Apple opens Siri AI App Intents preview to all developers
Apple opened its Siri AI developer preview to all developers on November 3, alongside the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas. Apps can expose custom actions through App Intents. Apple says these new integrations will reach customers in a few months, with App Store submissions planned for spring.
For an app team, the immediate question is which useful tasks to make available and how to test them. A good integration needs a reliable connection between a spoken request, the right app data and an action the user actually intended.
Who can test the Siri AI preview now?
Apple’s preview documentation describes two testing routes. Neither makes the new integration available to everyone who downloads your app.
| Distribution route | Preview access |
|---|---|
| Xcode development build | Supported with development signing on a compatible device |
| Internal TestFlight testers | Supported |
| External TestFlight testers | Not supported |
| App Store download | New preview functionality is unavailable |
Developers do not need a separate program or an additional entitlement. Apple Intelligence hardware is required, and Apple warns that APIs and Siri behavior may change during the preview. Siri can use only the custom content and actions an app explicitly exposes.
Context also affects discovery. The app need not be named when it is in the foreground, has an active Live Activity, or remains part of a Siri conversation that mentioned it. For background requests, people identify the app by name. Apple illustrates this with checking a saved product or changing a ride pickup location.
App Intents become tools with explicit boundaries
The implementation guide introduces the @AppIntent(.tool) annotation for eligible custom intents. Titles, descriptions and parameter descriptions tell Siri what an action does and what information it needs. Existing intents in schemas that already support Siri should not receive the extra tool annotation.
Apple’s shopping example exposes an action that adds a product and quantity to a cart. That is a useful distinction for developers designing their own integrations: an action should have a clear purpose and a result that can be checked. Cart editing and purchase approval deserve separate treatment in a test plan.
For content, Apple describes Spotlight indexing, contextual associations with visible views, and transferable representations for data passed between apps. Developers can apply authentication policies to entities or individual properties and keep selected properties internal. If an intent returns sensitive information, Apple says authentication must be set at the intent level. Property policies alone apply to Siri’s responses.
App skills add guidance, not a fixed script
Apple’s app skills guide addresses requests that need several actions. A skill can explain how tools fit together, which action to prefer, what to avoid and when to clarify a request. Apple’s example covers a shopping flow from finding products through reviewing a cart and confirming a purchase.
These instructions live in Markdown files inside an Xcode skills bundle. Siri first reads a skill’s metadata to decide whether it is relevant, then reads the body when it selects that skill. Apple supports a subset of the Agent Skills format, with a name, description and Markdown body. Supporting scripts and reference files are not included in that format.
The important limit is behavioral. Apple explicitly says skills cannot enforce deterministic execution, and Siri does not simply run every listed tool in order. Developers should therefore treat skill wording as guidance, with consequential checks implemented in the app itself. That is our engineering interpretation of the documented limit, not a guarantee about Siri’s reliability.
Test outcomes before promising an experience
The preview includes a Siri Integration Advisor in Xcode and a Siri Evaluations framework for scripted requests and response checks. Apple also recommends trying different request wording when validating skills on a device or through internal TestFlight.
A useful test set should cover more than the happy path. Can Siri select the intended item when names are similar? Does a request with missing details trigger an appropriate question? Does a failed action leave the app in a clear state? These are suggested evaluation cases, not results from ByteForward testing.
Apple lists six supported languages initially, with regional limits. Siri AI is unavailable in the EU on iOS, iPadOS and watchOS, and is unavailable to children under 13. Some server features have daily limits. Additional terms and usage policies are still forthcoming. Those conditions belong in launch planning alongside the code.
For related coverage, read about Apple’s Siri AI home device rollout. For now, the useful deliverable is a tested app integration, with public availability communicated only when its distribution path is ready.
Featured image: AI generated editorial illustration by ByteForward, showing conceptual app actions rather than an actual Siri interface. Apple logo reference, original design by Rob Janoff. Apple, the Apple logo and Siri are trademarks of Apple Inc. ByteForward is an independent publication and is not sponsored or approved by Apple.
