Local-first AI for product teams, and why your security review keeps saying no

Product managers in fintech, health, and most of Western Europe run into the same wall. The tool is good, the team wants it, and then security asks where the recordings live. The honest answer is usually "in our cloud, in a region you do not choose, processed by subprocessors we may change."

Three separate objections, usually collapsed into one

1. Where the data sits

A meeting transcript is one of the densest confidential artefacts a company produces: roadmap, pricing, named customers, unreleased strategy, sometimes personal data about employees. Uploading that to a third party is a real transfer of risk.

2. Who else is in the room

A recording bot joining a call with an external participant is a consent problem, not just a storage one. That participant never signed your vendor agreement, and in several European jurisdictions their consent is not yours to give.

3. What the model does with it

"We do not train on your data" is a policy, and policies change with a version bump and an email you will not read.

The shape that answers all three

Keep the work on the machine that already has clearance:

• Record locally. No bot joins, so no external participant is processed by a third party, and the audio never leaves the laptop.
• Store locally. Transcripts, the commitment ledger and the context file live on disk, on a machine already covered by your device policy.
• Send the minimum. Matching a commitment across meetings needs language understanding, so the commitment line goes to a model. The transcript does not.
• Use the account you already cleared. If the company already approved a Claude or OpenAI-compatible subscription, running through it adds no new vendor to the review.

What this costs you, honestly

Local-first is not free of tradeoffs, and pretending otherwise is how trust gets lost:

• No cross-device magic. The ledger lives where the app lives. Sync is a feature you must design, not a default.
• Not fully offline. Cross-meeting matching still calls a model. Anyone promising both full offline operation and semantic matching is describing something else.
• Your machine does the work. Local recording and transcription use local resources.

The question to put in front of a vendor

Not "are you secure," which everyone answers yes to. Ask instead:

If I uninstall your product tomorrow, which copies of my meeting content still exist, on whose infrastructure, and how do I delete them?

For a cloud note taker the honest answer involves a retention window, a support tier, and probably a ticket. For a local-first tool the answer is: you already deleted them, because they were only ever on your disk.