AI note-taking for meetings: stop writing up the minutes

Of everything I set up with clients, AI note taking for meetings is the one nobody needs persuading about. There is no philosophical debate to have. You sat through the meeting, you took half a page of notes you cannot read, and now you owe six people a summary by tonight. Handing that to a machine is not a leap of faith.

What surprises people is which part of it actually saves the time, because it is not the part they expect.

The transcript is not the point

Every tool in this category will record the call and give you a transcript. That part is solved and has been for a while. It is also close to worthless on its own. A sixty-minute meeting becomes nine thousand words, and nine thousand words is not a record of the meeting, it is the meeting again, in a format you will never open.

The value sits entirely in what happens after the transcript: a summary short enough to read on the walk to the lift, and a list of who agreed to do what by when. If a tool gives you the transcript and calls it done, it has handed you the raw material and kept the work.

A transcript is not a memory. It is a longer version of the thing you already forgot to read.

AI note taking for meetings: the setup worth keeping

The version that survives contact with a real week is duller than the demo:

• One tool, on one calendar. Pick the one that joins your actual meeting platform, whatever your firm already pays for. A second tool for a second platform means you will use neither.
• A house format you decide once. Decisions, owners, deadlines, open questions. Four headings, always in that order, so your eye knows where to go.
• A five-minute edit before it goes out. You correct the two things it misheard and delete the paragraph about the weather. This is the step people skip, and skipping it is how a summary quietly becomes wrong.
• One place the notes land. A folder, a channel, a document per client. Notes scattered across an inbox are not a record, they are litter.

That is it. The saving is not glamorous. It is thirty to forty minutes per meeting-heavy day, and the more useful half is that you stop carrying six unwritten summaries around in your head all week. If you want the summarise-and-extract prompts rather than another subscription, they are in the free prompt library.

The Hong Kong part nobody mentions

Three things reliably break this in a Hong Kong office, and none of them appear in the product marketing.

The first is consent. Recording a call is not a neutral act, and a good rule is to say out loud that the meeting is being transcribed before it starts. It costs you four seconds and it removes the entire problem.

The second is what your firm considers a record. If you work in banking, law, or anything with a regulator attached, assume a recording of a client conversation is a business record rather than a private note, and that it is discoverable. Before you point a consumer tool at a client call, find out where the audio is stored and who else can reach it. I have written separately about whether ChatGPT is safe for work data, and the same questions apply here with more force, because a transcript captures things people say that they would never write down.

The third is the one that catches everyone out. Half the meetings in this city switch between English and Cantonese, sometimes inside a single sentence, and most transcription models handle that badly. They will not tell you they handled it badly. You get a clean, confident, subtly wrong paragraph where the important qualification used to be. If your meetings code-switch, test the tool on a real one before you trust it in front of a client.

Where it still fails you

It cannot tell that the most important moment in the meeting was a pause. It flattens tone, so a reluctant "fine, we can look at that" and an enthusiastic one read identically on the page. It will attribute an action item to whoever spoke last rather than whoever agreed to it. And it captures what was said, not the personal detail you would want in front of you next time, which is a different job and one I have covered in using AI for client relationships.

None of that is an argument against using it. It is an argument for reading it before you send it.

The reason this is usually the first thing we fix

I spent thirteen years in meetings at a bank, and the notes were never the work. They were the tax on the work, paid in the hour when I was least able to think clearly. Automating that hour is worth doing on its own, but the real reason it goes first is that it is small enough to be running by Thursday.

Once one habit has visibly given time back, the conversation changes from whether this is worth it to which of the other four things should go next. That second conversation is the one worth having, and it is specific to how you actually work. If you want to have it, start here.