AI Agents Are Becoming the Customer. Here Is the $100 Billion Market Nobody Is Building For.

The Shift Nobody Is Naming Clearly

For the entire history of the internet, the end user was a human being. You built websites for humans. You built apps for humans. Your entire product strategy assumed a human on the other end.

That assumption is changing. AI agents are becoming the customer.

Not replacing human customers. Joining them. An agent browsing on behalf of a user, making purchases with approved budgets, filing support tickets, processing refunds, routing communications , all without a human initiating each individual action.

What Agents Need That Humans Do Not

The infrastructure for human internet users was built over 30 years: login systems, checkout flows, inbox interfaces, identity verification. Almost none of it was designed with an automated actor in mind.

Agents need different infrastructure. Five categories Greg Isenberg identifies:

Identity. Who is this agent acting for? How does a website verify that an agent has legitimate authority from the human it represents?

Inbox. Where do responses, OTPs, documents, and replies land for an agent? AgentMail , backed by YC , is already building email inboxes specifically for AI agents, not humans.

Wallet. What can the agent spend? Stripe recently launched infrastructure to give agents wallets with spend caps, approval rules, and audit trails. A purchasing agent that can buy software on your behalf needs exactly this.

Memory. What does the agent know about your preferences and rules? The agent managing your inbox needs standing instructions that persist across sessions.

Receipts. What did the agent see, decide, change, and buy? The audit trail is the accountability layer.

The Opportunity That Is Already Open

Every category in that list is underbuild. The companies building the infrastructure for agent identity, agent inboxes, agent wallets, and agent audit trails are not competing with Google or Microsoft. They are building plumbing for a new type of user that the existing internet was not designed for.

The comparison Isenberg uses: think of agents like a new employee. As trust grows, you give them more. First a task. Then access to tools. Then a budget. Then a corporate card with limits. The infrastructure that manages that progression for millions of AI agents simultaneously does not fully exist yet.

The opportunity is in the infrastructure, not the agents themselves.

And the window in which that infrastructure is underbuild is closing.