Google AI Edge Foresight: A Local Meeting Note Taker, and When Your Board Should Use One

Google AI Edge Foresight is a Mac app that transcribes meetings, writes notes and answers questions without sending audio or notes to the cloud. Google calls it experimental. It arrived on October 6, 2026, alongside EmbeddingGemma 2, an open model that Foresight uses for search.

For nonprofits and associations, the interesting part is where the data sits. Board and committee meetings cover budgets, staff issues and member records. A note taker that keeps all of that on one laptop changes the questions you ask before you press record.

The short version: Foresight is worth testing for confidential meetings run by one note taker on an Apple silicon Mac. It is not ready to be your organization’s standard tool. Google has not published system requirements, a price policy, or how its Google Drive connection works. And because it runs quietly on one computer, the people in the room will not see a notice unless you give one.

This post covers what the app does, which models do the work, what Google has and has not said, when local makes sense for a board, and how it compares with Gemini in Meet, Zoom, Otter and Granola as of October 2026.

What Google AI Edge Foresight does

Google’s Foresight product page describes three features.

• Notes from your shorthand. You type quick bullets during the meeting. Foresight listens in the background and turns them into full notes using the transcript.
• A local knowledge base. You connect project folders, reference files and Google Drive. It reads PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks.
• Live answers. You can ask questions in a chat. A Live Assistance mode listens and shows answer cards, with citations, when a topic connects to your files.

It captures your Mac’s microphone and system audio. That means it works with in-person meetings and with video calls. The FAQ names Google Meet and Zoom. Microsoft Teams is not named, but Google’s blog says it works with any meeting platform.

There is no meeting bot. Nothing joins the call as a participant.

What runs on the Mac, and what stays there

Google’s claims here are clear. The Foresight FAQ says "All AI inference runs 100% locally on your device using on-device Gemma models." The product page says files stay on your Mac. Google’s developer blog post says "your sensitive data never leaves your device."

The app also works offline. The FAQ mentions flights and air-gapped facilities.

One feature needs a closer look. The product page says you can connect Google Drive and that Foresight reads calendar details. Google does not explain how that connection works. Reading files from Drive normally means signing in and going online. If your goal is to keep a meeting fully off the network, test what Foresight does with Drive connected and with it disconnected.

The page also says nothing about usage statistics or crash reports. The local inference claim covers the AI work. It does not tell you what else the app reports. Ask your IT lead to watch its network traffic during a test meeting before you rely on it for a closed session.

Which models do the work in AI Edge Foresight

Early coverage, including TechCrunch’s October 8 report, said Foresight uses EmbeddingGemma 2 to capture notes. Google’s own pages describe a split.

EmbeddingGemma 2 is an embedding model. It turns text, images, video frames and audio into numbers that make search work. It does not write sentences. Gemma 4 does the writing.

The Google DeepMind announcement adds detail. EmbeddingGemma 2 has 740 million parameters, an 8K-token context and an Apache 2.0 license. It is built on the Gemma 4 architecture. Weights are on Hugging Face and Kaggle. On a Pixel 11 Pro, Google measured about 567MB of memory for the full multimodal model.

Google does not say which Gemma 4 size Foresight runs. So nobody outside Google can yet say how much memory a Mac needs. Our post on local AI hardware prices in 2026 explains why that matters before you buy.

What Google has not said yet about AI Edge Foresight

Several basic facts are missing from Google’s pages as of October 10, 2026.

• System requirements. The FAQ says Foresight is “optimized for Apple Silicon.” No minimum macOS version, memory or disk space is listed. Intel Macs are not mentioned.
• Price. The download is a direct disk image from Google. Google’s pages list no price and no paid tier. Android Authority reports it is free.
• Status. The developer blog calls Foresight “an experimental app.” The product page does not repeat that. Experimental apps can change or end without much notice.
• Distribution. It is not in the Mac App Store. It downloads from Google’s developer site.
• Data controls. Nothing on export, deletion, retention settings or admin management.
• Languages and speakers. No list of supported languages. No mention of telling speakers apart.

TechCrunch also reports Foresight comes from the team behind an offline dictation app Google released in April. It suggests the app may mainly exist to show off Gemma models. That is a reasonable read of an experimental release, and a reason not to build a process around it yet.

When a local note taker makes sense for a board

Most association meetings do not need a local tool. A staff stand-up or a webinar debrief is fine in Gemini, Zoom or Otter, as long as your settings are right. Local starts to matter when the content is sensitive and the group is small.

Good fits for AI Edge Foresight:

• Executive sessions. Closed board sessions on the CEO’s review, legal matters or a merger.
• Personnel and finance committees. Salaries, audits, investment discussions.
• Member data reviews. Ethics complaints, credential appeals, award juries.
• Offline settings. Retreats and hotel ballrooms with poor Wi-Fi.

Poor fits:

• Meetings where many people need the notes. Foresight keeps notes on one Mac. Sharing them is a manual step, and a copy then lives somewhere else.
• Organizations on Windows. It is Mac only.
• Anything that must be supported long term. Experimental means no promise it will be here next year.

We mapped this same local-versus-cloud question for AI agents in where each AI agent actually runs. The pattern holds. Local reduces who can see your data. It also makes one device, and one person, responsible for it.

Five things to check before AI Edge Foresight records

None of this is legal advice. It is housekeeping that keeps a local note taker from creating a new problem.

1. Tell people. Gemini in Meet shows every participant that notes are being taken. Foresight runs on one laptop with no bot, so nobody sees anything. Have the chair say it at the start, and note it in the minutes. Check your bylaws or counsel if you are unsure what your organization requires.
2. Decide what is kept. Minutes are usually the official record. A full transcript is a different, much longer record. Decide whether transcripts are deleted once minutes are approved, and who does it.
3. Know where the files sit. Notes on the Mac can still end up in Time Machine backups or a synced folder. Check before a closed session, not after.
4. Lock down the Mac. Turn on FileVault disk encryption and a strong login. A lost laptop holding executive session transcripts is a bigger problem than a lost laptop holding email.
5. Plan for staff changes. If the board liaison leaves, their Mac holds the transcripts. Add it to your offboarding checklist.

Local tools carry their own risks too. Our post on local AI agent security covers why running a model on your own machine does not make it secure by default.

How AI Edge Foresight compares with cloud note takers

Here is how the main options handle data, based on each vendor’s own pages as of October 2026.

Sources: Google Meet help, Zoom’s June 2026 blog post, Otter’s privacy page and Granola’s security page.

The cloud tools have one big advantage: admin controls. Meet notes follow the retention policy you already set in Workspace. Our guide to Gemini defaults in Google Workspace shows where those settings live. Foresight has nothing comparable yet.

Foresight’s advantage is simpler. The audio and transcript never go to a vendor. For a closed session, that removes a whole category of questions.

If you also use transcripts for captioning, note the different job. Our post on transcription model retirement dates covers the cloud speech models agencies build on.

Frequently Asked Questions