I don't want AI meeting notes to tell me everyone agreed. I want them to preserve the bit where someone didn't.
That's a much harder job than producing a tidy summary. These tools turn a spoken meeting into written notes, often with suggested tasks attached. Useful. But a neat account of a conversation isn't permission to spend money, promise a customer something or start a project.
My objection isn't to the note-taking. It's to quietly promoting the note-taker into management.
Even a perfect transcript can't settle whether “interesting” means “promising” or “please stop talking so we can have lunch”. British business has survived for years on that distinction. I'd rather not let a summary tidy it away.
The awkward bit belongs in the notes
Google Meet's note-taking feature can put a Decisions section into its output, with outcomes labelled Aligned, Needs Further Discussions, Disagreed or Shelved. That's a useful distinction. A summary that preserves disagreement is more valuable to me than one that merely sounds decisive.
But those labels are still generated by software. Google also warns that meeting summaries can be incomplete or inaccurate. Microsoft gives a similar warning for Teams' AI-generated recap content, which can include notes and follow-up tasks.
The warning matters because the dangerous mistake isn't necessarily an invented sentence. It's a real sentence with its condition missing.
Imagine a supplier discussion. Someone says we could offer a customer a longer payment period if finance approves it. Another person offers to ask finance. Nobody approves the actual offer.
If the resulting task becomes “send the customer revised payment terms”, the software hasn't just shortened a conversation. It's crossed the line between exploring an option and authorising it.
That's a hypothetical, not a claim about a particular tool's output. It's also the test I'd put ahead of how elegantly the thing formats its headings.
Congratulations: you've invented a project nobody authorised, with excellent minutes.
The useful question isn't whether the summary contains the words everyone said. It's whether it preserves what they're still waiting for.
Who gets the tidy version?
Google lets hosts choose who receives notes, including all invited guests, internal invitees, or hosts and co-hosts only. Its definition of invited guests includes people on the calendar invitation who didn't actually attend.
There's nothing inherently wrong with that. Catching up absent colleagues is a sensible use of notes. But it means the written version can reach someone who never heard the hesitation, the objection or the condition attached to a proposal.
Microsoft's intelligent recap can also draft an Outlook email containing a summary and follow-up tasks, ready to share with participants, including people outside the organisation.
I'd keep the distinction between “draft available” and “decision confirmed” painfully obvious. A recipient shouldn't need to reconstruct the meeting to discover that a confident-looking action was still awaiting approval.
Otherwise we've saved someone an hour in a meeting and given them an afternoon undoing it. A generous exchange, particularly for the calendar industry.
The paid version of the story
When the pitch is that AI will make meetings productive, I assume I'm hearing the version someone's paid to tell. Microsoft and Google both benefit if the useful record of a meeting stays inside their software.
Microsoft's intelligent recap sits within Teams Premium, an add-on licence, or a Microsoft Copilot licence. That doesn't make it bad. It does make the beneficiary rather less mysterious than the phrase “the future of work” suggests.
My read is that the more important contest isn't who writes the nicest paragraph. It's whose record becomes the one a business relies on afterwards.
Once notes become the place people look for tasks and commitments, the distinction between recording a decision and making one starts to matter commercially. I don't think a better-looking summary deserves that authority by default.
Nor do I want another expensive layer of software solemnly approving the output of the first expensive layer. We've already got committees. They're remarkably good at being available in additional layers.
Keep the useful bit, refuse the promotion
For a business owner, I'd judge an AI note-taker on an ordinary, slightly uncomfortable meeting. Not the demonstration where everyone agrees about the launch and nobody mentions the budget.
Give it a discussion with a rejected idea, a conditional commitment and an unresolved objection. Check whether each survives the summary. Check whether the person assigned a task actually accepted it. Check who gets the document before anyone's reviewed it.
Those are buying criteria, not a demand for perfect software. Human notes need checking too. The standard I'd insist on is simpler: don't let the quality of the prose disguise the status of the decision.
Let the machine draft the account. Keep an explicit human confirmation for consequential commitments, with the outstanding condition written beside them. If nothing was decided, “nothing decided” is a perfectly respectable result. Better that than a fictional breakthrough with three owners and a deadline.
The notes can be automatic. Agreement still needs a person.
Who gets paid for saving the meeting, and who gets paid for knowing what it actually decided?