How AI Can Assist with Physical Meeting Minutes

On a video call, each person's voice arrives on its own channel from a microphone about forty centimetres from their mouth. The platform knows who is speaking because it knows which stream carried the audio.

In a room, one microphone hears everyone at different distances, plus the air conditioning, the corridor, the chairs and the two people who start talking at once. Speaker separation gets harder, quiet participants get lost, and the platform has no account identity to attach a name to.

Nothing about that is unfixable, but it changes where the effort goes. On a call, the work is configuration. In a room, the work is placement, announcement and review — three things a person does, not a setting anyone can toggle.

Sources: ^[1,4]

Read the vendors' scope carefully before assuming your existing licence covers the boardroom.

Google does support this case explicitly. Its help page for "Take notes for me" for in person meetings says you can capture notes for an existing meeting space from a scheduled meeting, or for impromptu and casual meetings that are in person, and that you do not need to be in a meeting room to use it. The flow is to open meet.google.com, tap Take notes, then Start taking notes, and set your device nearby so Gemini documents the meeting; Pause and Stop are available while it runs. Notes arrive by email and in the Drive folder of whoever turned the feature on. The limits carry over from the online version: an eligible Google Workspace edition or Google AI plan, the same supported languages, one language at a time, and a summary that Google says can be incomplete, inaccurate or not generated. Google also notes that if someone who is not in the room wants to join, you transition to a normal video call.

Microsoft's transcription and captions documentation is scoped to Teams meetings and events. Transcription is a per-organiser and per-user policy, and transcripts are stored alongside the meeting recording in OneDrive and SharePoint — all of which presupposes a Teams meeting exists. Microsoft separately states that Copilot will not work in meetings hosted outside the participant's organisation, which rules out the common in-room case of a client or partner session on someone else's tenant.

The honest summary: Google has a documented in-person path with plan and language conditions; the Teams path assumes a Teams meeting. Anything else in the room needs a device that records locally.

Sources: ^[2,3,4,5]

Placement is the highest-leverage decision in the whole workflow, and it costs nothing.

Memora's public staff guide gives three instructions for a clearer transcript: place the device or microphone near the centre of the table, keep the browser tab open and the screen awake while recording, and choose EN for English meetings or BM for Bahasa Melayu meetings. Google's in-person page gives the equivalent instruction in one line — set your device nearby.

Assign the device to a person, not to the room. Whoever owns capture that day does four things: puts the device down before people sit, starts capture, watches that it is still running after any break, and ends it deliberately. Memora's guide notes that pause and resume are available whenever needed, that you tap End Meeting when the meeting is over and keep the page open until the upload finishes, and that a silence warning asks whether to keep recording or end the meeting — it never stops the meeting automatically. Google's in-person flow similarly offers Pause and Stop, with a confirmation before stopping.

One practical rule that prevents most lost recordings: upload completion is not the end of the meeting for the person holding the device. Wait for it.

Sources: ^[1,2]

On a call, the platform can tell everyone. Google's Meet documentation describes exactly that: Google Meet informs all meeting participants that notes are being taken, a pencil icon appears on participants' screens, and an administrator may require all participants to give explicit consent before features like "Take notes for me", Record or Transcribe start — a setting Google says is off by default and can be turned on.

That mechanism lives in the meeting interface. The people sitting around a table are not in that interface, so in a room the announcement is a human step. Say it out loud before you start, say what is being captured, and say who will receive the output. If someone objects, that is information you needed before the meeting rather than after it.

Where your organisation runs a formal policy on recording, apply it here the same way you would on a call. This is an operational courtesy and a governance habit, not a legal opinion — check your own organisation's policy for the rules that bind you.

Sources: ^[2,3]

Room audio produces a different failure profile from call audio, so review it for different things. Overlapping speech tends to merge, quiet speakers at the far end of the table get thinner coverage, and speaker labels have no account identity behind them — so names are the field to check first, not last.

Work through six items against the recording rather than from memory: names, figures, dates, decisions, action owners and deadlines. Memora's guide puts this review after the upload finishes: open Meetings and check the audio, transcript, summary, action items and decisions.

Google's own caveat applies in the room as much as on a call — a summary can be incomplete, inaccurate or not generated, and Google adds that the recommended meeting length for a usable summary is at least fifteen minutes.

The review step is not administrative overhead. NIST's AI Risk Management Framework is described by NIST as a voluntary framework intended to improve the ability to incorporate trustworthiness considerations into the design, development, use and evaluation of AI products, services and systems. Evaluation of the output is part of using the system, not a separate favour to the compliance team.

We do not publish an accuracy figure for Memora in physical rooms, and no first-party measurement exists to quote. Run your own room, with your own accents and your own table, and judge the transcript yourself.

Sources: ^[1,2,3,6]

Six steps, the same every time, and it takes about ninety seconds of extra effort per meeting.

1. Before. Name the person who owns capture. Confirm the device is charged and the language setting matches the meeting.
2. On arrival. Put the device near the centre of the table, before people sit down around it.
3. Announce. Say that capture is running, what is being captured, and who will receive the output.
4. Start, then check. Start capture, and glance at it once after the first break to confirm it is still running.
5. End deliberately. End the meeting in the app and wait for the upload to complete before closing anything.
6. Review and approve. One named person checks names, figures, dates, decisions, owners and deadlines against the recording, then approves and distributes.

For the template that the reviewed output goes into, see AI Meeting Minutes Template: A Practical Format for Clear Notes. For the wider picture of what can be automated across platforms, see How AI Can Automate Your Meeting Minutes.

Sources: ^[1,2,6]