An AI assistant can help with a private meeting without making a remote model the next recipient of the conversation.
Google AI Edge Foresight puts that idea into a familiar setting: a Mac, a meeting, and the files you use to do your job. Its arrival is a useful signal for business owners watching the move toward local AI. More of the technology is being designed to work near the information it needs.
The opportunity is practical. A firm can ask where a task should run before deciding which assistant should do it.
What Google Foresight does
Google announced Foresight on October 6, 2026 as an experimental Mac meeting companion. It combines Gemma 4 with EmbeddingGemma 2 to help expand shorthand notes and retrieve relevant material from conversations and files. Google describes local AI processing, offline operation, and no cloud subscription costs.
The Foresight product page says the app processes meeting audio and transcripts on the computer, can draw on local folders and Google Drive, and can surface context during a conversation. It is optimized for Apple silicon. These are Google's published descriptions; we have not tested the application or audited its network behavior.
The interesting part for a business is the placement of the processing. Your meeting and reference material can be useful to an assistant on the machine where you are working.
Privacy becomes a choice about the system
Consider a project meeting that includes a client's budget, an unresolved staffing problem, and a proposed commercial concession. The team wants a useful record of decisions. That does not necessarily require sending the full discussion to an outside inference service.
A cloud service's retention and training policies matter whenever information is sent to it. Local processing offers another option: keep that processing step within the environment the firm has chosen for the material.
This can reduce one form of exposure. It also gives the business a more concrete question to ask: which information crosses which boundary, for what purpose?
The answer needs to cover more than the model. Documents, transcripts, search indexes, generated notes, application logs, integrations, and backups can each have different destinations. A local model inside an app that uploads its notes is a different arrangement from a workflow whose notes remain on the device.
Google's Foresight FAQ says personal files, recordings, and transcripts are not sent to external cloud servers for its described local workflow. That statement should not be expanded into a claim that the application makes no network connections or collects no telemetry. Google Drive access also involves a cloud service. Review the actual configuration and behavior before using any app for confidential work.
The broader trend is local AI becoming part of ordinary tools
Foresight is one example of a wider product direction. Google also describes built-in AI in Chrome: browser-managed models that can perform supported tasks on the user's device. Its documentation identifies local processing, offline use, and reduced server round trips as benefits, while retaining server-side options for workloads that need them.
Apple's third-generation foundation-model announcement likewise describes a family spanning on-device models and models running in Private Cloud Compute. Local and remote processing coexist in that design.
Our reading of these developments is that local execution is becoming a normal product choice. These announcements do not establish how many businesses have moved off cloud AI, or that every workload is about to do so. They show major platform vendors investing in useful work that can happen on the user's hardware.
For a business owner, that expands the options. A meeting assistant, a document search tool, and a drafting workflow do not all need the same deployment model.
What “local” should mean for your firm
The term can describe several arrangements. Be specific about who uses the system and where its information goes.
| Arrangement | Where inference happens | What the business needs to decide |
|---|---|---|
| On-device | On an employee's Mac, PC, or phone | How devices, saved files, access, and backups are managed |
| On-premises | On a machine or server operated at the firm's premises | How staff share capacity and how permissions separate their work |
| Hosted service | On infrastructure outside the firm's premises | Which data is sent, the provider's terms, and the available service controls |
| Hybrid workflow | Different steps run in different locations | Which steps may leave the local environment, and how that rule is enforced |
A single-person desktop app can be useful without being a shared system for the whole firm. Shared retrieval adds questions about who can access each source, how permissions change when someone leaves, and how a system recovers after a failure.
Those are implementation decisions. They should be part of the brief from the beginning.
Start with work that benefits from staying close to its source
Meeting notes are a good illustration because the input is specific and the result can be reviewed. A sensible pilot might use synthetic conversations and sample project files to check whether decisions, owners, and deadlines are captured accurately.
Then ask the system to answer questions whose supporting material is known. Include a question the files cannot answer. Useful behavior includes finding the right evidence and recognizing when it is missing.
Other candidates include classifying incoming documents, extracting fields from standard forms, and drafting internal summaries. These are examples of workflows to evaluate, not claims that Foresight provides every one of them.
Choose a task with a clear acceptance standard. Compare the output with the team's existing process, measure how much correction it needs, and check performance on the hardware people actually use. A private answer still needs to be correct enough for its purpose.
Local processing changes operating costs and responsibilities
When a model runs entirely on hardware you control, that inference does not consume a hosted model provider's token allowance. Repeated work can be planned around the capacity of that equipment. We explored that distinction in our account of encountering a Dot usage limit.
The machine still has limits. Longer documents and simultaneous requests can increase processing time and memory use. Someone must manage updates, access, recovery, and support. Local deployment can make those responsibilities more visible and controllable; it does not eliminate them.
Similarly, an offline processing step does not make an entire business workflow independent of the internet. Retrieving a cloud file, joining an online meeting, or sending the final note can still require connectivity.
A useful direction for business AI
Foresight makes a compelling idea tangible: useful AI can work alongside private information on a device, with the processing close to the task.
At Looski, we see the same principle extending to firm-wide workflows. Start with the work people do, decide where its information should remain, and build the system around that boundary. Use local processing where it fits the quality and operating requirements. Make any external processing an explicit choice.
Our guide to designing a testable data boundary for private AI develops that approach. The question for your next AI project is straightforward: what needs to leave the firm for this job to get done?
Frequently asked questions
Does local AI mean that an application never connects to the internet?
No. Inference can run locally while integrations, updates, telemetry, or backups use the network. Evaluate those paths separately and test the configuration you intend to use.
Does this mean business AI is abandoning the cloud?
These product announcements show investment in local execution, not evidence of a wholesale move away from cloud services. A firm can assign different steps to local or hosted systems according to quality, data access, capacity, and operating needs.
Is a desktop AI app the same as an on-premises system for a team?
No. A desktop app operates in an individual device’s environment. A shared system also needs team access controls, capacity planning, recovery, and a way to maintain permissions across its source material.