How Accurate Are Automatic AI Meeting Notes in Indonesian, and What Are the Risks

The meeting ends, and then someone gets stuck writing the minutes. It's no surprise that automatic meeting-notes tools sound like the answer. But for Indonesian, especially Indonesian mixed with English mid-sentence, there are three things to check first: accuracy, tool support, and privacy.

How accurate is it for Indonesian?

The usual yardstick is WER (word error rate): the percentage of words transcribed wrongly. One public number you can hold on to comes from the Whisper paper, OpenAI's speech recognition model. In the Whisper paper (Radford et al., 2022), the large-v2 model scored a 7.1% WER for Indonesian on the FLEURS dataset, versus 4.2% for English and 3.0% for Spanish. That's roughly seven wrong words in every hundred.

Wrong words per 100 on FLEURS (Whisper large-v2), lower is better(WER %)

• Spanish3.0%
• English4.2%
• Indonesian7.1%

A few important notes before using that number to picture your own meetings:

• FLEURS is read speech, not a meeting. The audio is clean and one person speaks at a time. Real meetings have overlapping voices, noise, internal jargon, names of people and products, and mixed languages. The result is likely to be worse than 7.1%.
• The data is from 2022. Newer models are likely more accurate, but I couldn't find a comparable public figure for Indonesian meetings mixed with English, so test it with your own meeting recordings.
• What matters isn't all the words, but certain words. A wrong connecting word is no big deal. A wrong name, number, date, or term is what makes meeting notes misleading.

Which tools support Indonesian

This is what trips people up most: meeting summary features have their own language lists, and Indonesian isn't always on them. According to each vendor's official documentation at the time of writing:

Two notes. For Teams, I only checked the transcription feature. Copilot's summaries have their own licensing requirements and language list, so check before you buy. And every vendor's language list moves fast, so treat this table as a starting point, not a promise.

If the built-in tool doesn't support your language, there's another route: record the meeting, transcribe it with a speech recognition model through an API, then summarize it with a language model. That gives you more control, with more setup.

Risks that often get missed

1. Privacy and participant consent

Between August and September 2025, several class action lawsuits were filed against Otter.ai in federal court in Northern California and consolidated as In re Otter.AI Privacy Litigation. According to a NatLawReview summary, the complaints allege that its notetaker joins video meetings and records participants without notice or consent, and uses the data to train models. Those are still allegations, and that article says no substantive rulings had been issued when it was written, so the status may have changed since. But the lesson is clear: don't let anyone be recorded without knowing.

In Indonesia, meeting recordings and transcripts can contain personal data: names, salaries, performance reviews, even personal financial data, which the PDP Law (Indonesia's Personal Data Protection Law) classifies as specific personal data. This isn't legal advice, but the basic habit is simple: tell every participant before the meeting starts.

2. A summary summarizes, it doesn't witness

AI can misattribute who said what, or turn "let's consider it" into "decided". Treat the summary as a draft, and for decisions, tasks, numbers, and dates in particular, check against the source before sending it out.

3. Meetings that shouldn't be recorded

Meetings about employee evaluations, legal matters, sensitive financial data, or material under NDA shouldn't go into an automatic notes tool unless you truly understand where the data goes. The risk is the same as the one I covered in Want to Summarize Your Company's Financial Data in ChatGPT? Is It Dangerous?.

4. AI notes aren't official minutes

If a meeting produces decisions with legal or contractual consequences, the official minutes should still be confirmed by the participants. An AI summary helps draft them, but it's not a substitute for sign-off.

5. Where the recording and transcript go

In Teams, for example, transcripts are stored with the recording in OneDrive and SharePoint. That means anyone with access to that folder can read what was said in the meeting. Decide first who should have access, and how long the file needs to be kept.

How to use it so the result can be trusted

1. Announce it in the invite and at the start of the meeting that the meeting is recorded and summarized by AI.
2. Choose which meetings qualify. Project coordination, operations, and brainstorming are a good fit. Sensitive ones, no.
3. Prepare a glossary. Put names, products, and key abbreviations in the agenda, and say them clearly. One person speaking at a time helps a lot.
4. Review the summary within 24 hours. Check decisions, tasks (who, what, when), numbers, and dates. Fix them before sharing.
5. Move tasks to where they're actually used, don't leave them buried in a notes document. Reminders can be automatic, like in the team task reminder system.
6. Set a retention period. Delete the recording once the minutes are approved, unless it genuinely needs to be kept.

A realistic outcome: less time writing notes, replaced by time checking them. That's a good trade, as long as the checking actually happens. How to split what can run alone from what a human must check is covered in At What Point Should Humans Still Check AI and Automation Results.

Wrapping up

Automatic meeting notes are useful, but their accuracy in Indonesian isn't something to assume, and tool support varies. Test with your own meeting recordings first, tell the participants, and treat the output as a draft.

If you're weighing a tool or a meeting-notes workflow for your team, tell me what you need first, and I'll help think through what makes sense for your language and how sensitive your meetings are.