Apple’s ‘New’ Siri AI Was Exciting 2 Years Ago. Not Anymore

On the homepage of Apple Intelligence, Apple calls it “AI for the rest of us”. 

Instead, the ‘us’ inadvertently gestures towards a user base that spent years frustrated by a stagnant assistant—and the ones who even sued Apple for the delay, a lawsuit the company eventually settled. 

But at this year’s developer conference, Apple finally unveiled a completely overhauled Siri, earning early praise in iOS betas.

While the next iPhone will cost significantly more due to ongoing memory issues, the question is whether that price will at least come with something worth paying for. 

With Apple Intelligence and the new Siri, rebranded as Siri AI, Apple is promising a meaningful change in how one billion iPhone users experience the ecosystem.

Today, Siri AI operates at the operating system level, allowing it to access your Messages, Mail, Photos, and on-screen content in real time without switching apps. 

It can now execute tasks contextually. Pointing the camera at a plate of food yields a nutritional breakdown, while a single prompt can split a bill via Apple Cash

In the Passwords app, Apple Intelligence can automatically go to individual websites to change and fix insecure passwords on your behalf. 

Powering all of this is a dual-model architecture: Apple’s own Foundation Models (AFM), running on-device, paired with Google’s Gemini for more demanding tasks, with the AFM Cloud Pro model, built in collaboration with Google, running on NVIDIA GPUs in Google’s cloud for the heaviest workloads. 

But the question worth asking is whether Apple just delivered on what it had already promised or built something genuinely new. The two are not the same. 

Siri AI Was Exciting 2 Years Ago, But Not Today

Jesai Tarun, an iOS developer and Founder of Sentient, tells AIM, “I think Siri is exactly what Apple promised. Nothing more, nothing less.”

“What they promised was impressive two years ago. It’s no longer impressive today,” he adds. 

He notes that while Siri AI is limited to basic, app-specific tool calls, a truly exciting shift would be agentic AI, like Gemini on Pixel devices, which can autonomously navigate the entire Android OS by interacting directly with the screen.

For such agentic capabilities, Apple doesn’t have such powerful models yet. Apple Photos, for instance, illustrates the shallow limits of the tool-call approach, serving as a prime example of how AI is restricted to only a fraction of an app’s features. 

System prompts from the ‘Apple Intelligence Report’ reveal that the voice-editing demo shown at WWDC—such as adjusting brightness and rotation—accounts for virtually all of Siri’s capabilities. 

Because a model faces a hard ceiling on how many tool calls it can process at once, developers can only expose a handful of basic commands, leaving the vast majority of deeper app functions completely out of reach.

This same constraint shapes Siri’s on-device architecture. While the local model handles simple tasks like summaries or basic actions, any query requiring meaningful reasoning is offloaded to Apple’s Private Cloud Compute (PCC) servers. 

If that server-side model needs personal context, such as a text message about a shopping list, it requests the data from the iPhone. 

The local device then runs a retrieval search, extracts the text snippet, and sends it back to the cloud. Ultimately, the on-device model isn’t doing the reasoning but functions purely as a retrieval engine.

While Apple Intelligence uses on-device models for basic writing tools, advanced local processing is restricted to high-tier hardware. 

On 12GB RAM devices like the iPhone 17 Pro, Apple deploys a multimodal LLM specifically to improve dictation and speech-to-text, allowing it to process audio directly and infer meaning from unclear speech. Beyond this, however, its local capabilities remain thin.

Nowhere Close to Gemini on the Pixel Series

The deeper bottleneck is Apple’s dependency on Google for foundational model capabilities. 

Lacking a web crawler like Google’s or an automated synthetic data loop like Anthropic’s, Apple pays Google an estimated $1 billion annually to use Gemini as a ‘teacher’ model. 

Apple distils Gemini’s outputs to train its own leaner, proprietary models, which run securely within Apple’s Private Cloud Compute infrastructure. 

Yet, because the ceiling of Apple’s AI is bound to the specific Gemini version it last distilled, Apple remains trapped in a cycle of retraining whenever Google upgrades its models.

Since March, Gemini on the Pixel 10 has been able to autonomously navigate app interfaces to complete multi-step tasks like ordering food or booking a ride. 

It operates on two layers: AppFunctions, which lets developers explicitly expose app actions, and a UI automation fallback, allowing Gemini to read accessibility trees and click through apps that have zero formal integration work.

This raises an obvious question: since both Apple and Google rely on Gemini-derived models, why can Android do this while Siri cannot?

The bottleneck isn’t the model but the operating system. 

Apple’s App Intents framework requires developers to explicitly pre-declare every single action Siri is permitted to take. Without an intent, Siri has no fallback. 

Furthermore, iOS’s strict sandboxing architecture prevents apps from accessing data or processes outside their own container, making Android-style screen automation structurally impossible unless Apple opens up the OS layer. 

Google, on the other hand, allows Gemini on the Pixel 10 to actively parse accessibility trees and overlay controls directly over third-party apps. 

Apple is likely to avoid doing this because granting an AI agent systemic permission to ‘read’ and control everything on a user’s screen creates massive security risks, like data scraping or accidental authorisation, that fundamentally clash with iOS’s core privacy ethos,

Siri can only trigger pre-defined functions. It cannot read a screen and start tapping, and no amount of model training closes that architectural gap.

It is also worth noting that while both ecosystems point to Gemini, they are not utilising it in the same way. 

Google has direct, native access to the most high-end, unconstrained versions of its frontier models. Apple, by contrast, does not run Gemini out of the box. Instead, it uses a process called model distillation—essentially using the massive reasoning outputs of Google’s raw models as a ‘teacher’ to train Apple’s own smaller, highly specialised Apple Foundation Models (AFM). 

While this allows Apple’s models to run natively and privately on Apple Silicon, a distilled student model structurally lacks the deep, raw emergent reasoning capabilities of the massive frontier model it was cloned from. 

Enterprise Agents to Follow Siri AI’s Architecture? 

Having said that, experts across the ecosystem have continued to praise Apple’s orchestration process, despite the limitations of features. 

Max Weinbach, an analyst from Creative Strategies, shared in a post on X, “I can near guarantee you enterprise agents will follow Apple’s Siri architecture. It’s the perfect mix of on-device and cloud.” 

Based on official documentation, Apple Intelligence prioritises privacy by defaulting to on-device processing. When a user makes a request, the local hardware handles the initial processing to determine whether the task can be completed locally, keeping personal context entirely on the device. 

For complex reasoning tasks that exceed local capabilities, the device routes a structured request to Apple’s PCC.

To protect this data in transit, your device scrambles the payload so it can only be unlocked by a certified Apple server. 

Since the cloud model must temporarily read the data to fulfil your request, true end-to-end encryption is not possible. Instead, Apple relies on a highly restricted server design that forces the system to handle the data securely without ever saving it.

PCC is designed for stateless computation: the personal data contained in the request is used exclusively to fulfil that specific prompt and is completely erased after the response is returned, preventing any data retention, logging, or unauthorised access.

Why Siri AI Won’t Change Your Life

While Apple has provided a framework called App Intents to bridge the gap between Siri and third-party apps, the system places an immense burden entirely on the developer. 

In practice, Apple possesses no native ability to integrate with external applications on its own. Instead, developers are forced to do an exhaustive amount of manual heavy lifting to integrate with the Siri AI index.

According to Tarun, making App Intents function properly requires an overwhelming amount of backend work. 

If an app contains dozens of features, a developer cannot simply let the AI figure them out; they must sit down and explicitly expose and map every single function for Siri to use. 

“I have to re-architect my back end, spend so much time thinking about what feature I should expose to Siri, and every time I make some changes to the app, I have to then make sure to do that again,” he says. 

“It is basically a crazy amount of work.”

Besides, is it in app developers’ best interests to integrate their features to work with Siri AI? 

For one, we have the big tech apps. The return on investment for someone like Meta in integrating WhatsApp with Siri is basically about improving Siri. Would the company want to make Siri a better assistant so that users would prefer using it over Meta’s AI in WhatsApp? 

“Apple is not gonna be able to integrate Siri AI heavily with Gmail and Google. It’s in Google’s best interest not to make Siri AI fancy, because they have Gemini inside Gmail,” says Tarun. 

This dynamic was already proven with Apple’s Vision Pro, which lacked native Netflix or YouTube apps at launch because competitors had no incentive to enrich Apple’s ecosystem while building competing hardware platforms of their own.

Besides, even if app developers without competing AI products may still prefer users to spend time inside their apps rather than automate actions using Siri, since automation reduces engagement. “It’s in everyone’s best interest to just get you into their app to see advertisements,” says Tarun.

However, this defensive stance from developers may ultimately be forced to shift as market dynamics evolve.

Avi Greengart, Founder and Lead Analyst at Techsponential, tells AIM that while many apps rely on advertising, that business model is already under pressure from Apple’s privacy initiatives, which allow users to prevent apps from tracking their usage data. 

Developers will inevitably have to find other ways to generate income, including charging directly for their apps and services.

Greengart notes that ecosystem stakeholders still have a brief window to adapt because Siri’s current agentic capabilities are restricted, and the broader consumer market has yet to fully upgrade its hardware. 

“They have some time; Siri AI has limited agentic capabilities today, and most of the iOS base doesn’t support Apple’s best on-device models today,” he explains.

While he expects several holdouts to resist, clinging to the old app-centric paradigm, he argues that economic reality will eventually force their hand. 

Put simply, if a user asks Siri to book a ride or order groceries, Siri will naturally route that business to the apps it can actually communicate with, leaving non-integrated competitors effectively invisible.

“Once Siri pushes more business to an app or service that supports the Intents API over alternatives that it can’t access, that will force competitors to support agentic APIs as well or lose out on both direct and indirect engagement,” Greengart warns.

“In my tests of the Apple Intelligence beta, Siri will soon do most of what today’s mainstream consumers actually want to do with AI,” says Greengart. 

Will Siri Pose a Threat to ChatGPT? 

Does this mean Siri will pose a threat to the nearly 68 million daily active users on ChatGPT who are on the iOS platform? 

For one, Greengart says that users who are already paying for an AI service on Google, OpenAI or Anthropic may not replace it with Siri for any task. Besides, the broader ecosystem comes into effect. 

“Even if you have an iPhone, if you’re a heavy Gmail, Workspace, and Maps user, it may still make sense to stick with Google directly for AI interactions.” 

Tarun believes Apple’s window to position Siri as a legitimate threat to OpenAI closed the moment it failed to launch immediately after ChatGPT’s debut. By now, consumer habits have ossified, and ‘ChatGPT’ has become completely synonymous with AI queries. Breaking that ingrained behaviour and persuading users to proactively choose Siri is a massive uphill battle. 

Google faced the exact same headwind. Despite a rapid rollout, standalone Gemini adoption lagged behind OpenAI. 

Its true scale today relies on ‘invisible’ distribution, embedded silently inside Gmail, Docs, and Search Overviews.

“On average, everyone will just continue using ChatGPT,” says Tarun.