The most interesting thing happening in consumer and business apps right now is a retreat from the cloud. Three recent launches point the same direction: AI features are being pushed onto the device, handed real tasks to execute, and trained on an individual's own context rather than a generic model's. The competitive ground in apps is shifting from who has the biggest model to who controls where the computation happens and what it is allowed to do.
Offline as a Feature, Not a Limitation
Google's new AI Edge Foresight app is the clearest example. As TechCrunch reported, it is a local-first competitor to Granola that transcribes conversations, generates notes and answers questions using on-device AI. That framing matters for the Apps & Services beat because it changes what an app is selling. Meeting note-takers have been a cloud product by default: audio goes up, a transcript comes back, the vendor pays for inference on every minute recorded. Foresight moves that work to the phone or laptop.
For US consumers, the practical pitch is reliability and privacy. A note-taker that works without a connection is useful on a plane, in a basement conference room, or anywhere the Wi-Fi is unreliable, and it does not require handing a recording of an internal meeting to a third party. For US app companies, the economics are just as important. On-device transcription removes a per-minute inference cost that scales with usage, which is precisely the cost structure that has made AI features hard to price in consumer subscriptions. Google is not the first to try this, but a platform owner shipping an offline competitor to a well-known startup signals that local processing is now a mainstream product decision rather than a technical curiosity.
Agents Get Jobs, Not Just Answers
The second thread is autonomy. TechCrunch reported that Google is turning Gemini into an AI agent that can plan, execute tasks and work across business apps and systems. The agent can delegate work to subagents, use multiple AI models and gets its own workplace identity, complete with an email address.
That last detail is the one to dwell on. An email address is not a feature; it is an identity. It means the agent can be addressed, assigned work, held accountable in a directory and granted permissions the way a human employee is. For US businesses evaluating software, this raises questions that ordinary app procurement never had to answer: who owns the agent's mailbox, what happens to it when an employee leaves, and how it is audited. It also means the agent competes for seats in the same way a colleague does. Google starting with businesses, rather than consumers, reflects where the willingness to pay and the tolerance for imperfect automation currently sit.
Subagents and multi-model routing point to the same conclusion. The value is no longer in a single model's output but in the orchestration layer that decides which model does what and how work gets broken up. That is an app-layer business, and it is where the differentiation is being built.
Personal Context Is the New Moat
The third launch is smaller but fits the pattern. TechCrunch reported that Zach Yadegari, the teen co-founder of the Cal AI calorie tracking app, raised $10M for a new personal AI agent startup competing with Instinct, Muse and Bee. A young founder with a consumer track record moving into personal agents suggests the category is attracting talent and capital on the strength of a simple premise: an agent that knows your habits and history is more useful than a general assistant that does not.
