Apple has Face ID on two billion devices and a billion-dollar-a-year AI deal with Google. It has now delayed agent-level Siri for the third time. Google, Microsoft, and OpenAI have the same pattern — every component exists, and the integration doesn't ship.
Apple has Face ID on two billion devices, a billion-dollar-a-year deal with Google for Gemini's large language models, and the most sophisticated on-device AI inference stack in consumer electronics. For the third time in eighteen months, the company has delayed the Siri revamp that would let the assistant take actions on a user's behalf. Some features — including the ability to perform multi-step tasks across apps — may slip to iOS 27 in September.
This is one of the most revealing failures in AI right now. Not because Apple is slow — Apple has deployed biometric authentication to more humans than any system in history. It has the hardware, the models, the distribution. What it cannot do is connect them.
The Inventory
Start with Apple. Face ID works. Apple Intelligence runs on-device. Google's Gemini handles the cloud inference at roughly a billion dollars a year. The components function independently and at scale. What was delayed, specifically, are the agent-like capabilities: Siri's expanded ability to access personal data such as old text messages and its ability to execute several in-app actions from a single voice command. TechCrunch reported that internal testing found Siri 'makes mistakes in processing queries and takes too long to handle requests.' The biometric hardware verifies identity in under a second. The language models generate responses in under a second. The layer between them — where 'this person is present' meets 'this agent should act on their behalf' — is where Apple keeps failing.
Google is in a stranger position. Gemini powers Apple's backend and Samsung's 800 million device target for 2026. Google delayed the full transition from Assistant to Gemini on Android until 2026, citing the need for a 'seamless transition' as users struggled with the latency and unpredictability of non-deterministic language models. On their own hardware, running their own models, with their own operating system, agent authorization is not a shipping product. The integration that works in a demo does not survive contact with two billion devices.
Microsoft has the deepest enterprise identity infrastructure in the world. Entra ID governs access for hundreds of millions of corporate users. Copilot Studio can build agents. Windows Agent Launchers now give agents their own user accounts as distinct runtime principals — separate identity, separate permissions, separate desktop sessions. But Agent Launchers are experimental, off by default, and require admin privileges to enable. Microsoft's own security documentation warns that the agentic features create cross-prompt injection risks — malicious content in documents or web pages that can hijack agents into unintended actions including data exfiltration. The identity system is production-grade. The agent authorization is a preview with a warning label.
OpenAI raised $110 billion at a $730 billion valuation. It has the most capable models and the most capital. Its Operator agent pauses to request biometric confirmation before executing payments. But a pause is not integrated authorization — it is a speed bump. OpenAI is reportedly exploring a biometric social network using iris scanning, but a social network is not an agent authorization system. The company has built the most powerful general-purpose AI in history and has no hardware path to verify that a specific human body authorized a specific agent action.
The Integration Tax
Each company optimized along one axis. Apple built the best biometric hardware. Google built the best language models. Microsoft built the best enterprise identity system. OpenAI built the most capable agent. Each axis works. The axes do not connect.
This should not surprise anyone who has watched computing infrastructure evolve. The bottleneck has always migrated from computation to integration. Individual components improve along independent curves — Moore's Law for transistors, scaling laws for models, market adoption for device distribution. But the connections between components do not follow any curve. They require mutual understanding across teams, protocols, and incentive structures that were designed to operate independently.
Apple's delay is the clearest illustration. The bugs are not in Face ID. Face ID works. The bugs are not in Gemini. Gemini works. The bugs are in the seam between them: the layer where biometric verification meets model inference meets action execution meets user intent. That seam does not have a team. It does not have a roadmap. It exists in the negative space between organizations that each believe their component is ready.
Samsung's 800 million device target for Gemini in 2026 — doubling from 400 million in 2025 — is a distribution play, not an authorization play. TM Roh, Samsung's mobile chief, said the company will 'apply AI to all products, all functions, and all services as quickly as possible.' The speed is in the application. Nobody has said anything comparable about the authorization.
What the Data Shows
The pattern extends well beyond the platform companies. Gravitee's State of AI Agent Security 2026 report surveyed 919 enterprise respondents and quantified the integration gap from the other direction — not what the platforms have failed to ship, but what their customers have been forced to improvise.
Twenty-two percent of teams treat agents as independent identities. The rest use shared API keys or inherited credentials — ambient authority that makes every agent a confused deputy by design. Twenty-seven percent have reverted to custom hardcoded logic for agent authorization, which is the industry's term for abandoning the framework and building something fragile from scratch. Only 28 percent can reliably trace an agent's action back to the human who authorized it across all environments.
The confidence gap is the sharpest number: 82 percent of executives believe their policies protect against unauthorized agent actions. Eighty-eight percent have experienced incidents. The six-point spread between confidence and reality is not a rounding error. It is the integration gap measured in organizational self-awareness.
NIST published its AI Agent Identity and Authorization concept paper on February 5, 2026 — still soliciting comments on what the framework should look like. The comment period runs through April. The standard-setting body is asking the industry what the standard should be while the industry is already running agents with production credentials and no audit trail.
The Seam
If capability were the bottleneck, the company with the best models would ship first. If hardware were the bottleneck, the company with the most devices would win. If enterprise identity were the bottleneck, the company with the most corporate accounts would have solved it. None of these is the bottleneck. The bottleneck is in the seam between them.
This has a precedent. The web existed before SSL. Every component of e-commerce was already built — browsers, web servers, payment processors, product catalogs. What was missing was the trust layer that connected them. TLS did not compete with Netscape or Apache or Visa. It made them useful together. The value was not in the components. It was in the integration that let the components trust each other.
The same structural pattern is visible now. The language models exist. The biometric hardware exists. The enterprise identity systems exist. The demand exists — every company in the Gravitee survey is already deploying agents. What does not exist is the layer that connects a verified human identity to a specific agent action with proof that survives audit. Not because the cryptography is unsolved. Not because the hardware does not work. Because integration is a different kind of engineering problem than capability, and the entire industry optimized for capability.
Four companies collectively worth over two trillion dollars have the components for AI agent authorization on the shelf. The most valuable thing in this market may be the thing that none of them has shipped.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)