AI Edge Foresight: Offline Meeting Notes With Governance Gaps

Google has released AI Edge Foresight, a free, experimental Mac app that transcribes meetings and turns rough notes into polished summaries without sending audio to the cloud. That is a meaningful privacy shift for AI meeting assistants. The same local-first design, however, raises a workplace question the launch materials don't settle: who knows the app is listening, and who controls it?

What Google shipped

Google launched Foresight on Tuesday, October 6, 2026, alongside a new open model, EmbeddingGemma 2. 1 The app runs on a 740-million-parameter version of that model, is optimized for Apple Silicon, and can work with no internet connection. 25 Google's FAQ highlights offline use as a feature for travelers and people with spotty connectivity. 4

Google calls Foresight a "context-aware meeting companion" powered by "fully local AI processing." It captures both system audio and the microphone, so it can work with any meeting platform. 3 It also works in person, not just on video calls. 25

The main feature is turning shorthand into notes. You jot terse fragments, and Foresight fills them in with details it pulls from the live conversation. 31 The layout uses a split screen, with your manual notes on one side and AI-generated notes on the other. Several outlets compared this directly to Granola. 25

During a meeting you can review the running transcript or query an assistant built on Gemma 4. That assistant draws on your notes and any documents you have added to the app. 5 Google also describes natural-language search across images, documents, transcripts and notes. It says the app can detect questions raised in a meeting and answer them in real time. 3

Why local processing matters

The architecture is the real story. Most AI note-takers, including Granola and Google's own Gemini notes in Meet, capture audio on the device but process it on company servers. 1 Foresight keeps the audio, transcript and analysis on the machine. The Verge described the app as transcribing "entirely offline." 1

For regulated industries, legal teams, and anyone wary of sending confidential conversations to a vendor, this changes the risk. There are fewer third-party processors, no cloud retention policy to vet, and no transmission to intercept.

The governance blind spot

The same design also removes the friction that normally signals recording. Cloud assistants often join calls as a visible participant, and enterprise tools can come with centralized oversight. Foresight, by Google's own description, simply taps the system audio and microphone of whatever meeting is happening. 3

None of the coverage describes a mechanism that alerts other participants, or a way for IT administrators to manage or limit deployment. That is an inference from what Google has published, not a confirmed specification. Still, it points to a real gap. Privacy for the person running the app is not the same as transparency for the people being recorded.

The local-first approach protects the user's data from Google, but it does nothing to tell a colleague or client that their words are being transcribed and mined for live answers. In-person use raises the stakes further, because there is no meeting interface to display any indicator at all. 25

The business-focused analysis from Increscotech frames this as a policy problem. It urges organizations to adopt a clear rule before AI starts listening to their meetings. 3 Progressive Robot's coverage likewise flags "privacy limits" alongside the app's strengths. 1 Both suggest that on-device processing narrows the data-exposure risk while leaving consent and disclosure squarely with the user and the employer.

Experiment or product?

Outlets differ slightly on how seriously to take Foresight as a Granola rival. ProPakistani and ua.news present it as a direct competitor, echoing the interface comparisons. 25 TechCrunch is more skeptical. It notes that Google's earlier experimental dictation app never reached mainstream use, and suggests Foresight may mainly showcase what Gemma's offline models can do. 4 The "experimental" label Google applies supports that reading. 3

The takeaway

Foresight looks like a capability demo more than a finished enterprise product. Its significance is that it shows capable meeting AI can run fully on a laptop. That will likely push competitors toward local processing and reassure privacy-conscious users.

But it also shows that moving AI on-device shifts governance problems rather than solving them. Workplaces adopting tools like this should not treat "offline" as a substitute for disclosure norms and clear usage policies. The audio may never leave the machine, but the people being recorded still deserve to know it exists.