Zoom, Google Meet or Microsoft Teams · Practical walkthrough

Choose a meeting notetaker that fits your consent and retention rules

Someone wants an AI notetaker on every project meeting from next month. A feature list won't tell you who receives the notes, what happens if a guest objects, or when each copy disappears.

Task by Task14 min read + practice

What you'll make

a short evidence grid comparing notetakers on notice, objections, sharing and retention, with what the vendor documents, what your account is set to and what’s still unknown kept apart.

What you'll need

the candidate products, their current help pages, your organisation’s requirements, and an administrator or privacy lead who can confirm settings. About 45 minutes. Don’t connect a calendar or record a real meeting for this desk review.

Example result.

Row: what happens if someone declines (help pages read 10 October 2026)
Zoom: with the custom AI notice on, a participant who doesn't accept stays in but can't unmute or turn on their camera
Meet: internal participants can stop "Take notes for me" unless host controls are on; an external guest has no in-product switch
Teams: AI notes need at least transcription, so declining means asking the organiser to stop transcription
Our setting: custom notice on/off not yet confirmed by IT (owner: Sam, IT)

A notetaker creates several records: audio, a transcript, generated notes, a shared summary. Each can have its own recipients and retention. “Secure” beside both products leaves most of the useful questions unanswered.

1. Write the requirements before comparing tools

Pick one intended use, such as internal project reviews; client meetings, interviews and sensitive discussions may need their own assessment. Name who decides whether the use is acceptable. Then list the questions:

What is captured: audio, video, screen content, transcript or notes?
What notice do participants get, and what happens if someone objects?
Who can start or stop each feature? Can a calendar rule start it automatically?
Where is each output stored, and who owns it?
Who receives access by default, including invitees who don’t attend?
What controls retention, deletion, backups and exported copies?
What may the provider do with the data, including model training?
Which plan, region and admin settings apply?

Keep must-haves visible. If external sharing must be off and nobody can confirm that setting, the tool isn’t ready for that use, whatever else it scores.

For UK organisations, the ICO’s guidance on AI meeting tools connects them with your purposes, lawful basis, privacy information and accurate outputs. A participant notice or an “accept” button doesn’t make the whole workflow lawful; ask your privacy or legal lead to apply the requirements.

Why this works: the questions decide the comparison. Without them, the grid fills up with whichever features the vendor’s page leads with.

What you should see: one intended use, a decision owner and the eight questions, with must-haves marked.

2. Fill each row with documented and configured evidence

For each question, record what the vendor documents, what your account actually does and what’s still unknown:

Question:
Product, plan and account type:
Documented behaviour, with page and section:
Date read:
Our configured setting, and who confirmed it:
Smallest outstanding question and owner:

If the documentation is silent, write “not found in the pages checked” and list them. Don’t write “unsupported” unless the vendor says the feature is unavailable. Check what words like “data” cover: a transcript retention setting may not touch exported notes, a shared file or an email copy.

Why this works: public help pages show what a product can do. Only an admin can tell you which of those things your account does.

What you should see: every row with a dated source and a named person for the account setting, or a blank marked as a blocker.

3. Ask AI to find holes in your evidence

Paste the grid and short documentation extracts into a work AI chat (ChatGPT, Claude, Gemini or Copilot). Leave out confidential configuration details unless that account is allowed to hold them.

Review this meeting-notetaker evidence grid for [intended use], using only the grid and extracts I've pasted.
For each row, find: an unsupported or vague answer; confusion between a documented capability and our actual setting; an unclear plan, region, recipient or output; and missing questions about copies, deletion or automatic joining.
Return the issue, where it is, why it matters and one factual question for our administrator or the vendor. Don't fill gaps from memory, pick a winner or declare anything legally compliant.

[requirements, grid and extracts]

[intended use] is the use from step 1, such as “internal project reviews”.

Why this works: the capability-versus-configuration check is the one people skip, and the one that matters most.

What you should see: a list of questions you can assign, not another polished comparison.

Check the result

Every documented claim links to the page and section it came from, with a date read.
Every “our setting” cell names who confirmed it, or is marked unknown.
No row turns a control into a promise (“has a retention control” isn’t “deleted after 30 days”).
Must-have rows with blanks are marked as blockers, not averaged away.

The usual failure: a documented capability quietly becomes a statement about your account. “Zoom can delete transcripts after a set number of days” becomes “our transcripts are deleted after 30 days” when nobody has checked the setting. Spot it by looking for any “our setting” cell without a named person.

Your grid states [claim], but the source only establishes [what the documentation says]. Remove the conclusion about our account and record the configuration question that would resolve it. Don't infer a default from the existence of a control.

Worked example: three rows filled from documentation

Fictional example. An operations manager compares Zoom, Google Meet and Teams for internal project reviews. Her first draft of the retention row read: “Zoom: transcripts deleted after 30 days. Meet: deleted with the meeting. Teams: follows recording retention.” None of those was her account’s setting, and two weren’t what the help pages say. After the repair prompt and a reread of the pages (on 10 October 2026), three rows read:

Participant notice.

Zoom: admins can turn on Show a custom message when turning on AI in a meeting (Admin Center › Settings › Zoom AI › Meeting). Non-host participants then join with camera and microphone off until they accept a notice shown for 90 seconds, or leave. Zoom AI notices - Phone users can require dial-in users to press 1 to consent.
Meet: a pencil icon appears on everyone’s screen while “Take notes for me” runs. An admin can require explicit consent first; that setting is off by default.
Teams: AI notes need the meeting to be at least transcribed, so the notice people see is the transcription notice.

If someone declines.

Zoom: a participant who doesn’t accept stays in but can’t unmute or use their camera. Silence isn’t an opt-out; it’s a mute.
Meet: internal participants can stop notes, unless host controls are on, when only hosts and co-hosts can. An external guest who objects has no in-product switch.
Teams: ask the organiser to stop transcription.

Retention and storage.

Zoom: Delete transcripts after a specified number of days, under Retention and permissions in the meeting transcript settings.
Meet: the notes document is saved in the organiser’s Google Drive, in a “Google Meet” folder, and follows the organisation’s Meet retention policy.
Teams: AI notes and tasks expire according to the organisation’s policies; ask IT for the detail.
The draft confused capability with configuration. Zoom documents a dial; it doesn’t say where yours is set.
One product has a dial; two inherit from policies you already run. That isn’t a ranking, but it changes who needs to be in the room.
The declining row is the one that differs most, and it’s the one a buyer is least likely to find on a feature list.

For setting up the Zoom notice itself, see set up Zoom AI meeting summaries and tell attendees.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Notetakers that join as a guest

Tools that dial into a meeting as an extra participant take the recording off the platform your organisation already governs. Add three rows: does it join as a visible participant with its own name, and can that name be changed to something less obvious? Does it join automatically from a calendar connection, including meetings the user didn’t think about when connecting it? Is the connecting account an individual’s or the organisation’s, and what happens to the archive when that person leaves? For meetings that shouldn’t be captured at all, see review which meetings should use AI notes.

Follow every copy of the output

List each copy: audio, transcript, generated notes, shared document, email or chat notification, exports and backups. For each, ask who can open it, what governs its retention and whether deleting the original removes it. Don’t promise participants that one deletion erases everything.

Review this output inventory against our evidence grid. Identify copies whose owner, access or retention is unknown. Return the missing facts only; don't assume one deletion control covers every copy.

Test the objection route, and date the grid

Once a tool and a fictional test meeting are agreed, ask the admin to demonstrate the participant view for a late joiner, an external guest and a dial-in user; a notice on one desktop client doesn’t show what everyone sees. Write down the fallback, such as manual notes. Put the date you read each page beside its link. A grid without read-dates becomes indistinguishable from folklore in about four months.