The transcript first, because that is the input
The summary will never be better than what the machine heard. And here the gap between the marketing figure and reality is wide.
The "98% accuracy" claim applies to a clean recording: one speaker, a close microphone, silence. A real meeting room is a different matter. An independent challenge run on 315 real meetings measured this:
| How it was recorded | Baseline | Best system |
|---|---|---|
| One device on the table | 41.4% word error rate | 22.2% |
| Microphone array in the room | 28.3% | 10.8% |
So with a single laptop in the middle of the table, even the best research-lab system gets every fourth or fifth word wrong. With a microphone array the same task drops to half of that.
The lesson for the recipe: the room and the microphone matter more than which tool you pick.
Outside English the starting point is worse. The most recent system tuned specifically for Hungarian reached a word error rate of 14.18% on spontaneous speech — it gets every seventh word wrong. And that is not a noisy meeting room but a lab-quality recording: in a real meeting in a language like that, with people talking over each other, it can only be worse.
There is one separate, rarely mentioned type of error: in a study of widely used transcription models, roughly 1% of the transcripts contained an entirely invented sentence — not a mishearing, but text that was put there. More than a third of that was content capable of causing harm: invented personal data, violent phrasing or a false reference. It is most likely during a pause in the speech.
The summary: what it gets wrong, and what it does not
The most detailed recent measurement ran on 114 real meetings of three different kinds (a municipal council session, confidential business material, a press conference), with 680 grading runs. Three things were measured:
| What it measures | What came out |
|---|---|
| Precision — whether the claims in the summary are supported | 0.553–0.583 |
| Coverage — how many of the reference items made it in | 0.804–0.942 |
| Detail — how fully the things that made it in were spelled out | 0.814–0.886 |
Two things are worth reading out of that.
The first: precision and completeness move in opposite directions. The model with the best coverage was the least precise. Whoever writes down more hits more — and invents more. There is no setting where both are good at once.
The second, and more important: it does not lie, it adds. The structure of the errors gives it away. There were 5,619 unsupported claims and 3,224 invented facts — but only 114 claims that openly contradicted the recording. The summary almost never contradicts what was said. It states something that was not said: a specific number, a name, an owner, a deadline.
This is the most dangerous kind of error, because when you read it, it looks exactly like every other line.
A contradiction worth spelling out. Another research group, working on real meeting data, found that "factual errors are rare in real life". The two results do not rule each other out: "not supported" is not the same as "false". And both measurements were carried out by machine graders whose own reliability is questionable — the same research measured a correlation of 0.11 between the machine and the human score on that real meeting data. The order of magnitude of these numbers is usable; the second decimal place is not.
What there is no separate measurement for: I found no independent, standalone figure for the accuracy of extracting decisions, owners and deadlines. The coverage values give an approximation, but anyone telling you "it gets 97% of the action items right" is not citing a measurement.
Language: what works and what is not promised
This difference surprises a lot of people, and it is worth knowing about before you build a process on top of it.
The quality of machine speech recognition varies enormously by language, and the reason for the gap is no mystery: the amount of training data. In the training set of one widely cited model family there are 11,100 hours of Spanish audio and 379 hours of Hungarian — a difference of roughly twenty-nine times. Same model, same benchmark, word error rate in percent:
| Benchmark | English | Spanish | Hungarian |
|---|---|---|---|
| Read speech | 4.2 | 3.0 | 17.0 |
| Crowd-sourced recordings | 9.4 | 5.6 | 19.7 |
So Spanish is among the well-supplied languages — on the read-speech benchmark it even beats English — while Hungarian sits in the data-poor band. That is not vendor malice, it is arithmetic. Tool support roughly follows the same shape; open the one that applies to you:
Hungary — transcription works, automatic note-taking is mixed
Live transcription works in Hungarian in the largest office suite, and Hungarian has made it onto the supported-language list for intelligent recap — it is on the vendor's official list. That needs a premium or assistant licence, and a meeting of at least five minutes.
The other big suite's automatic note-taker, however, supports eight languages, and Hungarian is not one of them. The same feature handles one language at a time, so a mixed-language meeting is not covered there at all.
What to take from that: an automatic summary in Hungarian is now officially supported in one suite and still does not exist in the other. If a weekly report depends on it, that decides which tool you work with — and because of the error rates above, the transcript is worth skimming either way.
Spain — both big tools support it, but not mixed language
In Spanish, both large office suites support automatic summarisation: in one as part of the intelligent recap languages, in the other as one of the eight supported languages.
The limit lies elsewhere. The note-taker handles one language at a time: if the meeting starts in Spanish and switches to English — routine in an international team — the feature does not officially support it.
And the measurements are better, but those are laboratory figures. In a real meeting room, with a single tabletop microphone, even the best system errs on more than 20% of words, regardless of language. A language advantage does not cancel out the acoustics of the room.
The recipe
Five steps. The third one is the point.
1. THE RECORDING
A microphone array, or at least one close microphone.
This helps more than any choice of tool.
2. SAY IT AT THE START OF THE MEETING
We record, this is the note-taker, the audio goes here.
If anyone objects, this is the moment to say so.
3. ASK FOR THREE LISTS, NOT A SUMMARY
Decision / owner / deadline. In separate lists.
Anything you cannot file is an open question.
4. ASK FOR IT WITH THE QUOTE
Next to every decision, the sentence that was said.
Anything with no quote is invention — you delete it.
5. ONE PERSON REVIEWS IT BEFORE IT GOES OUT
Five minutes. They check the owners and the dates,
not the wording.
And a copy-paste prompt for the third step:
This is the transcript of a meeting. Write three lists
from it, then a fourth at the end — nothing else:
DECISIONS — what was decided. Next to each decision put
the exact sentence from the transcript that this
follows from.
OWNERS AND ACTIONS — who took on what. If the name was
not said out loud, write: "not named".
DEADLINES — what by when. If no date was said, write:
"no date was given".
Leave out anything you cannot back with a quote. Do not
fill gaps and do not interpret. At the end, write an
OPEN QUESTIONS list of what was raised
but not closed.
Asking for the quote is the only cheap defence against the error type above: if the summary adds a detail that is not in the recording, it cannot attach a quote to it. And spelling out "not named" and "no date was given" prevents the model from filling the gap with a plausible guess.
What to say at the start of the meeting
The legal treatment of audio recording differs from member state to member state, and the centre of gravity is not even in the same place: in some countries the recording itself is the question, in others the recording is fine and passing it on is not. Open the one that applies to you:
Hungary — recording requires consent
In Hungary, making and using an audio recording requires the consent of those involved — that comes from the personality-rights rule of the Civil Code.
On top of that, the recording is also data processing: in December 2024 the Hungarian data protection authority issued a separate statement warning about the spreading practice of covert audio recordings. The two rules run in parallel — the fact that a recording is defensible under civil law does not stop the data processing from being unlawful.
Spain — the recording itself is fine, distribution is the risk
The Spanish starting point is the reverse. Under a 1984 constitutional court ruling, reaffirmed since, a participant in a conversation may record it: there is "no secret" as against the person the communication is addressed to. In 2023 the supreme court added that it is enough for one participant to consent to recording by a third party.
Criminal law, on the other hand, catches distribution: Article 197(7) of the criminal code punishes anyone who passes on, without authorisation, a recording obtained with the subject's agreement, where doing so seriously harms their privacy. Paragraph 3 sets a heavier penalty for distributing unlawfully obtained material.
The workplace has a rule of its own: under Article 89(3) of the Spanish data protection act, recording sound at the workplace is permitted only where risks to security arising from the activity make it relevant, and then subject to proportionality and minimum intervention. A plain set of meeting notes does not meet that condition — it needs a different legal basis.
And there are actual fines: in a 2024 case the authority refused to treat as valid consent the platform's automatic message ("by joining, you consent to the recording"), and fined the company over the missing information. In a 2026 case it fined another company for forwarding a consented recording to outsiders in place of minutes.
In 2026 the authority published a two-part guidance specifically on AI voice transcription. Two sentences from it are usable anywhere: participants have to receive the information before the session starts, and consent given for one session expires when that session ends — it does not carry over to the next one.
And there is a third question that most places never ask: where does the audio go? In a US class action filed in 2025, the court in August 2026 partly upheld the claim that the note-taking service was not merely the meeting host's recording device but also used the conversations for its own business purposes, including model training. The case is not closed, but the lesson is already usable: it is not enough that the participants consented to the recording. They also need to know who gets it and what it is used for.
One sentence at the start of the meeting settles it:
"We are recording this and having it transcribed by machine. The recording goes into this tool and we keep it for this long. If that does not work for anyone, say so now."
Common misconceptions
- "Accuracy is 98%, the vendor says so." True for a clean recording. In a real meeting room, with one microphone on the table, even the best measured system errs around 22%, and outside the big languages it is already 14% on a lab recording.
- "If it gets something wrong, I will notice, because it will be nonsense." The overwhelming majority of the measured errors are not contradictions but additions: plausible, specific, unverifiable detail. That is exactly what does not stand out.
- "They will read it if it is nice and long." The other way round. Whatever is not a decision, an owner or a deadline does not get read — at best it gets searched.
- "It works the same in my language." Live transcription does. Automatic summarisation is officially supported for a far longer list of languages in one of the two biggest office suites than in the other — check which one you are on.
- "It is enough that I tell them I am recording." Beyond consent, it also matters where the audio goes and what it is used for — which is precisely what the pending lawsuits are probing.
What to take away
With meeting notes, the machine takes the typing off your hands, not the writing. And not the judgement: what counts as a decision is still said by a person.
That is why three lists with quotes work: you hand the machine what it is good at — collecting and ordering — and you structurally rule out what it is measurably bad at, which is adding things. The five minutes of human review are not a luxury; they are what holds the system together.
Where you can gain any time at all in an office day — and why time spent in meetings does not fall because of AI — is set out with measurements in the real time savings post; and from the inbox side, the inbox triage post gives you a recipe.