Local capture is only the first boundary
A product can record audio on-device and still send the transcript to a remote model. To understand the real privacy model, ask where every processing step runs.
Transcription and language-model inference are separate
Speech recognition turns audio into text. Summary generation and question answering reason over that text. Both can run locally, but they use different model runtimes and resource profiles.
Semantic search also has a privacy boundary
Cross-meeting retrieval can require embeddings and an index. A local-first design keeps the embedding model, vector data, full-text index, and retrieved evidence on the user's computer.
Model downloads are not meeting uploads
Downloading a model or checking for an application update requires network access, but it does not require transmitting meeting audio, transcripts, summaries, or screenshots to the model provider.
Recording consent still applies
Local processing improves data control but does not replace applicable laws, workplace policies, contracts, or participant-consent requirements.