Google Bets Your Meetings Should Never Leave Your Mac: AI Edge Foresight Works Entirely Offline

Most AI meeting assistants do their thinking somewhere else. Your audio travels to a data center, gets transcribed on someone else's servers, and comes back as a summary. Google is now testing what happens when that entire pipeline stays on your laptop. The company has released AI Edge Foresight, an experimental Mac application that records, transcribes, and organizes meetings entirely on the device, with no cloud processing and no internet connection required.

The release was reported by TechCrunch and The Verge on October 9, and it marks the first time a company at Google's scale has shipped a full meeting note-taker as a local-first product. It arrives at a crowded moment. Wispr and Calendly have shipped their own meeting assistants this year, Superhuman (the company formerly known as Grammarly) acquired AI note-taking startup Fathom in September, and OpenAI is pushing ChatGPT deeper into the office routine with its own Meetings plugin. Google's answer is to change the axis of competition entirely: not better summaries, but summaries that never leave the room.

What Foresight actually does

At its core, Foresight is a familiar kind of product with an unfamiliar architecture. It captures meeting audio, whether from a video call or an in-person conversation, and turns it into structured notes. The interface uses a split-screen layout: the user writes shorthand notes on one side while the AI generates a fuller version alongside. Full transcripts are available, and a built-in conversational assistant can answer questions about what was discussed. The design closely mirrors Granola, the independent note-taker that pioneered the bot-free approach of capturing audio from the computer instead of joining the call as a visible participant.

The difference is under the hood. Foresight runs on Google's EmbeddingGemma 2 model, which has 740 million parameters, and its question-answering assistant is powered by a Gemma 4 model. Both run locally, optimized for Apple Silicon, and Google says meeting audio, notes, and files never leave the user's computer. The app was developed by the same team behind Google's experimental AI dictation application released earlier this year, and it is listed on Google Labs as an experimental on-device meeting companion. Google has not said whether Foresight will become a mainstream product or be folded into its existing Gemini applications.

The knowledge base is the real bet

The transcription features, while competent, are becoming table stakes. The more consequential part of Foresight is its local knowledge base. Users can import PDFs, Google Docs, Microsoft Office files, plain text, Markdown documents, and saved web bookmarks. The assistant can then draw on those materials when a related topic comes up, for example pulling a detail from a proposal you uploaded while the meeting is still running. Documents and transcripts become one searchable memory instead of two separate systems.

This points at where workplace assistants are heading. Summarizing a conversation is now a standard feature across productivity software. Retrieving the relevant clause from a contract, comparing a current discussion with notes from a meeting three months ago, or answering a question from a collection of local documents is the harder and more valuable problem. Foresight's bet is that the assistant that already holds your documents will beat the assistant that only heard your call.

Google is not just selling a note-taker. It is arguing that trust, control, and offline resilience are features worth switching for.

There is a reason that argument is arriving now. Meeting data is unusually sensitive: customer plans, legal advice, hiring decisions, financial information, and internal strategy all surface in ordinary conversations. Professionals who cannot, or will not, send recordings to a third-party cloud service have had few good options. A local-first assistant speaks directly to lawyers, doctors, government workers, and anyone whose compliance team has opinions about where data goes. It also keeps working on a plane, which cloud assistants cannot.

What the local tradeoff costs

Running models on a Mac instead of a data center is not free. A 740-million-parameter model is small by modern standards, and smaller models are less capable than large cloud systems at reasoning, nuance, and long-context understanding. Long meetings and large document collections consume memory, storage, and battery power, and local processing can mean slower real-time responses on older machines.

Early commentary has been cautious. Coverage of the launch raised questions about transcription accuracy, the handling of overlapping speakers, battery consumption, response time, and the extent of the app's integration with other tools. Those concerns cut to the core of the product, because a polished interface cannot compensate for missing words, confused speakers, or answers drawn from the wrong document. Mac-only availability also limits the audience from day one: this is a product for a subset of knowledge workers, not a platform play.

Why this launch matters beyond meetings

The deeper significance is what Foresight demonstrates about Google's Gemma models. Google has spent the year building smaller models capable of running on personal devices, and Foresight is the first consumer-facing application that asks a Gemma model to carry an entire workflow without a network connection. The company is effectively using meetings as the proving ground for on-device AI as a category.

The history of Google's experimental local AI tools is mixed: the dictation app that preceded Foresight never reached mainstream usage. Technical demonstrations do not automatically become durable products, and Google's experimental software has a long record of quiet retirements. But the strategic framing is clear. Cloud systems will keep winning on scale, collaboration, and raw model power. Google is betting that a meaningful slice of the workplace market will choose the opposite tradeoff: local systems that win on trust, control, and the ability to work anywhere. Whether enough users care about those qualities to turn offline AI from a showcase into a practical alternative is now an open question with a real product attached. The answer will shape how the next generation of workplace software is built.