AI meeting notes are a draft, not a record

The summary lands before you have finished packing up. It is tidy, it has a decisions heading and an action items heading, and it is roughly right. The problem is the part that is not, which arrives in the same calm typeface as everything else and includes a line saying you agreed to the July date. Three weeks later somebody quotes that line back to you, and you discover that a document nobody read has become the official memory of a conversation five people attended.

This guide is for solo operators and small teams who already run AI notes on their calls, or are about to, and want the output to be safe to rely on. It covers the consent sentence you say before recording, the four ways these summaries drift from what happened, and the short verification pass that converts a plausible document into a shared record. It is orientation, not legal advice. It is also not the right document if you are recording in a setting with its own rules on capture and retention, such as regulated financial or medical advice to consumers, where the question is which statute applies to you and the answer comes from a lawyer.

Consent comes before the button, and the rules are not uniform

Recording is the step people skip, and it is the only step with a statute behind it. Under United States federal law it is not unlawful for a person not acting under color of law to intercept a communication where that person is a party to it, or where one of the parties has given prior consent, unless the interception is for the purpose of committing a criminal or tortious act [7]. That is the floor, not the ceiling. States can be stricter and several are. California makes it an offence to intentionally record a confidential communication without the consent of all parties, with a fine of up to $2,500 per violation, rising to $10,000 per violation for someone with a prior conviction, and it defines a confidential communication as one carried on in circumstances that reasonably indicate a party wants it confined to the people in it, excluding anything said in a public gathering or an official proceeding open to the public [8].

You do not need a memorised map of jurisdictions. You need a habit that stays safe under the strictest rule you are plausibly standing under. Announce it at the start of every internal call, in one sentence, and ask rather than announce on any external call with a client, a candidate or a supplier. An audible yes costs five seconds. Recording someone who would have said no costs the relationship, and the summary was never worth that.

The tools will help, and they will not do this for you. Google Meet informs all meeting participants that notes are being taken and puts a pencil icon on every screen, and Google’s own instruction to the host is still explicit: “Once you turn on this feature, let everyone know you’re using ‘Take notes for me.’” [1] Meet notifies participants when a recording starts or stops, including people outside your organisation and people joining from the mobile app, who receive the notification but cannot control the recording [2]. A Workspace administrator can go further and require participants to click through an explicit consent prompt before recording, transcription or note-taking starts, though that setting is off by default [2]. Worth knowing, and worth saying out loud: in Meet, any participant internal to the host’s organisation can stop note-taking at any time, unless host controls are enabled for the meeting, in which case only the host and co-hosts can start and stop it [1]. Telling people that is what turns an announcement into consent.

Four ways a summary drifts from the meeting

The failures are patterned, which is the good news, because a patterned failure is a checkable one.

Decisions get promoted. Somebody floats an option, the room murmurs, the conversation moves on, and the summary records a decision. The model heard decision-shaped language and rounded up. NIST has a name for the general case: confabulation, “a phenomenon in which GAI systems generate and confidently present erroneous or false content in response to prompts,” with the specific risk that users believe the false content because of the confident nature of the response and go on to act upon it [6].

Owners get assigned by adjacency. “Someone should chase the vendor” becomes “Priya will chase the vendor” because Priya spoke next. The summary format has an owner column, and a blank owner column looks like a failure, so it gets filled.

Numbers and names get smoothed. A figure heard slightly wrong, a product name close to a real one, a date shifted by a week: none of these arrive flagged. They arrive fluent, in the same register as the parts that are correct. Google says so plainly in its own documentation for the feature: “Sometimes, the meeting summary can be incomplete, inaccurate, or not generated.” [1] Microsoft says the same thing about Copilot output generally, and draws the conclusion for you: “The responses that generative AI produces aren’t guaranteed to be 100% factual. While we continue to improve responses, users should still use their judgment when reviewing the output before sending them to others.” [3]

Disagreement gets averaged. This is the one that costs the most and shows the least. A good minute-taker records that two people left the room with opposite understandings. A summariser writes the middle of the room. The transcript still holds the disagreement; the summary has resolved it on your behalf.

There is a fifth failure that is not the model’s fault and is easy to walk into: the conditions the feature was built for. Meet’s note-taker supports one language at a time and does not currently support multiple languages spoken in the same meeting, it is recommended for meetings between 15 minutes and 8 hours, and it works in eight languages, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish [1]. A bilingual call, or a nine-minute stand-up, will still produce a document. It will just be a worse one, and nothing on the page will tell you that.

A four-minute check, while the meeting is still warm

Do this before you close the tab, not tomorrow. Memory is the input, and it decays fast.

Start by reading only the decisions and the action items. Skip the narrative summary entirely; it is the part least likely to be load-bearing and most likely to sound convincing. For each decision, ask whether the room decided it or discussed it, and delete or downgrade anything that was discussed. For each action item, check whether the owner was named by a human or inferred by the machine, and whether the deadline was said aloud or invented to fill the field. Then re-say every number that matters, out loud if you were in the room and by checking the transcript if you were not. Names of people and companies get the same treatment, because a name that is 95% right is a different name.

Then send the corrected version to everyone who attended, that day, with a line saying what you changed. This last step is the one that does the actual work. A summary that participants saw and corrected within the hour becomes the shared memory of the meeting. A summary that shipped unread becomes a document someone will one day quote at you, and by then the only thing that can contradict it is a recording you may not have kept.

Four minutes is a real budget, not a rhetorical one. It is roughly what it takes to read two short lists and check five facts, and it is small enough that you will still do it in a bad week.

The transcript is the only thing that can settle an argument

The summary is for acting this week. The transcript is for the argument you have not had yet. Keep both, and know where they live before you need them. In Meet, the notes document is generated shortly after the meeting ends and saved in the meeting organiser’s Google Drive in the “Google Meet” folder, and the people you share notes with get an email containing a link to the document, a meeting summary and suggested next steps [1]. That is convenient and it is also a distribution event: the moment that email goes out, an unverified document has an audience.

Use the transcript as the audit trail rather than as reading material. Every claim in the summary should be traceable to a line somebody said, and a claim that cannot be traced should come out. This is the same discipline NIST recommends in general terms, to “deploy and document fact-checking techniques to verify the accuracy and veracity of information generated by GAI systems, especially when the information comes from multiple (or unknown) sources” [6]. Applied to a meeting, it is narrower and easier: search the transcript for the number, the name, the word “agreed.”

Archive transcripts where your documents already live, with the date in the filename, searchable. Three weeks later a transcript settles in thirty seconds a question that two memories would litigate for a day.

The notes become other people’s data the moment an outsider speaks

A transcript of a client call is a recording of a named person discussing their business. It sits under whatever terms your account has, and those terms differ more than the interfaces suggest. Otter.ai’s privacy policy states that it improves and monitors its services by “training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information),” and it puts the permission burden on you: “please make sure you have the necessary permissions from your co-workers, friends or other third parties before sharing Personal Information or referring them to us” [5]. Microsoft takes the opposite position for Copilot, stating that “prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot” [3]. Both are legitimate products. They are not the same commitment, and you cannot tell which one you are using by looking at the button.

The free tiers also cap the thing you are relying on, which matters if your habit quietly depends on it. Zoom’s assistant is now ZoomMate, and its free Basic tier includes meeting summaries for 3 hosted meetings per month, AI note-taking for 3 uses per month across Zoom and third-party meeting platforms, 20 AI queries per month, and 40 minutes per meeting, while the paid tier is metered in AI credits at 2,200 per user per month [4]. Google’s note-taker requires an eligible Google Workspace edition or Google AI plan [1]. The failure mode here is not a bill. It is the fourth client call of the month producing no notes at all, and nobody noticing until the notes are needed.

Spend ten minutes once, in writing: which tier each account is on, what its terms say about training on your content, who can see the transcripts, and how long they are kept. Then stop re-asking the question every time you paste something.

What the summary cannot hear

The notes record what was said. They do not attend the meeting for you. The client who agreed in words and hesitated in tone, the teammate who went quiet after a decision and has not actually accepted it, the four minutes at the end where the real conversation happened after the agenda ran out: none of that survives into bullet points, and none of it is a transcription failure. It is a category the format does not have.

This is worth stating because the tool is good enough to feel total. It is not a reason to take notes by hand again. It is a reason to keep the part of the job that was always the valuable part, and to stop treating a completed document as evidence that the meeting was understood.

What still goes wrong

The check catches what it is aimed at and nothing else. Reading only the decisions and actions finds inflated decisions and guessed owners, and it will not find the decision the summary left out entirely, because there is no line there to be suspicious of. The only defence against omission is the transcript, and reading a full transcript costs more than the summary saved. In practice that trade is worth making for the two or three meetings a quarter where money or scope changed, and not worth making for the rest.

Consent has a harder residue. Announcing a recording is not the same as everyone being comfortable with it, and the person who minds most is the least likely to say so on a call with their client. Vendor notifications reduce the risk of someone being recorded unaware, but they cover their own product’s behaviour, not your obligations. Google’s consent prompt is a Workspace admin setting that is off by default [2], and California’s statute turns on whether all parties consented, not on what your software displayed [8]. If recording is central to how you work, the hour spent confirming the rules where you and your clients actually are is the cheapest hour in this guide.

The last limit is the one you cannot engineer around. Every verification step here assumes someone was paying enough attention to notice that a line is wrong. If the notes are the reason you stopped listening, the check has nothing to check against, and the document becomes both the record and the only witness to it. Vendors are unusually direct about this. Microsoft’s framing is that these features “provide useful drafts and summaries” while “giving you a chance to review the generated AI rather than fully automating these tasks” [3]. The chance is the product. Taking it is still your job.