AI meeting minutes: what it drafts and what it cannot

AI can draft most of a set of minutes and cannot finish any of them. It is good at what was said, reliable at who said it, and useful for actions with an owner attached. It cannot record a quorum it was never told about, it cannot decide what to leave out, and it cannot sign anything off. Those are not gaps in a particular product. They are gaps between what a recording contains and what a minute is.

This page is about that boundary, requirement by requirement. If you are choosing what to buy, the categories are set out in meeting minutes software. If you want the method, there is a guide to taking meeting minutes and a separate one on board meeting minutes.

What does a set of minutes actually have to contain?

Minutes are a formal record rather than a write-up. The requirements are consistent across most bodies that keep them:

Read down the "no" rows and a pattern appears. Everything an AI cannot produce is something that was never in the audio: a rule, a previous document, or a future event.

Where AI genuinely earns its place

Actions. Pulling commitments out of a conversation and attaching a name is the single most reliable thing these tools do, and it is the part of minutes most often wrong in practice. People agree to things in passing and nobody writes it down.

Attribution. Knowing who said which part matters more in minutes than in a summary, because a minute sometimes has to record that a particular person dissented. Speaker labelling is imperfect and worth checking, but it beats reconstructing a two hour meeting from memory three days later.

The verbatim fallback. The value of a transcript is not that you read it. It is that when two people remember a decision differently, there is something to check.

Where it cannot help, and no product will change this

It does not know your constitution. Quorum, who may vote, whether a motion needs a seconder, what proportion carries it. These live in a governing document. A model reading a transcript has no access to them and no way to infer them.

It cannot decide what to leave out. This is the real work. Minutes fail in two directions: too thin and they do not evidence that the body applied its mind, too detailed and they create a disclosable record of a candid discussion. A model does not know which of those risks your organisation is carrying.

It cannot phrase a contested decision. "The board noted concerns" and "two directors opposed the proposal" can describe the same ninety seconds. Choosing between them is a judgement with consequences, and it belongs to the person whose name goes on the document.

It cannot approve anything. Minutes become the record when the body accepts them, usually at the following meeting. No amount of drafting quality shortens that.

So what is the honest workflow?

Record the meeting. Let the tool produce the transcript and a first draft. Then a person turns that draft into minutes: adds attendance and apologies, checks quorum against the constitution, phrases the decisions as decisions, carries forward what was deferred, and puts it up for approval.

That last step is not administrative tidying. It is where the document becomes a record.

What to check before you approve an AI-drafted minute

The draft will read well, which is the risk. Fluent prose is not the same as an accurate record, and a minute that reads smoothly is harder to challenge than one that reads badly. Six things are worth checking every time.

Attendance and apologies. Add anyone who did not speak. Check the apologies against the invitations rather than the audio.

Quorum. Against the constitution, not against the number of voices in the transcript.

Every decision, phrased as a decision. A model tends to report discussion. Look for anything written as "the group talked about" that should read "it was agreed that".

Names attached to actions, and dates. The action is usually captured. The owner is sometimes guessed from who spoke last, which is not the same as who accepted it.

Anything a speaker asked to be kept out. People say "off the record" mid-sentence and a transcript does not honour it. This is the most common way an AI-drafted minute causes a problem.

Numbers and figures. These are where transcription errors hurt most, because a wrong figure in an approved minute is a wrong figure in the record. Check them against whatever was on screen.

If that list feels long, it is still shorter than reconstructing the meeting from memory, and it is the difference between a draft and a record.

Where Weeve sits in this, plainly

Weeve is ours, and it is not minutes software. It records the meeting on a Mac, transcribes and summarises on the device, and gives you an accurate account to draft from. It produces a summary, not a minute, and the difference is everything on this page.

It is also Mac only, on Apple Silicon with macOS 14 or later, so it is no help on Windows. Accuracy drops on heavy accents, overlapping speakers and noisy rooms, which are exactly the conditions of a busy committee. And it has no shared workspace, so a company secretary circulating a draft is doing that in whatever they already use.

What it does solve is the part where you are expected to take part in the meeting and produce the record of it. The recording and the transcript stay on the machine that made them, which matters when the meeting was about a redundancy programme or a legal dispute.

Common questions

Can AI write meeting minutes?

It can write a good first draft of what was said, what was decided and who owns the actions. It cannot supply the things that were never in the room: quorum, the previous meeting's carried-forward items, or the approval that turns a draft into the record. Treat the output as a draft a person finishes.

Are AI generated meeting minutes legally valid?

Validity comes from the body approving them, not from how they were drafted. A set of minutes written by hand, typed from a transcript or drafted by a model all become the record the same way, when the meeting accepts them. If your governing document says something specific about how minutes are produced, that document decides.

What is the difference between AI meeting minutes and an AI summary?

A summary tells you what happened. A minute is a formal record with required elements and an approval step. Most tools sold for meetings produce summaries. Calling that output minutes is where the disappointment in this category comes from.

Will AI record who was absent?

No. It hears the people who speak. Someone who attended and said nothing, and someone who sent apologies, both have to be added by a person. On a board this is not a detail: attendance is often the first thing a regulator checks.

Can AI handle a vote?

Only if the vote is spoken. "Four in favour, one against, one abstention" said out loud will be captured. A show of hands will not be, because nothing about it reaches the microphone.

Does the recording have to go to a server?

That depends on the tool, and it is worth checking before a confidential meeting rather than after. Some process in a vendor cloud, which is normal and fine for most organisations. Weeve runs on the Mac itself, so the audio and the transcript stay on the device.

A model can tell you what was said. Deciding what the meeting will be remembered for is still a person's job.

If the meetings are the kind where the recording should not leave the building, Weeve's free Starter plan records and transcribes on your Mac, ten recordings a month. The board minutes template is free to copy either way.