Apple Files Patent for a Switchboard That Routes App Questions to the Right AI Model
Apple has filed a patent application for a go-between that lets apps hand questions to the AI model best suited to answer, then keep the conversation going. The model remembers the earlier exchange, so your device doesn't have to resend it.
What Apple's AI request router actually does
When you ask an AI a follow-up question, somebody has to supply the earlier conversation, and resending all of it costs data, processing power and battery. Apple's filing describes a middleman built into the device that handles this for apps.
An app tells the middleman what it wants (its intent, such as answering a general knowledge question). The middleman picks a suitable AI model, opens a conversation with it and passes along the question. The model keeps a copy of the question and its own answer on its end. When the app follows up, the model already has the earlier exchange, so only the new question has to travel.
The filing says the chosen model can run on your device or on a different one. For general knowledge questions, the device can first check whether a model on the device can handle the job.
identifying a model based on the intent; creating the session, using an API, with the identified model; receiving a second request with a second intent from the software application; transmitting the second request to the identified model via the session …
Translation: The system figures out the right AI for the job and opens a connection to it.
How the app, the router and the model talk
The filing describes a go-between layer on the device. An app sends it a request tagged with an intent (a label for what the person wants, such as a knowledge question). If no conversation is open yet, the layer picks an AI model based on that intent and opens a session (a standing connection to that model) using an API (the set of rules apps follow to ask the system for help).
Requests then flow through that session. The model writes a response and stores both the request and the response. When the app sends a follow-up, the model builds its next answer from what it saved.
Narrower claims add several options:
- Choosing the model by intent type, a database of models, user preferences, or whether the size of the job passes a threshold.
- For knowledge-seeking intents, checking first whether a local model can answer.
- Running the model on the same device or on a second, different device.
Other claims cover images and documents: an app registers a file first, the model stores it and returns an identifier, and later requests point to that identifier, with progress updates during the transfer. Requests can also include tool functions (actions the model can ask the app to run), such as gathering user data or changing the state of the app or device. The app runs the action and returns the result to the model.
… systems that route requests to a remote model or to a local model based on an intent type. This application also includes disclosure related to routing of knowledge-based requests to remote model or to a local model.
Translation: Requests go to either an on device AI or a cloud based AI depending on what you need.
Why follow-up AI questions could cost less
The practical payoff is lighter, quicker follow-up answers. The filing says that because the model already holds earlier requests and responses, the device sends less data and uses less processing power and battery. For you, that could mean AI features inside apps that respond faster to a second or third question without draining your phone.
It also points to a privacy angle. The description says checking whether a local model can handle a knowledge question first reduces how much information leaves your device. Apps get one consistent way to ask for AI help, and the system decides where the work happens, which matters as more apps want AI features.
Apple's fifth patent we've tracked since October in our AI models working together watchlist builds on earlier applications, including one swapping a failing model and one routing to AI or rules.
The central trade here is memory for bandwidth. Letting the model store each request and answer saves your device from resending the conversation, but it means the history lives with the model, which may be on another device. That is a real cost for anyone who cares where their questions end up, and the filing's local-first check for knowledge questions is the main thing offsetting it.
The second cost is the guessing. The system must choose a model from a label like an intent type, a user preference or a size threshold. A wrong guess sends a question to a model that is too weak or too heavy, and the person only sees a slow or poor answer.
The trade reads as worth it. Follow-up questions are the most common thing people do with AI, and shipping the whole backstory every time is wasteful. The design asks for trust in the router, which is a fair price for the savings.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Apple patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
33 drawing sheets from US 2026/0350061 A1 · click any drawing to enlarge
Show the other 25 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in