does not know the strings have changed yet.
The Next Leverage Point
The leverage point that MT did not address is the detection-and-delivery pipeline: the automated detection of string changes in the codebase, the immediate queuing to the translation layer, and the automatic merging of completed translations back into the release branch. This is not a translation quality problem -- it is a workflow architecture problem.
When string change detection is continuous (on every commit), translation queuing is automatic (no human-initiated TMS upload step), and locale file merge is triggered by translation approval rather than a scheduled export job, the latency between source change and translated release compresses to the translation step itself. At that point, the improvements to MT speed and quality that AI delivered become fully visible at the product level, not just the TMS budget line.
The pairing of AI translation quality improvements with a sync-first workflow architecture is where the full economics change. Either one alone is incremental. Together, they close the sprint gap.