What Happens to the Web When AI Agents Do the Clicking

From 2016 until about April 2023, a member who wanted to cancel Amazon Prime online had to find, and then finish, a process Amazon called “Iliad,” after Homer’s epic of the long Trojan War. The Federal Trade Commission’s amended complaint describes it as a four-page, six-click, fifteen-option process, while joining took one or two clicks. The complaint alleges that the length was the point: Amazon designed the flow to thwart the subscribers who entered it. It says an internal working group found customers abandoning the flow partway in the belief that they had already cancelled. It also says executives slowed or rejected changes that would have simplified it because those changes hurt the bottom line.1 A member who gave up partway stayed a member and kept paying. Amazon settled in September 2025, agreeing to a $1 billion civil penalty and $1.5 billion in refunds without admitting or denying the allegations.2

The flow was built for a person with eyes, hands and finite patience, and it worked because that person got tired.

For thirty years the web was built around that person. Ad impressions, checkout funnels, retention flows: the whole economy assumed persuasion happens at eye level, and that if you design the screen you’ve designed the user.

That assumption breaks once software can do the waiting and clicking, and the web becomes a post-user web, where whoever is at the screen may not be a person.

Institutions run on machine time. They can generate administrative demand at near-zero marginal cost and wait forever. People process sequentially, forget passwords, and get tired on hold. In that mismatch the interface stops being a tool for access and becomes the place where human slowness gets converted into abandonment, and abandonment into revenue.

Deploying against the web

The useful thing about a personal AI agent is persistence, not creativity: the ability to separate your intent from your attention. Web interfaces have required your attention to secure your rights. An agent breaks that link. You won’t walk the four pages. You’ll say “refund this,” and the agent will do the one thing the flow was built to prevent, which is not get tired.

That changes the economics of the interface.

Diagram

Your agent meets the four-page cancel flow

View source

%% title: Your agent meets the four-page cancel flow
%% caption: You hand a refund to an agent that never tires of the four pages. The site answers by shifting its layout to break the agent, settling the dispute bot to bot, and asking you to prove you are alive before you can reach a person.
flowchart TD
    A["You tell an agent:<br/>‘refund this’"] --> B["Your agent, which<br/>does not get tired"]
    B --> C["Site’s endpoint<br/>built for machines"]
    B --> D["The four-page,<br/>six-click flow"]
    D --> E["Layouts shift, code changes<br/>to break the agent"]
    E --> B
    C --> F["Refund negotiated in<br/>milliseconds, unwatched"]
    F --> G["Your agent against<br/>the site’s bots"]
    G --> H["Biometric check: prove<br/>a living person is here"]
    H --> I["Walled garden: you reach a<br/>person after proving you are one"]

Friction was a business model. The dropout rate, the share of people who abandon a claim, a cancellation or a dispute, was a line of revenue, and it depended on human patience being finite.

The institutional response is to make automation expensive.

Expect the human-readable web to degrade. Sites already go polymorphic, shifting their code, randomizing layouts and obfuscating data to break scrapers. The visual web becomes a heavy, shifting surface that’s hard for machines to parse and unpleasant for people to look at. A clear page made sense when the reader was a person and the person was the revenue. When the reader may be an agent filing a refund, a clear page is a weakness for the site.

Proof of life

As agents take over the negotiation of daily life (buying, scheduling, disputing), the interaction speeds up past human comprehension. Dispute resolution starts to look like high-frequency trading: millisecond negotiations between systems that nobody watches. When errors happen at that speed they happen at scale.

And if behavior can be spoofed, institutions stop trusting behavior. They demand the biological operator. Login starts to require proof that a living person is present, and biometric checks become the price of entry, so the site can tell the account holder from automated traffic.

The web splits. On one side, bot-to-bot infrastructure: fast, invisible, adversarial.3 On the other, walled gardens where people can talk to people, after proving, legally and biologically, that they’re real.

Sites are moving toward asking people only to authenticate, authorize, and occasionally absorb the consequences of decisions made elsewhere, and toward treating a person’s slower pace as a delay to route around. That follows from platform incentives running unchecked, refusal staying privatized, and systems built without having to answer to the human beings they exploit.