Apple Foundation Models: local AI is becoming an application platform
Apple’s Foundation Models framework is not just another system AI feature. It points to a more practical direction: local and private AI as a normal layer inside apps, exposed through Swift, structured output, tool calling, multimodal prompts, and the option to run on device or through Private Cloud Compute.
For companies, this is more interesting than a typical model announcement. If AI capabilities move into the operating system and app framework, some workflows no longer need to start with an external API call. They can start close to the user’s data, with lower latency and a clearer privacy story.
What Apple is opening up
Apple’s documentation describes the Foundation Models framework as access to models designed for Apple Intelligence, including on-device models and Private Cloud Compute. It also leaves room for other model providers that conform to the relevant Swift protocol.
Architecturally, that matters. A developer does not have to treat AI only as a remote chatbot. It can become a normal app capability: summarization, entity extraction, text and image understanding, suggestions, game dialog, or work with local context.
Three parts are especially practical:
- structured output through ’@Generable’, so model output can become Swift data structures,
- tool calling, where the model can call tools defined by the app,
- dynamic profiles, which let an app change instructions, tools, and model configuration during one session.
Where this fits in a company
The best use cases are not giant autonomous agents. They are small, frequent operations inside existing applications.
For example, a sales app on an iPad can summarize meeting notes, extract the company, contact person, budget, and next step, then send a structured draft to the CRM. A support app can classify a short report, suggest a category, and draft a reply. An invoice tool can prefill fields from a document while a human still reviews them.
These are not flashy demos. They are places where AI saves minutes in repeated work without every small task immediately leaving the device.
Why the on-device layer matters
Enterprise workflows often depend on details: where data flows, how fast the response is, how much repeated calls cost, and what happens when the user has poor connectivity.
A local model will not always be the best choice. Apple itself points to Private Cloud Compute or server providers when a task needs stronger reasoning or a larger context. But for the first automation layer, local AI is useful:
- fast classification and suggestions,
- more sensitive data that should not go to every external API,
- preprocessing before a more expensive model,
- user-facing features directly inside the app.
Good architecture does not mean “do everything on the device.” It means routing: simple tasks locally, harder tasks through private cloud or a selected server model, and critical decisions through human review.
What to watch
Foundation Models is not a replacement for an internal AI strategy. Apps still need permissions, auditability, evaluation, fallbacks, and quality checks. If AI prefills a CRM record, invoice, or customer response, the system needs to know when it is only a draft and when it changes real state.
I also would not expect a local model to replace top cloud models for difficult analysis. Its strength is different: being available, fast, built in, and good enough for small tasks that repeat all day.
How I would deploy it
I would start with one narrow workflow:
- lead triage in a CRM,
- extracting data from an email attachment,
- suggesting a support ticket category,
- summarizing a document before approval,
- checking a form before submission.
Then I would measure the basics: time saved, how often humans correct the output, when the workflow needs a stronger model, and which outputs must never be written without review.
That is healthier than trying to build an “AI app.” Apple is giving developers a way to add intelligence to existing apps in small, controlled steps.
Bottom line
Apple Foundation Models are not interesting because Apple suddenly beat every large model provider. They are interesting because AI capabilities are moving into the application layer where companies already work.
For CTOs and development teams, the question is simple: which small steps in our apps are we still sending to the cloud, or asking humans to do manually, even though a local model with proper review could handle them?
Sources: Apple Developer documentation for the Foundation Models framework, the Apple Intelligence developer overview, and Marek Bartoš’s video Apple omylem ukázal největší sílu AI.