Technical Deep Dive
The death of the internet business model is not a metaphor; it is an architectural inevitability. Traditional digital platforms operate on a two-sided market logic: they aggregate supply (content, products, services) and demand (users, buyers), and extract rent from the connection. Google Search monetizes the gap between a query and a result. Amazon monetizes the gap between a product listing and a purchase. Facebook monetizes the gap between a user and an ad. LLMs and AI agents collapse these gaps.
The Reasoning Layer as the New Bottleneck
The key technical shift is the emergence of the 'reasoning layer.' In the old stack, you had:
- Infrastructure layer (cloud, compute)
- Platform layer (search, social, e-commerce)
- Application layer (websites, apps)
The platform layer was the moat because it controlled the index of supply and the attention of demand. Now, LLMs introduce a new layer above the application layer: the reasoning layer. This layer takes natural language intent, decomposes it into sub-tasks, retrieves information from multiple sources (via tools like web search, APIs, or vector databases), and synthesizes an action or answer. The LLM does not just retrieve a link; it reasons about what the user actually needs.
For example, consider an AI agent like OpenAI's Operator (released early 2025) or Anthropic's Computer Use. These agents can navigate a browser, fill forms, compare prices across multiple e-commerce sites, and complete a purchase. They do not click on ads. They do not browse through sponsored results. They go directly to the most relevant source. The platform's ad-based revenue model is bypassed entirely.
Technical Mechanisms of Disintermediation
1. Tool Use and Function Calling: Modern LLMs (GPT-4o, Claude 3.5, Gemini 2.0) support function calling, allowing them to invoke external APIs—flight booking APIs, payment gateways, inventory systems—without a human intermediary. This turns the LLM into a direct transactional agent.
2. Multi-Agent Systems: Frameworks like AutoGPT (now at 160K+ GitHub stars), CrewAI (25K+ stars), and Microsoft's AutoGen (35K+ stars) allow multiple LLM agents to collaborate. One agent can negotiate a contract while another checks legal compliance. These systems replace entire B2B platforms.
3. Retrieval-Augmented Generation (RAG): RAG pipelines allow LLMs to pull real-time data from proprietary databases, public web sources, or internal documents. This means an AI agent can answer a question or complete a task without ever visiting a traditional website. The 'content' is consumed by the AI, not by a human with eyeballs that can be monetized.
4. Autonomous Web Navigation: Projects like WebGPT (OpenAI) and the open-source Playwright-based agents can execute JavaScript, handle cookies, and fill forms. They simulate human browsing but with perfect efficiency—no distractions, no ads, no impulse purchases.
Data Table: Performance of Leading AI Agents on Autonomous Task Completion
| Agent / Model | Task Completion Rate (WebArena) | Average Steps per Task | Cost per Task (API) | Human Baseline |
|---|---|---|---|---|
| GPT-4o + Operator | 78.4% | 12.3 | $0.42 | 92.1% |
| Claude 3.5 + Computer Use | 71.2% | 15.1 | $0.38 | 92.1% |
| Gemini 2.0 + Agentic | 69.8% | 14.7 | $0.35 | 92.1% |
| Open-source (AutoGPT + GPT-4) | 52.3% | 22.4 | $0.55 | 92.1% |
Data Takeaway: While still below human performance, the top proprietary agents are approaching 80% task completion. At this rate, by late 2026, AI agents will match or exceed humans on routine digital tasks. When that happens, the economic value of human attention on traditional platforms collapses.
Key Players & Case Studies
The race to own the reasoning layer has three distinct camps: the model makers, the platform incumbents, and the agent-native startups.
Camp 1: The Model Makers (OpenAI, Anthropic, Google DeepMind)
These companies are building the reasoning engines. Their strategy is to make the LLM the default interface for all digital interaction.
- OpenAI: With GPT-4o and the Operator agent, OpenAI is positioning itself as the 'operating system for agents.' They recently launched a 'GPT Store' for custom agents, taking a 20% cut of agent-to-agent transactions. This is a direct play to replace the app store model.
- Anthropic: Claude 3.5 Sonnet's 'Computer Use' feature allows the model to control a desktop environment. Anthropic is targeting enterprise workflows—automating data entry, CRM management, and procurement. They charge per task, not per user, bypassing traditional SaaS subscription models.
- Google DeepMind: Gemini 2.0 is deeply integrated with Google's own services (Search, Maps, Gmail). Google's strategy is defensive: keep the reasoning layer tied to its existing ad ecosystem. But this creates a conflict of interest—Gemini must decide whether to show an ad or give the best answer. Early tests show Gemini often suppresses ads in favor of direct answers, cannibalizing Google's core revenue.
Camp 2: Platform Incumbents (Amazon, Meta, Microsoft)
These companies are trying to retrofit their platforms for the AI era, with mixed results.
- Amazon: Launched 'Buy with AI' in early 2025, allowing agents to purchase on Amazon via API. But this reduces Amazon's role to a fulfillment provider, not a discovery platform. Amazon's ad revenue (over $40B annually) is at risk if agents bypass product search.
- Meta: Pivoting to AI-generated content. Their 'AI Studio' lets users create AI personas that interact on Facebook and Instagram. The business model is unclear—will Meta sell ads to AI personas? Or charge for AI-to-AI interactions?
- Microsoft: Copilot is embedded in Office 365, but the subscription model is under pressure. If an AI agent can generate a PowerPoint deck in seconds, why pay $30/user/month? Microsoft is experimenting with per-usage pricing, but this could reduce ARPU significantly.
Camp 3: Agent-Native Startups
These are the disruptors building on the new paradigm.
- Adept AI: Founded by former Google researchers, Adept is building an 'AI agent for the enterprise' that can use any software tool. They raised $350M at a $2B valuation. Their pitch: 'Your company doesn't need 20 SaaS subscriptions; you need one agent that can use them all.'
- Cognition Labs: Creators of Devin, the AI software engineer. Devin can autonomously code, debug, and deploy. This threatens platforms like GitHub (owned by Microsoft) and Upwork. Cognition recently announced a 'Devin-as-a-Service' model, charging per feature delivered, not per developer seat.
- Browserbase: An open-source platform (15K+ stars) that provides a 'headless browser for AI agents.' It allows agents to interact with any website programmatically. This is the infrastructure for bypassing traditional web interfaces.
Data Table: Business Model Comparison
| Company | Traditional Model | AI-Native Model | Revenue Impact |
|---|---|---|---|
| Google | Ad-based search ($240B revenue) | Per-query AI subscription (Gemini Advanced) | Potential 30-50% revenue decline by 2028 |
| Amazon | Marketplace fees + ads ($50B+ ad revenue) | Fulfillment-only + AI transaction fees | Ad revenue at risk; fulfillment margins thin |
| Salesforce | Per-seat subscription ($35B revenue) | Per-task AI agent pricing | ARPU could drop 60%+ |
| Upwork | Commission on freelancer earnings ($700M revenue) | AI agent replaces freelancers | Business model obsolete |
Data Takeaway: The shift from per-seat to per-task pricing is the single biggest threat to SaaS and platform companies. Per-task pricing aligns cost with value delivered, but it dramatically reduces total revenue for companies that previously charged for unused capacity.
Industry Impact & Market Dynamics
The disruption is already visible in market data. Digital advertising growth is slowing: after 12% growth in 2023, it dropped to 8% in 2024, and projections for 2025 are below 5%. Meanwhile, AI agent API usage is exploding: OpenAI reported a 300% increase in API calls from agent frameworks in Q1 2025 alone.
The Death of the Click
The core unit of the internet economy—the click—is being replaced by the 'task.' An AI agent does not click on a link; it executes a function. This has profound implications:
- Cost-per-click (CPC) advertising becomes meaningless. Advertisers pay for attention that never arrives.
- Affiliate marketing collapses. If an AI agent books a hotel directly via API, there is no affiliate link to track.
- Content monetization (paywalls, ad-supported media) fails because AI agents consume content without generating ad impressions or subscribing.
The Rise of the Agent-to-Agent Economy
A new market is emerging: agents negotiating with other agents. For example, a travel agent (AI) negotiates with a hotel booking agent (AI) for the best rate. These transactions happen in milliseconds, with zero human involvement. The value capture shifts to the platform that hosts the agent marketplace. OpenAI's GPT Store and Microsoft's Azure AI Agent Service are early examples. They take a small percentage of each agent-to-agent transaction, similar to a payment processor.
Market Size Projections
| Segment | 2024 Market Size | 2028 Projected Size | CAGR |
|---|---|---|---|
| Traditional digital advertising | $680B | $720B | 1.5% |
| AI agent transaction fees | $2B | $150B | 150% |
| AI agent subscriptions (B2B) | $5B | $80B | 100% |
| Traditional SaaS subscriptions | $250B | $200B | -5% |
Data Takeaway: The agent economy is growing from near-zero to a $230B+ market in four years, while traditional models stagnate or decline. This is not a cycle; it is a structural shift.
Risks, Limitations & Open Questions
The Reliability Problem
Current AI agents still fail on complex, multi-step tasks. The WebArena benchmark shows a 20-30% failure rate. In mission-critical domains like healthcare or finance, a single failure can be catastrophic. Enterprises are hesitant to fully automate procurement or legal workflows.
The Alignment and Safety Question
If agents act autonomously, who is liable when they make a bad decision? An AI agent that books a non-refundable flight on the wrong date has no legal standing. Current frameworks (like Anthropic's 'constitutional AI') are not designed for economic agency. Regulators are beginning to ask: should AI agents be required to have 'licenses' to transact?
The Monopoly Risk
The reasoning layer is likely to be controlled by a handful of companies—OpenAI, Anthropic, Google. This creates a new kind of bottleneck more powerful than any platform before. If OpenAI decides to favor its own services in agent recommendations, it becomes a de facto monopoly over digital commerce. Antitrust frameworks are not equipped to handle this.
The Data Paradox
AI agents need access to real-time, high-quality data to make good decisions. But the data sources (websites, APIs) are increasingly blocking AI agents. The New York Times, Reddit, and many others have sued or restricted AI crawlers. If data access is cut off, agent performance degrades. This could lead to a 'data war' where only the largest AI companies can afford to license data.
AINews Verdict & Predictions
Prediction 1: By 2027, the top 10 internet companies by market cap will include at least two 'agent-native' companies that did not exist in 2020. The current incumbents are too slow to pivot. Google's ad business is a cash cow they will not kill, but it will be killed for them.
Prediction 2: The 'per-seat' SaaS model will be dead by 2029. Companies like Salesforce, Adobe, and Atlassian will be forced to adopt per-task pricing, leading to a 50%+ revenue contraction. The survivors will be those that own the reasoning layer, not the application layer.
Prediction 3: A new regulatory category—'AI Economic Agent'—will emerge by 2026. Agents will need to register, post bonds, and comply with transparency rules. This will slow adoption but ultimately create a trusted market.
Prediction 4: The most valuable company in the world in 2030 will not be a search engine, a social network, or an e-commerce platform. It will be a company that owns the reasoning layer for enterprise transactions. My bet is on Anthropic, given their lead in safety and enterprise adoption, but OpenAI's distribution advantage is formidable.
What to watch next: The battle for the 'agent operating system.' Watch for Microsoft's attempt to make Azure the default runtime for enterprise agents, and for Apple's entry into the agent space with a privacy-first approach. The next 18 months will determine who controls the digital economy for the next decade.