Offline AI meeting notes: what actually runs on your machine

Nothing happens.

That is the honest answer to what Balance does to my computer during a call. The fans do not spin up. Zoom does not go choppy. The transcription does not quietly die twenty minutes in and leave me holding half a meeting. I run it on every call I take and I have noticed exactly zero machine tax.

I am also the worst possible person to ask.

I am running this on a computer that cost the better part of seven thousand dollars, with a GeForce 5090 in it. My results are not your results. Any founder who tells you the local AI thing runs beautifully, without telling you what it runs on, is selling you their hardware budget as a product feature.

So here is the actual map of what runs where.

What is running while you talk

Three things, all of them on the machine in front of you.

Your computer is already playing the other side of the call, and your microphone is already picking up your half. Balance takes both of those streams directly off the machine. Nothing joins the meeting to do it, which is a whole separate argument I have made elsewhere.

That audio hits a speech model running locally and becomes a transcript, live, while you are still talking. Then the meeting gets indexed locally too, so that searching across everything you have ever recorded is a thing your own machine does.

Capture, transcript, search. All of it on your side of the wire.

It sizes itself to your box

Here is the part nobody in this category talks about, and it is the actual answer to what runs on your machine.

When Balance starts, it asks your hardware what it can genuinely do, then picks a transcription model to match. Not what your driver claims. What actually loaded.

On my card, that is the large model. Around a gigabyte on disk, wanting roughly four gigabytes of video memory to run.

On an older laptop with four cores and no GPU worth using, it is the tiny model. Thirty one megabytes.

Same app, same install, very different transcript.

Underneath all of it there is a floor, and the floor is your CPU. If nothing else loads, that is what runs, and it always runs. Slower, smaller model, but a real transcript of a real conversation rather than an error message.

There is also a rung we deliberately do not take automatically. Balance can run on non-NVIDIA graphics, and you can switch that on yourself. Auto-select refuses to, because nobody has yet put a real meeting through that path on real AMD or Intel silicon. Promoting a path that has never transcribed an actual conversation is how somebody loses the first recording they ever tried to make. Slow and verified beats fast and assumed.

What this costs you if your laptop is not mine

Three things I would want to know before downloading anything.

Windows, on an NVIDIA card, is the entire supported surface today. macOS is designed and not shipped. If your team is on Macs, this is not for you yet, and I would rather you read that here than discover it during setup.

A weaker machine gets a smaller model, and a smaller model is a worse transcript. Everything downstream of the transcript inherits that. This is not a licensing tier or a paywall, it is arithmetic.

And the cloud is a switch, not a rescue. There is a cloud transcription option, it works, and you turn it on yourself with your own key. Nothing silently degrades to it when your machine struggles. You also give up the live preview when you use it, because that pass stays local.

The word offline is doing too much work

If you searched for offline AI meeting notes, you were probably asking one of two questions, and they have different answers.

The first: does a recording of my customer call end up sitting in a vendor cloud? It does not. The audio and the transcript are produced on your machine and they stay there.

The second: can I pull the network cable and have all of this still work? By default, no. This is where pages like this one usually start lying, so let me be plain.

Turning a transcript into notes, into CRM fields, into extracted commitments, into a name attached to each voice: that is an AI call. Out of the box it goes to Anthropic, under your own API key. The call never leaves your machine. That work does.

So the sentence that survives scrutiny is the call never leaves your machine, and not Balance is offline. It can be offline. That is a setting, and you should understand it before you flip it rather than after.

I run it further left than it ships

There is a privacy mode. Switch it on and every AI task is force routed to a model running on your own hardware, no matter what else is configured. Nothing goes out. That is not a roadmap sentence, it works today.

It is also how I run my own company. When I am using Balance as my actual relationship manager rather than testing it, I am running the whole thing on my own card with a local model and zero compute borrowed from anywhere else. No token bill. Nothing leaving the building.

Which means the default this product ships with is not the setting I personally use.

That seems worth saying out loud rather than letting somebody discover it. The default is the frontier model, because for most people, most of the time, it is the better answer. The local path is there for when it is not, and I am one of the people it is not.

One detail I care about more than it probably deserves: the local model address defaults to loopback. Your machine, and nothing else. Not a convenient address somewhere on the office network. Local should mean this box, because a network address would ship your transcripts to a different computer and feel identical while doing it.

Nothing here is tied to anyone

Local model, Anthropic, OpenAI, Gemini, or any endpoint you choose to point it at. You bring the key. We are not the vendor of the intelligence, and we do not want to be.

If something better or cheaper lands next quarter, that is a two minute change in settings. Not a migration. Not a support ticket. Not a renewal conversation where somebody explains why switching is complicated.

You pay the model provider directly, at cost, rather than paying a margin on top of it. That is the same argument I made about what a CRM actually charges you for, pointed at the compute instead of the seats.

Customer data is not yours to be casual with

Strip the privacy conversation of its usual anxiety and there is a simpler thing underneath it.

The data you hold about your customers is the most important thing you have. What they pay today, what they tried and abandoned, who inside their building actually decides, what they told you in confidence about a competitor. You are holding that on their behalf.

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

What you keep when you stop paying

Here is the test I would put to any tool holding your relationships.

Stop paying for it. What do you still have?

If Balance goes away, or you walk away, everything remains local on your device. The recordings, the transcripts, the history of who promised what. It was on your disk the whole time. Nobody has to grant you an export window, because nobody was holding it hostage in the first place.

That is the one claim in this category that a cloud tool cannot make no matter how good its intentions are. Not because they are villains. Because of where the files live.

Your machine already did the work.

It stays where you put it.