Three announcements landed in the same June week. Apple approved Poke as the first AI agent on Messages for Business, charging a per-user fee. Meta made its WhatsApp Business AI agent available globally, pricing it into Business Premium subscriptions with token-based charges for large enterprises. Microsoft launched Scout, an OpenClaw-inspired personal assistant gated behind a GitHub Copilot subscription.
The coverage treated each as a standalone product launch. They are not separate stories. They are the same story, playing out on three different platforms simultaneously: the messaging layer is becoming the primary distribution channel for consumer AI agents, and every major platform holder has independently converged on the same business model — the per-user toll.
This is the App Store tax replicated for the AI era, and it will define the economics of the agent industry for the next decade.
The pattern, made explicit
Apple’s Poke approval is the cleanest signal. The startup, valued at $300 million, now operates on Apple’s Messages for Business platform — a channel previously reserved for businesses communicating with their own customers, not third-party AI agents. Poke co-founder Marvin von Hagen disclosed that his startup pays Apple on a per-user basis, calling the fee “significantly lower than Meta AI, after [Meta] increased fees in response to EU regulation.” The key detail is not the price — it is the structure. A per-user toll, applied at scale, turns Apple’s messaging platform into a revenue pipeline that scales with agent adoption.
Meta’s WhatsApp Business AI agent, available globally as of June 3, follows the same logic but with a twist. Meta charges businesses through WhatsApp Business Premium subscription tiers, with large enterprises paying based on token consumption. The business model mirrors the messaging revenue that WhatsApp already generates — except now the product being sold is not a message, it is an autonomous agent that handles customer support, lead qualification, and appointment booking. Meta is not just distributing AI agents; it is charging for every interaction they facilitate.
Microsoft’s Scout is the most expensive toll of the three, requiring a GitHub Copilot subscription to access. Built on the OpenClaw framework, Scout is an always-on agentic assistant embedded in Microsoft 365, operating across desktop and browser. The subscription gate means Microsoft collects its distribution tax at the account level, not the interaction level. The model is different in mechanics but identical in function: pay to access the agent through the platform’s distribution channel.
What the dominant narrative misses
The coverage has focused on each announcement’s product details — Poke’s approval process took months, Meta’s agent can book appointments, Scout comes with a “policy conformance system.” None of that is the story.
The story is that three of the world’s largest platform companies, acting independently and within days of each other, all concluded that the correct way to monetize AI agent distribution is through a platform toll. Apple chose per-user. Meta chose per-token (for large businesses) and per-seat (for SMBs). Microsoft chose per-subscription. The pricing unit varies. The architecture does not.
This convergence is not coincidental. It reflects a structural reality about AI agents that the industry is only beginning to grapple with: agents are not apps. They do not fit the existing distribution models that platforms have built. An app is a bounded artifact — you download it, it lives on your device, it has defined capabilities. An agent is an unbounded service that acts on your behalf, interacts with other services, and consumes compute continuously. The App Store’s one-time purchase and subscription models do not map cleanly onto agent economics. The platform toll — per user, per token, per interaction — is the simplest solution that works.
The trust bottleneck as a distribution moat
Apple’s Poke approval process took months and required the company to verify it could offer live human support, clearly label its AI agent, customize its interface to Apple’s guidelines, and submit testimonies from messaging providers. Von Hagen described the timeline as something any future entrant should expect: “It will also take them a couple of months to get through this approval process.”
This is not a bug. It is a distribution bottleneck that favors incumbents. Apple’s approval process for AI agents is not a security checklist — it is a gating mechanism that determines who gets access to the iMessage distribution channel and who does not. The per-user fee is the economic toll. The months-long approval process is the time toll. Together, they create a moat that only well-funded, patient startups can cross.
Meta’s approach is the mirror image. Its WhatsApp Business agent is available globally, but Meta controls the platform layer — it decides which enterprises can build custom agents, which systems those agents can connect to (Shopify, Zendesk, Shopee were named), and how those agents are billed. The openness of the distribution channel is inversely proportional to the platform’s ability to extract rent from it.
The WWDC question
WWDC 2026 starts next week, and Apple is rumored to announce an AI agent App Store. The MacRumors report, citing The Information, revealed that Apple has started contacting developers about integrating app capabilities into the revamped Siri. Some developers are already worried about commission fees. Apple is telling them it does not plan to charge during early stages — but that fees are “a possibility in the future.”
The developer skepticism is warranted. Apple’s playbook with the App Store was to establish a low-commission relationship with developers, then progressively extract more through paid search, in-app purchase requirements, and subscription fees. The agent App Store will follow the same trajectory. The only question is how long the “no fees” window lasts and what the eventual toll rate settles at.
The counter-narrative that changes the calculation
The dominant view is that the agent platform war will be won by the best models — that whichever company builds the most capable, most reliable, most trustworthy AI will own the agent ecosystem. This is wrong.
Model capability is already commoditizing. OpenAI, Anthropic, Google, and Meta are all producing frontier models within striking distance of each other. The switching cost between models is approaching zero as agent frameworks abstract model selection behind unified APIs. The real moat is not the model — it is the distribution layer. The platform that controls the messaging channel through which users access agents controls the economics of the entire ecosystem.
Apple controls iMessage. Meta controls WhatsApp. Microsoft controls Outlook, Teams, and the Microsoft 365 ecosystem. Each has a captive user base that already conducts daily interactions through these channels. AI agents are not displacing these channels — they are being layered on top of them. And the platform that owns the channel collects the tax.
What this means for agent builders
If you are building a consumer AI agent today, the economics just got more complex. Every platform you distribute through will take a cut — per user, per token, or per subscription. The Poke model (pay Apple per user) suggests the platform fee is manageable at small scale but becomes a significant cost line at scale. The Meta model (pay per token for large businesses) means your cost of goods sold includes not just model inference but platform distribution. The Microsoft model (gate behind an existing subscription) means your addressable market is limited to users who already pay for the platform.
The strategic implication is clear: agent builders need a direct distribution channel that bypasses platform tolls. Poke operates over SMS, Telegram, and WhatsApp in some markets — but Apple’s iMessage was the approval they sought because that is where the users are. The tension between platform distribution and platform dependency is the defining strategic question for every AI agent startup.
The editorial stance
The agent distribution tax is not a bug in the market. It is the market. The App Store tax was not introduced because Apple needed the revenue — it was introduced because Apple controlled the distribution channel and could set the terms. The same dynamics are now replicating in AI. Apple, Meta, and Microsoft all independently concluded, in the same week, that the distribution layer is where the value in AI agents will be captured. They are not wrong.
The open question is whether a viable alternative distribution model emerges — a decentralized agent marketplace, an open protocol for agent discovery, or simply agents that live entirely in the browser and bypass platform distribution entirely. None of these exist at scale today. Until they do, the platform toll is the price of admission.
The agent platform war is not about intelligence. It is about access. And the companies that control access have already decided how much it costs.
No comments yet