In the mid-1990s, a company called ProCD sold a national phone directory on CD-ROM, priced low for households and higher for businesses. Matthew Zeidenberg bought the household version and sold access to the listings online. The license forbidding commercial use came inside the box, and the software displayed it and refused to run until he accepted. Judge Frank Easterbrook upheld the license with a formula courts would repeat for decades:
"Notice on the outside, terms on the inside, and a right to return the software for a refund if the terms are unacceptable."
Easterbrook was solving a real problem. Nobody could print a full license on a shrink-wrapped box, and the cheap household price only worked if the seller could enforce its terms. So his test asked what the seller had shown and whether the buyer could walk away. It did not ask what the buyer understood, which meant agreement could be established by examining the packaging.
The web inherited the test and turned it into a question of layout. In 2002, a federal appeals court refused to enforce Netscape's license because the notice sat below the download button, where only people who scrolled would see it. Terms had to be "reasonably conspicuous," and assent had to be clear. Companies redesigned their screens to pass, and judges measured each new screen against the ones already approved, until a few layouts hardened into standard practice: the click-to-agree box, the blue link under the sign-up button.
Those screens showed something narrower than agreement: that a person was present, could have read the terms, and chose to proceed. The American Law Institute's 2024 Restatement of consumer contract law is an influential summary of how courts treat these agreements, and in its notes the scholars who drafted it conceded that "informed consent to standard contract terms is, by and large, absent in the typical consumer contract." The arrangement held together anyway, because the person who could have read the terms was also the person who clicked.
AI shopping agents, programs that browse and buy on someone's behalf, separate those two roles. The agent can read every clause, which almost no person does, but the person it works for is the one bound, and that person may never see the page.
Contract law already has a term for such a program. A 1999 model statute, adopted by most states, lets programs form agreements that no human reviews, treating the program as a "tool" of the person using it. In August, the Ninth Circuit used the same word for an AI shopping assistant, then left open whether a user is bound when an agent accepts a website's terms.
A published checkout specification requires the user to finalize a purchase on a trusted screen unless they have signed advance permission. In form, that is Easterbrook's arrangement again: a look before committing and a chance to decline. But Easterbrook's buyer saw the license itself. The agent's user confirms an order, perhaps beside a link to the checkout terms, while other terms may have been accepted earlier and out of sight, when the agent created an account. A person could approve a printer while the seller's records show a printer and an arbitration clause.
The incentives likely widen that gap without anyone planning it. Agents are measured on completed tasks, and sellers benefit when fewer shoppers pause over their terms. To software working its way toward checkout, an "I agree" box is one more step between the request and a finished task.
One remedy would be a final confirmation that lists every contract the agent accepted along the way, putting the arbitration clause right beside the printer. Easterbrook asked whether the buyer had a fair chance to refuse. Once an agent is doing the clicking, the record would also need to show that the buyer authorized it to say yes.
-
The November 20 hearing: In its motion to dismiss Amazon's amended complaint, Perplexity is fighting a new claim of tortious interference with customer contracts, which brings agent shopping one step closer to the contract question the Ninth Circuit left open.
-
Signed advance permission rules: The AP2 Mandates Extension requires that mandates never be created "without explicit user consent from trusted, deterministic channels," and it lists terms-of-service links among the checkout details to display, which makes it the clearest current attempt to record what the buyer actually authorized.
-
A platform's confirmation rule: Google's Gemini API terms, effective March 23, 2026, tell developers building computer-use agents that they "will not automatically bypass any requests for human confirmation," a contractual commitment that sits awkwardly beside agents measured on completed tasks.
-
The 1999 error protection: UETA's individual-error provision lets a person undo a mistake made while dealing with someone else's electronic agent, but its official comment confines that protection to human error, so it offers no general escape when your own AI clicks yes.

