Apple tried to build a smarter Siri on its own. It spent two years developing in-house models, announced the upgrade at WWDC 2024, ran ads for capabilities that did not yet exist, and then had to publicly admit it needed more time. The in-house models were not good enough.
So Apple did what Apple almost never does — it went to a competitor.
iOS 26.4, shipping to all compatible devices this month, delivers a fundamentally reimagined Siri powered by Google's Gemini. Not a partnership announcement. Not a feature preview. A shipping product that lands on every iPhone 15 Pro and newer in the coming days, transforming the most widely deployed voice assistant on the planet.
This is not a minor upgrade. Siri can now see your screen and act on what it sees. It can chain up to 10 sequential actions from a single natural language request. It maintains context across months of emails, messages, and calendar events using a 1-million-token context window. And it does all of this through a 1.2-trillion-parameter model running on Apple's Private Cloud Compute servers — Google's brain, Apple's privacy architecture, your device's interface.
When the company that built its entire brand on doing everything in-house hands the most intimate piece of its operating system — the assistant that reads your emails, sees your screen, and manages your calendar — to a competitor's AI, the architecture lesson is clear. No single company, not even Apple, can build the best AI for every task. The model-agnostic era just reached 2.2 billion devices.
What Siri Can Actually Do Now
The iOS 26.4 Siri is not an incremental improvement. It is a category change.
On-screen awareness means Siri can see and interpret what is currently displayed on your device. If a restaurant is shown in Safari, Siri can make a reservation without you copying the name or address. If a flight confirmation email is open, Siri can add it to your calendar and set departure reminders automatically. If you are looking at a photo, you can say “send this to Sarah” and Siri identifies the content, finds the right contact, and sends it through the appropriate messaging app — without you touching the screen.
Multi-step task chains allow Siri to execute up to 10 sequential actions from a single request. “Book me on the next available flight to New York, add it to my calendar, and text Sarah my arrival time” executes as a single workflow rather than requiring three separate commands with multiple confirmation dialogs. This is the same kind of agentic execution that Copilot Cowork delivers in Microsoft 365 — but running on a mobile device in your pocket.
Personal context means Siri now maintains a massive working memory of your interactions — emails, text messages, calendar events — across months of activity. The 1-million-token context window allows Siri to recall and synthesise information with a level of continuity that was previously impossible for a mobile assistant. Ask Siri about your dinner plans and it pulls up the restaurant reservation you made via text last week, cross-referenced with your calendar availability.
These capabilities transform Siri from a command processor into an agent. It does not just respond to instructions — it understands context, reasons about your intent, and executes multi-step workflows autonomously. This is agentic AI, shipping to billions of devices this month.
The Architecture That Made It Possible
The technical architecture behind the new Siri is what matters most for enterprise leaders and technology strategists.
Apple calls the system Apple Foundation Models version 10. It uses a custom implementation of Google's Gemini operating at approximately 1.2 trillion parameters. But the architecture is not simply “Siri powered by Gemini.” It is a hybrid execution model with three distinct tiers.
Simple tasks — setting timers, playing music, basic device controls — execute locally on the device using Apple's own smaller models, running on the Neural Engine in Apple Silicon. No data leaves the phone. Latency is minimal.
Moderately complex tasks — summarising emails, drafting responses, searching personal data — execute on Apple's Private Cloud Compute servers. These are Apple Silicon servers running in a stateless environment where data is never stored and is inaccessible even to Apple's own engineers. The processing happens on Apple infrastructure, under Apple's privacy controls, using Apple's security architecture.
The most demanding tasks — complex reasoning, multi-step planning, on-screen visual interpretation, natural language understanding across long context windows — are handled by the Gemini reasoning layer. But critically, this processing also runs through Apple's Private Cloud Compute with a privacy buffer layer before any data reaches Google's model. Apple controls the interface, the data routing, and the privacy enforcement. Google provides the reasoning capability.
This three-tier architecture is model-agnostic in practice even if it currently routes to a single external provider. The orchestration layer — deciding which tasks run locally, which run on Apple's cloud, and which require frontier reasoning — is the same architectural pattern that defines enterprise AI deployment. Route each task to the infrastructure best suited for it. Apply governance consistently regardless of where processing happens. Optimise for privacy, performance, and cost simultaneously.
Our Minnato orchestration platform operates on precisely this principle. Different tasks route to different models and different infrastructure based on what each task requires. Sensitive operations stay on-premises. Complex reasoning goes to frontier models. Routine operations run on cost-efficient local models. The orchestration layer makes the decision automatically, applies governance consistently, and captures value from every model improvement.
Why Apple Chose a Competitor
The backstory matters because it validates a principle every enterprise should internalise.
Apple spent two years trying to build Siri's reasoning engine in-house. Internal testing revealed the models were not performing at the level required. Apple then evaluated its options. It already had a partnership with OpenAI for ChatGPT integration in Siri, but that served a different function — handling complex world-knowledge queries rather than powering the core assistant.
By August 2025, Apple revisited Google's Gemini and found the technology had improved dramatically. Google also offered favourable financial terms. A September 2025 court ruling preserving Apple and Google's existing $20 billion annual search deal made the partnership less risky commercially. The companies finalised the agreement in November.
Apple reportedly invests approximately $1 billion annually for access to Gemini. The implementation is white-labelled — no Google branding is visible to users. From the user's perspective, this is still Siri. It is just dramatically more capable.
This is the same pattern playing out across every major technology company. Microsoft built Copilot Cowork on Anthropic's Claude rather than relying solely on its $13 billion OpenAI investment. Nvidia made NemoClaw hardware-agnostic rather than Nvidia-exclusive. Google embedded Gemini into its own Workspace while supporting MCP for external agent connectivity. And now Apple has handed Siri's reasoning to Google rather than insisting on in-house models that could not compete.
The lesson is universal: no single company, regardless of resources, can build the best AI for every function. The companies that thrive are the ones that select the best model for each task and wrap it in their own governance, privacy, and user experience framework. That is model-agnostic architecture. And it is now the operating principle of every major technology company on Earth.
What This Means for Enterprise Voice AI
The new Siri sets a consumer expectation that directly impacts enterprise voice AI.
When 2.2 billion device owners experience a voice assistant that sees their screen, chains multi-step actions, remembers months of context, and executes tasks autonomously, their expectations for every voice interaction shift permanently. Customer support calls, internal communication systems, enterprise assistants — every voice-based enterprise system will be measured against the new Siri's capabilities.
This creates both pressure and opportunity for enterprises.
The pressure is straightforward: customers who interact with Gemini-powered Siri every day will not accept clunky IVR systems, rigid menu trees, or voice bots that cannot maintain context across a conversation. The gap between consumer voice AI and enterprise voice AI, which was already narrowing, effectively collapses with iOS 26.4.
The opportunity is equally straightforward: the same architectural principles that power the new Siri — multi-tier execution, context-aware reasoning, agentic task completion, privacy-preserving processing — are available for enterprise deployment today.
Our Dewply voice AI platform delivers these capabilities in an enterprise context. Voice interactions are processed by AI agents that understand context, adapt to sentiment, remember conversation history, and execute multi-step tasks — booking appointments, escalating issues, retrieving account information, processing requests — without requiring customers to navigate rigid menus or repeat information. The same agentic voice capabilities that Apple is shipping to consumers, Dewply delivers to enterprise customer interactions.
The Arabic-language capability is particularly relevant for Gulf enterprises. While Apple's Siri has historically struggled with Arabic language processing, our Dewply platform processes Arabic voice interactions natively — understanding dialect variations, cultural context, and code-switching between Arabic and English that is common in Gulf business communication. As consumer expectations rise with the new Siri, enterprises that deploy voice AI matching those expectations — in the languages their customers actually speak — capture a meaningful competitive advantage.
What This Means for Enterprise Document Processing
The on-screen awareness capability in the new Siri points directly to where enterprise document processing is heading.
Siri can now look at a screen showing a flight confirmation and automatically extract the relevant information — dates, times, confirmation numbers, airline details — and act on it. This is visual document understanding applied to consumer use cases.
Enterprise document processing operates at a fundamentally larger scale. Invoices, contracts, regulatory filings, purchase orders, insurance claims, shipping documents — the volume of documents that enterprises process daily dwarfs what any individual Siri user encounters. But the underlying capability is the same: AI that can see a document, understand its structure, extract relevant information, and take action.
Our Vult document intelligence platform applies this at enterprise scale. Documents in any format — scanned PDFs, photographed invoices, handwritten forms, multilingual contracts — are processed by AI agents that extract, validate, route, and act on the information they contain. The same visual understanding that lets Siri read your flight confirmation lets Vult process thousands of invoices, contracts, and regulatory filings per day — with confidence scoring, human review triggers for ambiguous entries, and full audit trails for compliance.
The convergence between consumer AI capabilities and enterprise requirements means the technology gap has effectively closed. What remains is the governance gap — enterprise-grade security, compliance, audit trails, and human-in-the-loop controls that consumer products do not provide. That governance layer is where enterprise value is created and protected.
The March Pattern Is Complete
Stand back and look at what happened in March 2026.
Week one: Google embedded Gemini into every document, spreadsheet, and presentation through Workspace. Microsoft built Copilot Cowork on Anthropic's Claude and announced Agent 365 as the enterprise agent orchestration platform.
Week two: Nvidia revealed a $1 trillion AI infrastructure roadmap, launched NemoClaw as the enterprise agent operating system, unveiled the Groq 3 inference chip, and announced DGX Spark desktop AI factories. The Nvidia State of AI survey showed 88% of enterprises reporting AI-driven revenue gains.
Week three: Apple ships a Gemini-powered Siri to 2.2 billion devices, putting agentic AI capabilities in every consumer's pocket.
In three weeks, AI moved from enterprise infrastructure to enterprise platforms to consumer devices. The full stack is now in place: the chips (Groq 3 LPU, Vera Rubin), the platforms (NemoClaw, Agent 365, Minnato), the models (Gemini, Claude, GPT-5.4, Nemotron, Qwen 3.5), and the consumer surface (Siri, Copilot).
Every layer is multi-model. Every layer is model-agnostic. Every layer selects the best available capability for each specific task. The architecture debate is settled at every level of the stack — from silicon to consumer interface.
“Apple spent two years trying to build a smarter Siri alone. It could not. So it gave Google's Gemini the reasoning engine, wrapped it in Apple's privacy architecture, and shipped it to 2.2 billion devices. Microsoft built Copilot Cowork on Anthropic's Claude. Nvidia made NemoClaw hardware-agnostic. March 2026 settled the architecture question at every level of the stack: the companies that select the best model for each task — and govern it within their own framework — win. The ones still building everything in-house are already behind.”
What to Do This Week
Update your devices. iOS 26.4 delivers genuinely transformative capabilities. Experience what agentic AI feels like on a consumer device — then ask how your enterprise voice, document, and workflow systems compare.
Audit your customer experience against the new Siri. Your customers will interact with a Siri that chains multi-step actions, maintains months of context, and sees their screen. How does your customer support voice system compare? How does your document intake process compare? The consumer baseline just moved.
Evaluate your enterprise AI architecture. Apple, Microsoft, Nvidia, and Google have all independently adopted multi-model, model-agnostic architecture. If your enterprise AI stack is locked to a single provider, March 2026 has given you every signal needed to change that.
Plan for the agentic consumer. When 2.2 billion device owners experience AI agents that execute multi-step tasks autonomously, their expectations for every digital interaction shift. Enterprise systems that cannot match that expectation will lose customers to competitors that can.
The Signal Is Clear
Apple joining Google, Microsoft, and Nvidia in adopting a competitor's AI for its most important product is not a trend. It is a settled principle — confirmed independently by the four most valuable technology companies on Earth, within a single month.
No single model is best for every task. The orchestration layer that selects the best model, governs it within your framework, and routes each task to the optimal infrastructure is the most valuable piece of AI architecture for the next decade.
March 2026 proved it. From silicon to smartphone. The model-agnostic era is here — in your data centre, in your enterprise platform, and now in your pocket.