The Agentic Internet: When Machines Become Your Customers_
The Web Has New Primary Users — And They're Not Human
In April 2024, Imperva released a finding that should have been front-page news: automated traffic surpassed human traffic on the web for the first time, accounting for 51% of all web requests. Of that, 37% was bad bot traffic — scrapers, credential stuffers, fraud networks — and 14% was good bot traffic: search engine crawlers, monitoring services, and increasingly, AI retrieval agents.
By Q1 2025, the composition had shifted further. Retrieval bot traffic — AI agents fetching information on behalf of users and other systems — grew 49% quarter over quarter. These aren't traditional crawlers indexing pages for search results. They're agents retrieving structured data, comparing options, initiating transactions, and reporting back to orchestrating systems.
The implications are hard to overstate. For thirty years, the internet was built for human eyeballs. Every pixel, every click path, every conversion funnel was optimized for a person sitting in front of a screen, making decisions with their attention and their mouse. That architecture is now encountering a fundamentally different user: one that doesn't see pixels, doesn't get fatigued, doesn't respond to emotional design, and processes information at machine speed.
The web is becoming a machine-to-machine medium with human oversight. And most of the business models built on it assume the opposite.
From Prompt to Answer to Intent to Execution
The AI interaction model is evolving through a critical phase transition. The first generation — ChatGPT, Perplexity, Claude — operated in a prompt-to-answer paradigm. A human asks a question. The AI returns text. The human reads, evaluates, and decides what to do next. The AI is a tool. The human retains agency.
The second generation — now emerging — operates in an intent-to-execution paradigm. A human expresses an intent ("find me the three best CRM vendors for a 50-person B2B SaaS company, schedule demos with all three, and prepare a comparison matrix after each demo"). An orchestrating agent breaks this into subtasks, deploys specialized agents to each, manages dependencies, handles errors, and returns completed outcomes.
The human doesn't choose which websites to visit. The agent does. The human doesn't compare pricing pages. The agent does. The human doesn't fill out contact forms. The agent does. The entire click path — the journey that every SaaS company's growth team has spent years optimizing — gets bypassed.
This isn't theoretical. Anthropic's Model Context Protocol (MCP) is enabling agents to interact with external systems through standardized tool interfaces. OpenAI's function calling and Google's agent frameworks are doing the same. The infrastructure for agent-mediated commerce and operations is being built right now.
What Breaks
Seat-Based SaaS Pricing
The most obvious casualty is the per-seat pricing model that has defined SaaS economics for two decades. When a company pays $50/seat/month for a project management tool, they're paying for human access to a UI. When an agent can create tasks, assign resources, track progress, and generate status reports via API — without ever rendering a single pixel — the seat has no user.
This doesn't mean the underlying system has no value. The data model, the workflow engine, the permission system, the integration layer — these remain essential. But the pricing model that wraps them in a UI and charges per human accessor is fundamentally mismatched with agentic consumption.
UI-Only Tools
Products whose primary value proposition is a well-designed interface face existential pressure. If the core function is "make it easier for humans to do X," and an agent can do X without any interface at all, the product's moat evaporates. Design tools that help humans arrange pixels. Dashboard tools that help humans visualize data. Communication tools that help humans coordinate. All of these assume a human in the loop who benefits from visual presentation.
Agents don't benefit from visual presentation. They benefit from structured data, clear APIs, and reliable execution.
The Advertising-Attention Complex
The entire digital advertising ecosystem assumes human attention as the currency. CPM (cost per thousand impressions) requires impressions on human eyeballs. CPC (cost per click) requires human clicks. When agents are the ones retrieving information and making decisions, impressions and clicks lose meaning. Cloudflare reported blocking 416 billion AI bot requests in a single reporting period — requests that, if they'd reached their targets, would have consumed resources without generating any advertising value.
The Washington Post's partnership with TollBit represents an early attempt to adapt: licensing content to AI systems for a fee, rather than relying on ad-supported human readership. This points toward an "AI readership" economy where content is monetized by machine consumption, not human attention.
What Persists
Systems of Action
While UI layers face pressure, the systems underneath them — the ones that actually do things — become more valuable, not less. Database systems, workflow engines, transaction processors, identity providers, permission systems: these are the bedrock that agents need to operate. An agent can bypass a dashboard, but it can't bypass the database the dashboard queries.
The companies that own systems of action — not systems of presentation — are positioned to thrive in an agentic internet. Their value isn't in how they display information to humans. It's in what they enable machines to do.
Trusted Execution Environments
As agents gain the ability to take consequential actions — spending money, signing contracts, modifying production systems, communicating on behalf of organizations — the question of trust becomes paramount. Who authorized this agent? What are its permissions? What audit trail exists? What happens when it makes an error?
Trusted execution environments — systems that provide verifiable identity, scoped permissions, immutable audit logs, and rollback capabilities for agentic actions — don't exist as a mature category today. But they will be essential infrastructure. Every agent acting on behalf of a company needs to operate within a governance framework that the company controls, audits, and can explain to regulators.
Agent-Native Infrastructure
a16z's "agent-speed" workloads thesis identifies a critical gap: most existing infrastructure was designed for human-speed interaction. Databases expect query patterns that reflect human workflows. APIs have rate limits calibrated for human usage patterns. Authentication systems assume a human entering credentials.
Agent-speed workloads look fundamentally different. Thousands of concurrent requests. Sub-second decision cycles. Parallel execution across dozens of systems simultaneously. Infrastructure that can handle this — while maintaining governance, audit trails, and cost controls — is a greenfield opportunity.
The SaaS Archetypes Under Pressure
Not all SaaS categories face equal pressure from the agentic shift. The impact varies by how much value lives in the UI versus the underlying system:
| SaaS Archetype | UI Dependency | Agentic Pressure | Likely Outcome |
|---|---|---|---|
| UI-Heavy (design tools, dashboards) | Very High | Severe | Must rebuild as API-first or become agent tooling |
| API-First (Stripe, Twilio, AWS) | Low | Minimal | Already agent-native; pricing aligns with usage |
| Compliance-Heavy (GRC, audit, identity) | Medium | Opportunity | Governance demand increases as agents proliferate |
| Content Platforms (CMS, media, publishing) | High | Severe | Must monetize machine readership or lose distribution |
The pattern is clear: value migrates from presentation to execution, from interface to infrastructure, from human-readable to machine-executable.
Gartner's Sobering Forecast
Gartner's projections paint a nuanced picture of the agentic transition. On one hand, they forecast that 40% of enterprise applications will integrate AI agents by the end of 2026 — a remarkably fast adoption curve. On the other hand, they predict that more than 40% of agentic AI projects will be cancelled by 2027, primarily due to cost overruns and governance failures.
This isn't contradictory. It's the natural shape of a technology transition where the tooling outpaces the governance. Companies will rush to deploy agents because the productivity promise is irresistible. Many will discover that agents without proper guardrails — cost controls, permission scoping, error handling, audit capabilities — create more problems than they solve.
The cancelled projects won't prove that agentic AI doesn't work. They'll prove that agentic AI without governance infrastructure doesn't work. And they'll create massive demand for the governance layer itself.
The "AI Readership" Economy
The Washington Post's deal with TollBit signals the emergence of a fundamentally new economic model: content monetized not by human readers but by machine consumers. When an AI agent retrieves an article to synthesize a research brief for its user, that retrieval has economic value — but the traditional advertising model captures none of it.
New monetization models are emerging:
- Licensing fees for AI access to proprietary content (TollBit, licensing.ai)
- Tiered API access with different pricing for human versus machine consumption
- Provenance premiums where AI systems pay more for verified, high-quality sources
- Structured data feeds optimized for machine consumption rather than human reading
This represents a profound shift in how value flows through the information economy. Content that is authoritative, structured, and machine-accessible becomes more valuable — not less — as AI agents become primary consumers.
Governance-First, Not Feature-First
The conventional approach to building software products is feature-first: ship functionality, acquire users, iterate on the interface, optimize conversion. In an agentic world, this sequence inverts. The first question isn't "what features do users want?" It's "what governance framework allows agents to operate safely, accountably, and within defined boundaries?"
This is why AI Gens builds governance-first. Every venture in our portfolio is designed with agent interaction as a primary use case, not an afterthought. Permission models, audit trails, cost controls, and accountability frameworks are foundational architecture — not compliance add-ons bolted on after launch.
The agentic internet doesn't reward the product with the best UI or the most features. It rewards the product with the most trustworthy execution environment. When your customer is a machine acting on behalf of a human, trust isn't a nice-to-have. It's the entire value proposition.
The companies that understand this — that build for machine customers with human accountability — will define the next era of enterprise infrastructure. The ones that keep optimizing click paths for human eyeballs will find themselves serving an audience that's shrinking by the quarter.
The web's new primary users have arrived. The question is whether your infrastructure is ready to serve them.