Are AI notetakers secure? You are asking the wrong question

You have never once asked it on the call.

But you have typed it into Google, probably late, probably after a meeting where something confidential got said and the recorder was running.

are ai notetakers secure

I have been selling for a long time and I want to tell you something that took me a while to notice. Nobody asks that question in the room anymore.

The question did not go away. It went quiet.

A few years ago people pushed back on recorders. You would start one and somebody would shift in their seat. That mostly stopped. The tools got normalized, everyone has one now, and objecting to a notetaker feels about as current as objecting to email.

So on the call, silence. Total acceptance.

And yet hundreds of people a month still quietly type that exact phrase into a search box. The worry did not disappear when the objections did. It just went underground. It became the thing you wonder alone, at night, and never say in front of the buyer.

Which means it is still a live question. It just stopped being a spoken one.

So let me answer the private version, honestly, because the public version is not being asked.

Secure is the wrong word

It is doing too much work. Secure against what, exactly?

There are three completely different questions hiding inside that one, and they have different answers.

Who can see this. Where does it physically live. And does it get fed into a model.

Most people asking are ai notetakers secure never split those apart, so they get a vague reassuring yes from a vendor and go back to worrying anyway, because the yes did not actually touch the thing they were afraid of.

The one that actually matters

Here is the question I would lose sleep over, and it is the one almost nobody asks.

How much of your customer conversation have you handed to a frontier model, and did you decide that, or did a default decide it for you?

That is the real question underneath secure, and it is sharper than any of them, because it is not about whether a company is trustworthy. It is about how much you shipped out of your own control without ever choosing to.

Four tools, four completely different answers

So I sat down and read the policies, the actual ones, not the padlock pages. And the thing I found is that AI notetaker is not one security posture. It is four different ones wearing the same label.

Otter states in its own privacy policy that it trains its proprietary AI on de-identified recordings and transcriptions, and publishes no opt-out. Your sales calls are training material. I wrote about the exact clause, and where to find it.

Fireflies says the plain opposite in its policy: it does not use your content to train models, and it contractually blocks its vendors from doing so either.

Granola does not even keep the audio. It is discarded once the transcript exists.

Read that back. Same three words on the box. One trains on your calls with no way out, one refuses to train at all, one throws the recording away entirely. That is not a category with a security answer. That is three products with three opposite answers, and you agreed to whichever one you happened to install, on a screen you clicked through.

Everything in that paragraph was true when I read it in July 2026. Policies change. The point is not to trust my summary. It is to go read the one for whatever is recording your calls right now, which you almost certainly never have.

This is custody, not compliance

Strip the anxiety off the question and there is a simpler thing underneath it.

The data you hold about the people you sell to is the most important thing you have. What they pay today. What they tried and quietly abandoned. Who actually signs. What they told you in confidence about a competitor, or about their own boss.

You are holding all of that on their behalf. They did not consent to a vendor you chose.

That is a custody question, not a compliance question. Compliance is the answer you give an auditor. Custody is what you owe the person who told you the thing.

So what does Balance do, honestly

I am going to answer the real question and not the flattering one, because the real one is the whole point of this piece.

Balance does not train models. Not as a policy, as a fact: there is no model of mine for your calls to improve, because I am not in the business of building one.

The recording and the transcript are produced on your machine and they stay there.

And then the honest part. Turning that transcript into notes, into CRM fields, into extracted commitments, is an AI pass, and by default it goes to a frontier model under your own API key. So yes: by default, text from your call leaves your machine. I told you the real question was how much you hand to a frontier model, and I am not going to pretend my own default answer is zero.

Two things make that different from the tools above.

It is your key and your account. You are the model provider customer directly, reading their terms yourself, not a line item bundled inside a notetaker that decided for you.

And you can set it to nothing. Privacy mode runs the entire AI pass locally, on your own hardware, and then nothing leaves at all. I mapped exactly what runs where, and what it costs you, in its own piece.

So the defensible sentence is that the call never leaves your machine. How much of the text does is a dial, and unlike everywhere else, you are the one holding it.

The question to actually ask

If you are going to wonder are ai notetakers secure, stop asking the version that has no answer.

Do not ask whether the vendor is trustworthy. You cannot verify it, and every one of them will say yes.

Ask instead: how much of my customer conversation does this tool send somewhere I do not control, and can I turn that down to nothing? That question has a real answer, it is different for every tool, and it is the only one that maps to the thing you are actually afraid of.

Then go read the policy of whatever is running on your calls today.

One more honest note, because this piece is about not hiding things. Balance runs on Windows with an NVIDIA card today. macOS is designed and not shipped.

You have been asking this quietly for years.

Ask it out loud.