Deep Dive — Siri AI app intents open to every app in iOS 27.3

For two years, the most common complaint about Apple's rebuilt assistant was narrower than "Siri is dumb." It was that Siri could only act inside apps that fit a template Apple had already imagined. If your app was a calendar, a messaging client, a to-do list or a workout tracker, you were in. If it was anything else, the assistant would politely decline to touch it.

On November 3, Apple removed that ceiling. In the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas, any app can expose its own actions and content to Siri AI through App Intents, and the assistant will call them as tools when a request needs them. Apple's own phrasing, in the developer announcement, is that the preview "enables any developer to integrate their app with Siri AI." There is no new entitlement, no separate developer program, and no application to join.

What a reader would type to find this: "Siri AI app intents", "Siri AI app skills", "iOS 27.3 beta Siri developers".

The timing is the interesting part. Customers get this "in a few months," App Store submissions open "in the spring," and Apple has not yet published the terms and usage policies that will govern what developers build. The platform is open before its rulebook exists.

Every app is a tool now

The old mechanism was schema domains: a fixed list of task types Apple defined in advance, each with expected parameters. A developer adopted the schema for "send a message" or "add a calendar event" and inherited Siri's understanding of it. The benefit was that Siri knew what the action meant. The cost was that anything Apple hadn't modeled simply did not exist — 9to5Mac described the first version of Siri AI as limited to "a much smaller set of supported schemas," which is the polite way of saying most developers hit a wall in week one.

The replacement is the pattern the rest of the industry already uses. A developer marks an intent with a Swift macro, @AppIntent(.tool), which makes it conform to a new ToolIntent protocol and exposes it to Siri as a callable tool. Intents that already match a Siri schema keep working; ones built on a Shortcuts-only schema can be promoted with a schema argument rather than rewritten. Apple's documentation spells out what the assistant can do with them: infer which app a request belongs to when that app is in the foreground, has an active Live Activity, or was mentioned earlier in the same Siri conversation; pull content out of one app, hand it to another and act there; and show results in card snippets built from the developer's own views. Siri can also surface any entity an app places in the on-device Spotlight semantic index, not only content shaped like one of Apple's templates — per Apple's Siri AI preview documentation.

Two details there matter more than they look. One is an entity type whose values the model can generate itself rather than look up, which is what lets "add a flight to Lisbon on Friday" create structured data instead of demanding a menu. The other is tooling: a Siri Integration Advisor in Xcode that reviews an implementation, and a Siri Evaluations framework that scripts multi-turn requests and asserts which intent Siri called. That second one is the admission that this is a probabilistic system. You do not test it by tapping buttons; you test it by running scenarios and watching the pass rate move.

The skills are markdown files, and that is the tell

Complex requests need more than one tool. For those, Apple added app skills: a folder inside a Skills.skillsbundle in the Xcode project, containing a file called SKILL.md. The frontmatter carries a name and a one-line description, capped at 1,024 characters; the body is free-form Markdown up to 10,000 characters, telling Siri how several tool intents fit together, which one to prefer when more than one applies and what to avoid.

If that sounds familiar, it should. Apple's documentation says app skills "use a subset of the Agent Skills format" — the name and description fields plus a Markdown body — and that Siri does not support the other frontmatter fields or supporting files like scripts and references that the full format allows. In other words, the company that spent a decade insisting integrations run through typed, compiled, Apple-defined schemas has just shipped prompt engineering as an app resource, in the same file format the agent tooling world converged on over the past eighteen months.

That is a bigger concession than the API list suggests. Schema domains were Apple saying: tell us what your app does in our vocabulary and we will let the assistant understand it. Skills are Apple saying: explain it in English and the model will figure it out. The quality of an integration now depends on how well a developer writes prose about their own product — which is exactly the kind of thing a large language model is good at reading and exactly the kind of thing that is impossible to review at App Store scale.

Apple is careful to frame skills as guidance rather than automation. They cannot enforce deterministic behavior, and Siri does not simply execute the listed intents in order — the skills documentation calls them information the model "may use" in a conversation. Developers who wanted a scripting language got a style guide instead. That is the right design for a model-driven assistant and a frustrating one for anyone who has ever shipped a checkout flow.

Why this is bigger than a developer API

Apple's pitch for the preview is distribution, and the announcement says it plainly: "People around the world are turning to Siri AI to get things done every day, which means more customers can discover and use apps just by asking." No figure accompanies the claim, because the honest figure does not exist yet.

But the logic is sound. If Siri becomes the surface where people ask for things, then the description field Siri reads — Apple says it uses the description to decide whether a skill applies "without reading the body" — becomes the new App Store listing. Search optimization moves from screenshots and keywords to a 1,024-character sentence about what your app does and when to use it. The ranking rules that Apple has not published will matter more than the API it just published.

This is the fourth expansion of the same door. SiriKit, in 2016, let apps handle a short list of intents — messaging, payments, ride booking, workouts. App Intents, from 2022, let apps declare their own actions and made them reachable from Shortcuts, Spotlight, widgets and the Action button. Schema domains, shipped with Siri AI this year, gave the assistant a shared vocabulary for common tasks. Tool intents make the vocabulary optional. Each step moved the gatekeeping from "may your category integrate at all" to "whose description wins when two apps could serve the same request," and Apple has kept the router through all four: the System Orchestrator decides, the App Toolbox executes, the Spotlight index supplies the personal context. Apple does not need to own the app. It needs to own the routing.

The comparison is instructive. OpenAI's apps platform asks developers to build inside ChatGPT's own surfaces; Google's agent tooling asks them to adopt a separate skills SDK. Apple's asks for something developers already wrote — App Intents, the same code behind their widgets, Shortcuts actions and Live Activities. Apps that already had those needed little new code to become Siri tools. The lowest-effort integration path usually wins, and Apple's is now the lowest.

The part Apple hasn't written yet

The announcement ends with a sentence that deserves more attention than the features above it: "we will follow up with more details about the terms and usage policies for experiences enabled through Siri AI."

Nothing in the November 3 post describes a payment mechanism, a revenue share, a review process for skills, or a rule about whether Apple's own apps get preference when Siri has to choose. Those are not implementation details. They are the product. A Siri that can call any app's actions is a marketplace with one employee — the router — and the router's rules decide who gets the customer.

Consider what a developer is being asked to believe. Their app's most valuable interaction surface becomes a sentence a model reads and ranks. That model's judgment is not inspectable, its failures are not appealable, and the company operating it sells competing apps in Maps, Music, Notes, Mail and Messages. Apple is already defending an antitrust case about how it controls access to the iPhone. A voice assistant that decides which app answers "book me a table" is the same question with a microphone.

The cost question is equally open. Siri AI runs partly on-device and partly in Private Cloud Compute, and Apple has been explicit that demand can outrun capacity — the assistant already carries daily usage limits. If a third-party skill triggers cloud reasoning on Apple's infrastructure, someone pays for those tokens. Apple has not said who.

The preview cannot reach users

The fencing around this developer preview is stricter than the announcement's tone suggests. Integrations work in development-signed builds on supported devices and for internal TestFlight testers. External TestFlight testers and App Store users do not get them, and Apple warns that both the API and Siri's behavior may change during the preview.

Read that as a shipping constraint. A developer can write and submit code today that will do nothing for customers until Apple turns the feature on. Apple's timeline — customers "in a few months," submissions "in the spring" — puts the first real integrations past the 2026 holiday season and into the spring release cycle. Anyone building now is betting engineering time on an interface declared unstable in advance, which is why the smart play is to expose existing intents as tools and write skills, not to redesign an app around Siri.

It is also the second time this year Apple has shipped its flagship assistant with a beta label attached. iOS 27 went out on September 14 with Siri AI marked as a beta, as we noted in Apple ships iOS 27 — its new Siri AI is still a beta. Seven weeks later the developer platform arrives in the same condition. The pattern is consistent enough to be a strategy: ship early, label it, and let the ecosystem absorb the risk of the rough edges.

Geography, and the Google-shaped hole

"Any app" has an asterisk the size of a continent. Siri AI is available in English, French, Japanese, Korean, Portuguese and Spanish, with Traditional Chinese, Norwegian, Turkish and Vietnamese promised. It is not available in the European Union on iOS, iPadOS or watchOS. So the platform that just became open to every developer is closed to every user in twenty-seven countries, and Apple's DMA compliance work — not its developer relations — will decide when that changes.

China is a separate path entirely: Apple Intelligence there runs through Alibaba's Qwen rather than the Apple Foundation Models that power Siri elsewhere, the arrangement we covered in Apple Intelligence in China — Qwen comes to Siri on November 4. A third-party skill written for one market will not automatically behave the same way in the other.

Then there is the dependency nobody at Apple put in the announcement. The Apple Foundation Models behind Siri AI were custom-built in collaboration with Google's Gemini models, and the most capable of them — AFM Cloud Pro, which Apple says handles agentic tool use and complex reasoning — runs in production on Nvidia GPUs inside Google's cloud. Craig Federighi said so at WWDC in June, in a session 9to5Mac transcribed. Requests still run through Private Cloud Compute and are not stored, so the privacy claim holds. But the reasoning layer that now routes every third-party app action crosses infrastructure owned by a competitor. That is a strange foundation for a platform pitch about owning your own stack.

What skeptics say

The most useful skepticism is not about whether the API works. It is about whether Apple can write good instructions for its own assistant. Federico Viticci, who has covered Apple's software longer than most, went through the skills Apple shipped for its own apps and found them thin: skills to teach Siri how to select and insert text in Notes, or to explain which file types the Files app supports — what he called "a lot of examples of what NOT to do (sign of an old model)."

That matters because Apple's own skills are the reference implementation. If the company's best effort produces instructions that describe plumbing rather than intent, the long tail of third-party skills will be worse, and Siri's routing accuracy — not its model quality — becomes the ceiling on the whole platform. The failure mode is not "Siri can't call your app." It is "Siri called the wrong app, confidently, and the user blames you."

The counterargument is that a developer preview only has to exercise the format, not win a benchmark. Both readings fit the documentation. The next six months decide which one is true.

What to watch

The terms, not the API. Apple has promised usage policies for Siri experiences. Whether intents can be monetized, whether Apple takes a cut, and whether there is any ranking disclosure will determine if this is a distribution channel or a feature. Watch for the policy before the spring submission window.

Whether App Store review gains intent-level rules. Skills are prose, and prose is hard to reject. If Apple starts reviewing the description fields — the text Siri reads to decide which app answers — that is the moment this becomes a regulated surface rather than an open one.

The first customer ship date. "A few months" is not a date. If Siri AI's third-party capabilities slip out of the 27.3 cycle, developers who built early will have spent a quarter on an unstable API, and the enthusiasm visible in developer forums this week will cool.

Adoption by the top apps. The number that matters in six months is not how many developers tried the preview, but how many of the hundred most-used App Store apps ship a skill. Discovery arguments only hold if the apps people already want are the ones Siri can reach.

For developers, the practical read is unglamorous: expose the intents you already have, write the description like ad copy, write skills like onboarding docs, and put the Siri Evaluations framework in your test suite before you ship. The platform is real. The rules are not written, the API is allowed to change, and users cannot see any of it yet.

Is a Siri that picks which app answers your request a feature or a gatekeeper? Write to us via our contact page.