There's a moment in every demo of a cloud-based AI scribe that nobody dwells on. You hit record, the session happens, and then the audio goes somewhere. A server, in a data center, owned by a company you signed up with three weeks ago.
Everything else about these products — the note quality, the pricing, the integrations — is downstream of that one moment. And it's worth being precise about what it costs, because the industry has gotten very good at making it sound like it costs nothing.
What uploading actually commits you to
Once a recording of your session leaves your office, four things become true that weren't true before.
You now need a Business Associate Agreement. The vendor is creating, receiving, and maintaining PHI on your behalf, which makes them a business associate under 45 CFR 160.103. That's not a formality — it's a contract that has to be in place before they handle any of it, and you're responsible for making sure it exists.
Their breach is your breach. If the vendor is compromised, the notification obligations run to your clients. You will be the one making those calls, explaining a company your client has never heard of.
Their retention policy is now your retention policy. How long they keep audio, where backups live, which subprocessors get access, what happens to the data if the company is acquired — these are now facts about your practice, and most of them live in a terms page that can be revised.
You've added a dependency to a clinical workflow. No internet, or a vendor outage, means no notes that day.
None of this makes cloud tools reckless. Plenty of them are run carefully by people who take it seriously. The point is narrower: this is a real and permanent set of obligations, and it is being accepted in exchange for something that no longer requires it.
The assumption that expired
The reason every one of these tools started as a cloud service is that, until recently, it had to be. Speech recognition and language models needed hardware that didn't fit under a desk. Uploading wasn't a business decision; it was the only way the thing could work at all.
That stopped being true. An Apple-silicon Mac with 16GB of memory will run a capable speech-to-text model and a language model large enough to write a competent clinical note — locally, in seconds, with no network involved. The hardware in front of you is enough now. The architecture just hasn't caught up, because it's genuinely harder to ship a model than to call an API, and because per-seat cloud subscriptions are a much nicer business to be in.
What changes when the audio stays put
If the recording never leaves the machine, the four obligations above don't get managed better. They stop existing.
There's no BAA to negotiate, because no vendor receives PHI. There's no vendor breach that could expose your clients, because there's no vendor-side copy. There's no retention policy to audit, because the records are files on your own disk. And the workflow keeps working on a plane, in a basement office, or during an outage.
It also changes the honesty of the privacy conversation. A cloud vendor promising not to misuse your data is making a policy commitment — sincere, probably, but revisable, and only as durable as the company. "We can't access it" is a structural claim. We can't leak, sell, subpoena, or get breached out of data we never receive.
The honest trade-offs
Local isn't free of costs, and you should know them before switching.
You need reasonably current hardware — an Apple-silicon Mac, ideally with 16GB of memory or more. There's no web version and no Windows build, because there's no server doing the work. Setup includes a one-time model download of a few gigabytes. And your backups are genuinely your responsibility; nobody else has a copy, which is the entire point and also the risk.
For a lot of clinicians those are easy trades. For some they aren't, and that's a legitimate reason to choose a cloud tool with a solid BAA and read their security documentation carefully.
What isn't legitimate is the framing that uploading is simply how this works. It was, for a while. It isn't anymore, and it's worth asking any vendor why they still need to.