The Summary Doesn't Leave Either: Why MinuteAI Runs Meeting Notes Through a Local LLM

Transcription Was Never the Whole Job

A raw transcript isn’t what anyone actually wants out of a meeting recorder. What people want is the three-line summary, the list of decisions, the action items with names attached — the part that turns forty minutes of dialogue into something you can read in twenty seconds. We’ve written about how MinuteAI keeps transcription entirely on-device, running Whisper through Core ML so audio never leaves a Mac or iPhone. But transcription is only half the pipeline. The summary is the other half, and it’s the half where a lot of otherwise privacy-conscious tools quietly change their story.

Where the Rest of the Market Still Draws the Line

Search around the current crop of AI meeting assistants and a pattern shows up quickly: “local” and “private” usually describe the capture step, not the whole pipeline. Several 2026 comparisons of tools like Granola, Otter, Fireflies, and Fathom describe a common shape — audio capture happens locally or silently in the background, but the transcript still gets sent off to a cloud LLM to generate the actual summary, action items, and notes. Granola is a good example of this pattern specifically: it captures notes locally without a bot joining the call, but the enrichment step that turns raw notes into a polished summary runs in the cloud. That’s a meaningfully different privacy posture than never having a transcript leave the device at all, even when the vendor is careful about retention policies and opts you out of model training.

The concern that’s pushing back against that shape isn’t hypothetical. Coverage of local-AI meeting tools this year has pointed to data sovereignty as one of the top objections IT and compliance teams raise about cloud-based meeting assistants — the worry isn’t just “does this vendor train on our data,” it’s “does our meeting content leave our infrastructure at all,” which a no-training policy doesn’t actually answer. Healthcare, legal, and financial teams in particular have been named as the earliest adopters of fully local tools specifically because a server-side promise isn’t the same guarantee as an architecture with no server in the loop.

MinuteAI’s Local Summarization Step

This is the gap MinuteAI’s architecture was built to close on both sides, not just the transcription half. As we described in the on-device transcription piece, MinuteAI’s pipeline splits work across two different local runtimes because transcription and summarization are different kinds of workload: Core ML handles the Whisper-based speech-to-text model, while a GGUF model running through llama.cpp handles the LLM-based post-processing — including summarization. Both stages run through the same Model Manager, loading lazily and unloading after inactivity, and both write into the same on-device, SQLite-backed result store. There’s no point in that chain where a transcript crosses a network boundary to get turned into a summary, because the model doing the summarizing is sitting in the same memory-mapped weights on the same device that did the transcription.

That’s the same instinct that shaped MinuteAI’s Chrome Extension, which streams captured audio to localhost instead of any remote server, and its speaker diarization feature, which ships pyannote.audio as an optional local download rather than a cloud API call. Every stage of turning a recording into usable notes — capture, transcription, speaker attribution, summarization — happens inside the same on-device boundary, not most of them.

What “Half-Local” Gets You (and Doesn’t)

None of this is a claim that cloud-based summarization is technically worse. The open-weight model families most local meeting tools lean on for this task in 2026 — Llama’s 3.x and 4.x generations chief among them — are smaller than the frontier cloud models a Granola or Fireflies can call on, and a quantized GGUF model running on a phone’s Neural Engine is not going to out-summarize a full-size hosted model on every edge case. That’s a real trade-off, not a marketing footnote, and it’s the same trade-off the diarization pipeline makes when it chooses a smaller local model over a larger hosted alternative.

What the local approach buys instead is architectural: there’s no step in the process where a company’s confidentiality policy, an NDA, or a client’s expectation of privacy has to be weighed against a vendor’s retention promise, because there’s no upload for that promise to apply to. For a meeting that touches deal terms, medical detail, or anything under an NDA, that’s not a nice-to-have — it’s the difference between a tool legal will sign off on and one they won’t.

Where This Fits in the Portfolio

Keeping the entire pipeline local — not just the parts that are easy to keep local — is the same discipline running through the rest of the AityTech portfolio. PrivateAI applies it to browser-based tools instead of a native app; Legacy Dragon applies it to parsing legacy source code. The common rule is the same one MinuteAI applies to its own summarization step: don’t stop halfway through the pipeline just because the second half is harder to keep local.

---

MinuteAI is available on the App Store for macOS and iOS, or you can learn more at getminute.app. Questions about how it fits your workflow? Reach out at [email protected].