Microsoft says changes to Copilot have improved Japanese meeting transcripts, including raising recognition accuracy for Japanese names from about 40% to 66%. The work also cut transcription errors for English content spoken in Japanese meetings from about 90% to 27%, addressing a very everyday test of AI: can it follow a bilingual meeting without mangling the people in it?
Microsoft AI Watch analysis
What happened
In a Microsoft Inside Track article published on 5 November, digital experience lead Kevin Kotani describes work with Microsoft’s Copilot product group and Japanese language experts. Alongside speech-model updates, Microsoft says a fix stopped Copilot treating the Japanese honorific “san” as part of a person’s name. That fix was rolled out worldwide. Microsoft also says the system reduced mistaken language-switching in mixed-language settings by about 2.5%.
The work went beyond transcription. Microsoft describes tuning Copilot’s tone and interface text for more natural professional Japanese, and a shared process for routing localisation issues to engineering teams. The company says daily active use among its employees in Japan rose from an earlier baseline to a peak of 93%, and that 87% of surveyed Japanese employees said Copilot met or exceeded expectations in August 2026, up from about 76% in 2025.
Why it matters
For workplace AI, language support is not just a matter of translating prompts. Names, honorifics, meetings that switch languages and professional tone can determine whether a recap is useful or becomes another thing someone must correct. Microsoft’s figures put concrete measures against some of those problems.
The adoption and satisfaction numbers come from Microsoft’s own employees and survey, so they describe the company’s experience rather than an independent assessment of Copilot across Japanese workplaces. Still, the reported accuracy gains make this a more useful account than the usual promise that AI will simply “break language barriers”.
Our read
The strongest lesson is that localisation needs product fixes, native-speaker feedback and a clear route from reported problem to engineering owner. A better translation alone would not have fixed a meeting recap that turns an honorific into someone’s surname. Microsoft says Fujitsu expanded its Copilot deployment to more than 100,000 employees after improvements, but the account does not establish that localisation alone drove that decision.
What to watch
- Whether the reported gains hold up in independent tests and across more Japanese workplaces.
- Whether Microsoft publishes comparable results for other languages, including Korean.
- Whether the localisation process and proposed testing agents produce measurable improvements beyond Microsoft’s own users.
Discussion spark: For workplace AI in another language, should companies prioritise measurable accuracy gains in high-friction tasks such as meetings, or broader cultural fluency across the whole product?
Sources and evidence
- Inside Track – Beyond translation: How we’re making Microsoft Copilot work better in Japanese – Microsoft (5 November 2026, 16:45 UTC)
not affiliated with or endorsed by Microsoft