AI Meeting Notes: The Real Question Is Where Your Data Goes

One of the first things we ever built was called Summit: automatic meeting notes, powered by speech-to-text, back when that still felt like magic. Years later, the world caught up. Today, AI meeting notes are a commodity — Otter, Fireflies, Granola, Loom, and the AI summaries now baked into Zoom and Teams. We use Loom ourselves.

So we don't sell a meeting-notes product anymore, and for most teams you shouldn't buy a custom one either: pick a good off-the-shelf tool and move on. But there's one question those tools rarely answer well, and for some of our clients it's the only one that matters: where does your meeting data actually go?

---

The hard part isn't transcription. It's privacy.

When you drop a meeting into an AI notetaker, the audio and the transcript usually leave your walls. They're uploaded to a vendor's cloud — often outside the EU — processed by models you don't control, and sometimes retained or used to improve the product. For a weekly standup, fine. For a board meeting, an M&A call, a patient consultation, a legal strategy session, an HR investigation, or anything covered by an NDA or the GDPR, that's a real problem.

"Just turn training off" isn't an answer when the data still crosses borders and lands on infrastructure you can't audit. The transcription is solved. The data governance around it usually isn't.

What we build when data can't leave your walls

For privacy-sensitive teams, we build automatic meeting notes that run entirely on infrastructure you control:

• Self-hosted Whisper for speech-to-text, so the audio never leaves your environment.
• Speaker diarization so the transcript knows who said what, not just what was said.
• Open-source LLMs running on your own or EU-hosted GPUs for the summary, decisions, and action items — no third-party API, no data sent to an external SaaS.
• Everything on-prem, in your own cloud tenant, or on EU infrastructure, with retention and access entirely under your control.

The output looks like any good notetaker — a clean summary, decisions, action items, a searchable transcript. The difference is that the data never crosses your perimeter. That's the version nobody sells off the shelf, and it's the version we still build.

Should you build this? Usually not.

We'll be honest: if your meetings aren't sensitive, use Loom or whatever your team already likes, and skip the custom build. AI meeting notes are a solved problem for the common case, and paying to rebuild a commodity is a waste.

Build custom only when the privacy constraint is real — a regulated industry, confidential data, GDPR exposure, or a client contract that forbids sending data to third parties. That's the line. If you're on the wrong side of it with an off-the-shelf tool today, you may have a compliance problem you haven't noticed yet.

If your meeting data can't leave your walls

That's exactly what we build: self-hosted, open-source, on your own infrastructure. If confidentiality is the reason the commodity tools don't work for you, reach out and tell us the constraint — we'll tell you honestly whether a custom build is worth it.

Related: how we benchmarked transcription on messy real-world audio, and why we run our own GPU inference in Europe.