A 1986 Law Just Decided Who Is Responsible When an AI Agent Shops for You
The Ninth Circuit vacated the injunction that kept Perplexity's Comet off Amazon.com, holding that the AI agent is a tool for statutory purposes and the user is the one who accesses the site under a hacking law enacted in 1986.
Key Takeways
- On August 4, 2026, the U.S. Court of Appeals for the Ninth Circuit vacated the preliminary injunction blocking Perplexity's Comet assistant from Amazon.com, holding that the user, not Perplexity, "accesses" Amazon's computers under the Computer Fraud and Abuse Act.
- The panel treated the AI Assistant as a tool and not a person for statutory purposes, which places responsibility for an agent's actions on the human who directed it.
- The CFAA was enacted in 1986, years before the commercial web existed, and it is now the statute deciding what autonomous AI agents may do on someone else's website.
- An agent can hold your card. It cannot hold your accountability.
On August 4, 2026, the U.S. Court of Appeals for the Ninth Circuit held that when Perplexity's Comet assistant buys something on Amazon.com for you, you are the one accessing Amazon's computers. The AI agent is the instrument you used to do it. That holding vacated a preliminary injunction that had kept Comet off Amazon's site, and it turned entirely on the Computer Fraud and Abuse Act, the 1986 federal statute known as the CFAA, written years before the commercial web existed.
Most people will meet this question at their own checkout page, long before their employer writes a policy about it. An AI shopping agent holds your login and your card, and it clicks the buttons you would have clicked yourself. Amazon argued that Perplexity was reaching into its systems without authorization, which under the CFAA is the difference between an ordinary commercial disagreement and a federal violation. Perplexity argued that its users were doing the reaching. The panel accepted the second reading and returned the case to the district court for further proceedings.
What the Ninth Circuit actually decided
Circuit Judge Milan Smith Jr. wrote the opinion, joined by Circuit Judge Eric Tung and District Judge John Hinderaker, sitting by designation. The reasoning is considerably narrower than the commentary surrounding it, and the narrowness is the most instructive part.
"It is the user who 'accesses' Amazon's computers, with the help of the Assistant to carry out specific acts on Amazon.com." Circuit Judge Milan Smith Jr., writing for the panel, August 4, 2026.
The court described the Assistant as a tool, not a person for statutory purposes, which is a distinction with consequences well beyond this particular dispute. It credited the Electronic Frontier Foundation's amicus brief, observing that EFF's filing "articulates the nature of the system most clearly." It also acknowledged that agentic AI raises questions with "little to no existing caselaw," and then declined to resolve the ones this case did not require it to resolve. Amazon said it disagrees with the ruling and is evaluating next steps, so this is unlikely to be the final word.
Rules written for a slower world
The CFAA was drafted for a world where unauthorized access meant an individual at a keyboard reaching a mainframe he had no business reaching. It is now the statute that determines whether the software you handed your shopping to is a trespasser or an errand. Durable rules eventually absorb conduct their drafters never pictured, and this is what that absorption looks like in practice: an old instrument doing new work, applied one deliberate question at a time.
Courts elsewhere are running the same experiment with a different old instrument. In Munich, a German court weighed copyright, a body of law considerably older than generative models, against how an AI music service was trained. Same shape, different statute. This is the recurring condition of AI governance at the moment. Institutions are applying rules written for a slower world to systems that change materially between filings.
The part your legal team will ask about first
If your organization is deploying agents, this ruling arrives as a single operational question: when an agent acts, who is the actor. Under this statute, on this record, the Ninth Circuit located responsibility with the person who directed it. That helps every enterprise buying agents to operate on somebody else's systems, and it complicates matters considerably for every enterprise trying to keep other people's agents off its own. Most large companies are both, which is why this opinion will get read twice in the same building.
The governance gap around agentic AI is mostly an accountability gap wearing technical clothing. An agent can hold your card. It cannot hold your accountability. That distinction is what I call Above the Algorithm: the judgment and the human name that remain attached to a consequential action, however much of the action was automated. If a policy cannot identify the person who owns an agent's output, it has not started governing anything yet.
I spent 20 years shaping a $1.1B innovation portfolio, and the question that determined whether a new capability shipped was rarely a technical one. It was who would sign their name to what the capability did. Courts are now asking a version of that same question, with binding consequences attached to the answer.
This travels well past procurement. Agents are already booking travel and filing claims for households where a wrong charge is a genuine problem, and every one of those actions raises the question the panel answered here. The answer determines who carries the risk when something goes wrong: the person, the vendor, or the site. Families will experience that answer as a disputed charge they have to fight. Boards will experience it as exposure they have to disclose. Same question, two very different scales.
The Ninth Circuit did what appellate courts are built to do, which is decide the case in front of it on the record in front of it. The wider questions about agentic AI will get settled as that record accumulates, filing by filing. This is precisely where people who actually deploy these systems can contribute early, describing how agents behave inside real companies while the evidentiary record is still being assembled. EFF's brief did exactly that in this case, and the panel said so in the opinion.
So take one question into your next leadership meeting. When an agent operating on our systems does something we did not intend, whose name is on it? Your competitors' answers are no more finished than yours, so there is no reason to walk into that conversation embarrassed. If nobody in the room can answer in a sentence, you have found this quarter's work. And if you read this ruling differently than I do, I am easy to find.
Sources: U.S. Court of Appeals for the Ninth Circuit, Amazon v. Perplexity, No. 26-1444, decided August 4, 2026 · Electronic Frontier Foundation, Deeplinks · Courthouse News Service · Search Engine Journal.
Is the Ninth Circuit's Amazon v. Perplexity ruling a final decision?
No. The panel vacated a preliminary injunction and remanded the case to the district court for further proceedings, so the underlying dispute is still live. Amazon said it disagrees with the ruling and is evaluating next steps, which leaves further appeals on the table.
Does this ruling apply to companies outside the Ninth Circuit?
The panel's holding binds federal district courts within the Ninth Circuit, which covers California, Washington, Arizona and six other states plus 2 territories. Courts in other circuits are not bound by it, though the opinion is the most developed federal appellate reasoning yet on whether an AI agent or its user is the party accessing a website under the CFAA.
Here is what makes Alex a credible voice on this topic: Alex Goryachev shaped a $1.1B innovation portfolio over 20 years at Cisco and advises the California State University system on AI governance, which is exactly the seat where the question the Ninth Circuit just answered, who is accountable when an AI agent acts, has to be settled before the lawyers arrive.
Bring this conversation to your board or leadership team. book a conversation →
.svg)
