Research · The Edit Distance Report · 2026
Where your translation team's time matters most
We measured 27 million translated sentences. People change about 8% of AI suggestions and 15% of machine translation.
27 million sentences measured · 14 sections · Published by wxrks, formerly Bureau Works
27 million sentences · August 2024 to September 2026 · No customer text was read
Updated 14 October 2026
Hi. We measured 27 million translated sentences. Here is what we found.
Four numbers from the report
- 47%less editing than machine translation
- 84 of 100AI suggestions go out unchanged
- 42% to 16%editing, month one to month ten
- 98 of 100sentences pass review unchanged
How much do people change?
How much people change
Never has 8% mattered so much
Half of all sentences go out untouched. The editing that's left is where a person's judgement changes the result.
- 52%Untouched
- 13%Light, up to 5%
- 8%Moderate, 5 to 15%
- 13%Heavy, 15 to 50%
- 14%Rewritten, more than half
So where does the time go?
The person who runs localisation
Your role matters more than ever
Editing a suggestion is only 7% faster per word. The time comes back in the reading, because most of it is already right.
So the work moves to deciding what gets read, what goes to review, and what the memory keeps. Whoever runs that flow decides how much time AI saves.
You run the flow
- 1SuggestAI suggests from your memory and glossary. 84 of 100 go out as written.
- 2ReadYou choose what gets a full read. Smells flags where meaning may have drifted.
- 3ReviewYou choose which passes run. Keep them where the risk is.
- 4KeepYou decide what the memory and glossary keep. As they fill up, editing falls.
Editing falls as the memory fills up
It depends on the language
- Spanish (United States)9%
- Latin American Spanish13%
- Brazilian Portuguese14%
- Korean (South Korea)18%
- Spanish19%
- Japanese (Japan)20%
- Chinese (China)20%
- European Spanish21%
- Dutch (Netherlands)21%
- French (France)22%
- German (Germany)22%
- Canadian French24%
- Brazilian Portuguese → American English25%
- Italian (Italy)26%
How did you measure it?
How we measured
- Data
- 27 million sentences on the wxrks platform, August 2024 to September 2026.
- Measure
- Edit distance, the share of words a person changes before a sentence ships.
- Privacy
- Group numbers only. No client is named and no text was read.
- Author
- wxrks, formerly Bureau Works.
What the data does not show
- It counts changes, not quality. Quality is the next study.
- It is production data, not a lab test.
- The learning curve follows accounts that stayed.
Can I read the whole report?
Yes, it's free. No form, no email.
- 2024 against 2026
- Editing by origin
- The learning curve
- Editing by language
- Review and in-country review
- Speed
- Method and privacy
Questions
Is it free?
Yes. No form, no email.
What is edit distance?
The share of words a person changes in a suggestion before it ships. 0% means it was accepted as it came.
Can I share it?
Yes. Send the link or the PDF to anyone.