Siri could take another step as early as 2027: to follow objectives and pursue tasks without waiting for an instruction at each operation. This perspective, reported by Mark Gurman at Bloomberg, remains a calendar hypothesis. Apple has already announced tools to act in applications, but not yet a stand-alone version of its assistant with a release date.
Since Siri AI’s presentation in June and its deployment in September, the issue has changed. It is no longer just a matter of making the assistant’s answers more natural. The challenge is to give him a task, let him mobilize several applications and find a usable result. Between these steps, however, we must distinguish between the functions available, the tools reserved for developers and the projects mentioned by the press.
2027, a possibility raised by Gurman
In its letter Power On of 8 November, whose information isreported by MacRumorsMark Gurman describes Apple’s willingness to make Siri more capable of pursuing goals. The reporter believes that developments in this direction could take place next year. He also reported on a recent meeting at which John Ternus had asked Craig Federighi, software engineer, to accelerate developments beyond the improvements already planned for Siri.
The status of John Ternus is established. Apple aannounced his appointmentwith effect from 1 September 2026. Tim Cook became Executive Chairman of the Board of Directors. On the other hand, the content of the meeting and the horizon of a more autonomous Siri are part of the journalistic narrative. They are neither a product sheet nor a commercial undertaking of Apple.
What a real agent would change
A classic assistant intervenes mainly on request: he answers, opens an application or executes an order. An agentic system has a wider margin of manoeuvre to achieve the desired result. He may choose an action, examine what it has produced, and then decide on the next. Apple already describes this principle of looping shares in itsdocumentation on agent functions.
The additional difficulty appears when the task lasts. A one-time request can be completed in seconds. An objective sometimes requires keeping its state, waiting for information, and then resuming. This continuity must also have an end: the user must be able to know what remains to be done and stop the work. These are criteria for judging an agent, not functions that Apple would have promised for a future edition of Siri.
The word « autonomous » deserves careful handling. It can designate a sequence of operations performed without detailed guidance, while maintaining human validation for important steps. It does not necessarily mean permanent surveillance, freedom of purchase or unlimited access to the information of the telephone owner.
Siri already knows how to connect multiple applications
Apple does not leave a timer-limited assistant. In itsseptember deployment presentation, the manufacturer describes Siri AI as able to search for information in messages, emails and photos, and then act in applications. One of his examples combines a suggestion received in Messages, a recipe found in an email and adding ingredients to a shopping list in Reminders.
Understanding the displayed content is another element. An application may relate to what the user looks at rather than to information that he or she must copy. Apple also presents a dedicated Siri app, with a history of synchronised conversations between devices. These capabilities facilitate exchanges and demand-related actions. They do not alone demonstrate that the assistant can manage a sustainable goal without further solicitation.
A formal step was announced on 3 November
The most recent verifiable novelty concerns developers. TheNovember 3, Apple opened a preversionallowing all applications to offer their actions and content to Siri AI in the betas d的iOS 27.3, diPadOS 27.3, macOS 27.3 and visionOS 27.3. This opening is based on App Intents, the mechanism that describes the operations that an application can perform.
Apple also introduces « skills », designed to teach Siri how to combine multiple actions to deal with a complex demand. The manufacturer cites, in particular, the following of a delivery, reservation or order as possibilities offered by the integrations developed by the publishers. It should not be inferred that each application already has these functions.
The published schedule is progressive: first deliveries to users in a few months, then opening submissions to the App Store in the spring. It is a formal program for application integration. This program is not valid for confirmation of a personal agent able to handle all uses in 2027.
How a sentence becomes an action
The operation is based on a useful separation. The language model interprets the application. The application shall set out actions and the data necessary for their execution. Theapp Intents documentationasks developers to represent these elements in a structured form, so that the system can discover and use them.
Adding an appointment means more than producing a convincing sentence: you have to identify a calendar, a date, a duration and, possibly, participants. If information is missing or if two persons bear the same name, a supplementary question may be required. Quality therefore depends as much on the correct identification of the data as on the fluidity of the conversation.
This approach also allows us to understand a limit: the ability to talk about an application does not prove that Siri knows how to do it. The publisher must provide suitable integration and usable information. An assistant can very well explain an approach while being unable to carry it out in the department concerned.
Confidentiality is not just about iPhone
Apple distinguishes several levels of processing. Itspresentation of Siri AI architecturecombines models executed on the device with models used on servers via Private Cloud Compute. The system also relies on Spotlight index and a local action toolbox. So it would be incorrect to sum it up to an artificial intelligence that never leaves the phone.
In atechnical publication of 8 juneApple described the extension of Private Cloud Compute to Google Cloud, with components from Google and NVIDIA. The company claims to maintain its guarantees, including the non-preservation processing of requests, the lack of privileged access to data during execution and the possibility of inspection by researchers. It specifies to keep control of the software accepted by its devices.
These commitments concern the protection of calculation. They do not answer all the questions that an action poses in a foreign service. If an operation leads to the sending of a message or an order, information must be sent to the recipient. For the user, it is therefore necessary to distinguish between the confidentiality of the processing, the access granted to the assistant and the transmission necessary for the task.
Confirmations still needed
Apple devotes atechnical session with confirmations and shared content. In particular, developers should indicate whether an item belongs only to the user, whether it is shared with other people or whether it is public. This information helps Siri decide when to request validation before acting.
The example of the calendar sheds light on the problem. Changing a personal note and changing an appointment involving several people do not have the same consequences. An action may seem trivial in the interface while sending a notification or changing the organization of others. A useful autonomy must integrate these differences, rather than remove all stages of confirmation.
Reliability is also measured after action. An announced result is not enough if the operation failed, if a service did not respond or if a data has become obsolete. For tasks in several stages, the monitoring of failures and the possibility of going back is as important as the success of a demonstration.
The risk of hidden instructions
Apple’s security documentation highlights another risk: indirect injection of instructions. Content consulted by the assistant, such as a calendar event or external text, may contain phrases designed to divert his/her behaviour. When the system can act, the error may exceed a bad response and lead to an unwanted operation.
Apple presents the resolution of this problem as an active research area. The protections described include the reduction of model data, confirmation of sensitive actions and the requirement of a unlocked device for certain operations. Server confidentiality and resistance to these manipulations are two distinct topics. Protecting the place where the template runs does not guarantee that it will correctly interpret each document.
Concrete limits for Francophone users
Siri AI isavailable in french since 26 octoberon compatible devices and in supported regions. Apple states that the service remains unavailable in the European Union on iOS, iPadOS and watchOS. This restriction should not be applied indiscriminately to the Mac and Vision Pro. Arabic is not among the six languages of Siri AI announced on that date.
Apple also indicates that functions using its server models are subject to varying daily limits. The material, language, country and application used therefore remain decisive. Ourguide to iOS 27.2details the practical settings of the iPhone; the prospect of a future agent does not change the current access conditions.
To measure Siri’s progress, the next test will be concrete: what actions will the applications actually propose, with what validations and what monitoring of results? The ambition reported for 2027 gives direction. The tools available and their behaviour in everyday usage will enable them to measure their scope.
Comments