For the past few years the AI conversation has orbited almost entirely around models — which one scored highest on benchmarks, which company released the best reasoning engine, which chatbot felt most natural in a side-by-side demo. It’s understandable. Models are visible, comparable, and easy to narrate as a race with a clear leaderboard.
Apple, by that metric, looks like it’s losing. Siri remains the punchline of a reliable joke. The company arrived late to generative AI and shipped much of what it announced at WWDC 2024 somewhere between twelve and eighteen months behind schedule.
But the model race is the wrong race to watch. And Apple might be running the right one.
The Platform Play vs. the Content Play
What Apple has been quietly assembling isn’t a better chatbot. It’s an orchestration layer — the plumbing that any AI model has to pass through to do anything useful on an iPhone. Models generate text, but acting on the world requires connecting that text to structured, executable capabilities inside real applications. That’s the layer Apple is building, and it’s a fundamentally different kind of bet.
The architecture has four pieces working together. At the user surface, Siri becomes ambient — not a destination app but a thin layer sitting between you and everything else on the device. At the developer surface, App Intents (Apple’s framework for letting apps expose structured actions to the operating system) and MCP — the Model Context Protocol, an open standard that lets AI models communicate with software systems — provide the API surface. On-device foundation models handle personal context and routing. Private Cloud Compute handles the heavier reasoning, with a privacy architecture designed so personal data never leaves the device for routing decisions.
The detail that most coverage missed: developer betas for iOS 26.1 contained code showing Apple building MCP support directly into the App Intents framework. That single technical decision changes how third-party AI models interact with Apple devices. Instead of each model building its own custom integration, Apple handles protocol compatibility at the OS level. Any model that speaks MCP — ChatGPT, Claude, open-source local models — can act on apps through Apple’s layer without a separate deal. The toll booth moved from “you need our App Store to distribute” to “you need our intent system to be actionable.”
The Question Worth Asking
The reframe that’s stayed with me is this: “Do we have an app?” is no longer the right question. “Can an agent use our app?” is.
This is already concrete. Apple has expanded App Intent domains across browsers, mail, photos, file management, word processors, spreadsheets, and more. When a user asks an agent to apply a filter to yesterday’s photo of a specific person, the agent talks directly to a structured intent — no UI navigation, no pixel-matching, no brittle screen-scraping. Apps that don’t expose intents simply don’t participate in that interaction.
The parallel to how we think about AI design is more direct than it might seem at first. Specification engineering — writing precise, structured specifications so an AI can act without ambiguity — is exactly what App Intents asks developers to do for their own software. The clarity of intent that makes a prompt reliable is the same clarity that makes an application agent-addressable. The skill transfers in both directions.
Why “Late but Integrated” Has a Track Record
Apple has run this playbook before. Late to smartphones. Late to tablets. Late to ARM processors. Won on integration depth every time. The pattern: arrive after the first-mover scramble settles, then build something woven so deeply into the surrounding system that switching becomes expensive.
What’s structurally different this time is the flywheel. Every developer who adopts App Intents makes Apple’s agent layer more useful. Every agent capability added through MCP makes App Intents more valuable to developers. That dynamic compounds, which is why the comparison to SEO matters: SEO gave you years to adapt as search rankings shifted. The agentic transition may compress that window considerably.
There’s also an asymmetry in the Gemini deal worth noting. Google provides cloud reasoning; Apple maintains control of the UX, the data pipeline, and the privacy architecture. Google can’t see the personal data flowing through Private Cloud Compute, but aggregate query patterns — what kinds of tasks people delegate, how complex they are, where the failures happen — are valuable intelligence for improving Google’s own agent products. Apple has already described this arrangement as a bridge while on-device capabilities mature, much as the Intel chip relationship was a bridge before Apple Silicon. The exit is already planned.
What to Watch
The meaningful question isn’t whether Siri can beat any given chatbot in a live demo. It’s whether the App Intents flywheel reaches escape velocity — whether enough developers expose structured intents before the agentic transition becomes the default interaction mode.
For anyone building software today, the architectural choice between structured intents and UI automation (where a model tries to navigate an app by recognizing screen elements visually, like reading a map instead of using coordinates) maps to a real near-term decision. Structured intents are more reliable but require developer adoption. UI automation works today but breaks whenever an interface changes. The tradeoff isn’t abstract.
The most durable position in the agent era probably accrues to whoever controls the layer between the model and the action — not to whoever built the most impressive model. That layer is being built right now, mostly by the company everyone’s writing off.
Text summarized and optimized using Anthropic’s models and reviewed by a human.