Every agentic AI pitch you have sat through this year describes software that goes off and does things for you. On Tuesday the Ninth Circuit answered the question the pitch decks skip. When the software does something, who did it? Not the software, and not the company that sold it to you.

THE LEAD PLAY

The Agent Didn't Access Anything. You Did.

Amazon sued Perplexity in the Northern District of California over the agentic assistant that signs into a user's Amazon account and shops on their behalf. Judge Maxine Chesney granted Amazon a preliminary injunction in March. On August 4 the Ninth Circuit vacated it, in a published opinion by Judge Milan Smith, joined by Judge Eric Tung and District Judge John Hinderaker sitting by designation.

Amazon's claim ran on the Computer Fraud and Abuse Act, which starts with somebody accessing a protected computer. Perplexity's answer was that it had never accessed anything. The panel agreed: “It is the user who ‘accesses’ Amazon's computers, with the help of the Assistant to carry out specific acts on Amazon.com.” And the line you will hear read back to you in vendor meetings for the next three years: “However advanced the Assistant currently is, it is a tool, not a person for statutory purposes.”

Amazon's claim under California's computer fraud statute failed the same way for the same reason. Injunction vacated, case back to Chesney.

The panel drew the box tight. It wrote that it was not establishing “a new legal regime governing agentic AI,” said nothing about whether Perplexity could avoid liability on other theories including tort, and expressly declined to reach the rest of the CFAA.

A CFAA claim is two questions stacked: did somebody access the computer, and were they permitted to. The Ninth Circuit answered the first one and put the answer on the human who pointed the agent. It left the second one wide open.

So think about where your agents already reach into systems that belong to somebody else. Medical records portals. Insurer claims portals. County recorder sites. Court e-filing. Carrier MVR services. The terms of use on most of those were drafted to prohibit automated access, and most of them are entered under a named human's credentials, sitting in a config file so the job can run overnight without anybody watching.

The firm owns the first element now. The entire fight moves to the second one, and it does not need the CFAA to hurt. The same facts support a breach of contract claim, a state computer-access statute, or a records vendor shutting off your account in the middle of a case.

Perplexity won this appeal by telling a federal court that its own product is not a person. Every agentic vendor selling into your firm has that argument available now.

The Play this week: Ask whoever runs your automations for two columns. Every system outside the firm that a bot, script, or agent signs into, and whose credentials it uses to get there. Not the tools you bought. The systems you reach into. Records portals, insurer portals, e-filing, recorders, MVR vendors, skip-trace services.

Then pull the terms of use for each one and find the automated-access clause. What you are looking for is whether the account you are using is permitted to be operated by software. Where it isn't, the fix is usually a phone call. Most records vendors already sell an API or a bulk retrieval product, and they would rather sell you one than find you in their logs.

SECOND CHAIR

He Shepardized every case. It didn't save him.

Judge Vernon Oliver sanctioned Hilary Miller, a sole practitioner, on August 4 in Barteca Holdings LLC v. TacoBarn Newtown LLC, No. 26-cv-250 (D. Conn.). Miller wrote the brief himself, conventionally. Then he ran it through Open Law, which he paid for, and through Claude and ChatGPT, for suggestions on arguments and additional cases. Then he cross-checked the citations and Shepardized every case. What he never did was verify that a pin cite pointed at the right page or that a quotation was real. Seventeen hallucinations across two motions, including fabricated quotes attributed to Ashcroft v. Iqbal, Landscape Forms, and Yurman Design. He pays $3,500 by Friday and goes to the Grievance Committee under Local Rule 83(c)(2). The sanction that will actually change how he works is the standing one: every authority he cites in any future filing in the case, including the corrected motion to dismiss, has to carry a hyperlink to the corresponding Westlaw or Lexis entry. From the order: “Rule 11 imposes an obligation on attorneys to, at a minimum, read the cases they cite to the Court.”

Two days later, Magistrate Judge Luke Evans publicly reprimanded Louis Ringger III in the Middle District of Tennessee and assessed $1,500. Evans also wrote the cleanest description of this problem I have seen from the bench. Artificial intelligence, he said, “combines the illusion of thorough review and analysis with the temptations of speed and low cost.”

The read: Four days before Barteca, the Connecticut Supreme Court sanctioned a lawyer who researched on Lexis, Shepardized, and then pasted the verified draft into ChatGPT to polish it. Same seam both times. A citator tells you whether a case is still good law. It does not tell you whether the page you cited says what you wrote.

Harvey added $100 million in a quarter. Nobody can say what it bought.

Harvey named Steve Zad chief revenue officer on August 4, and the number was sitting in the announcement: Q2 “marked a milestone quarter for Harvey, with the company adding more than $100 million in ARR.” Zad comes from Rubrik, where he helped take the business from $50 million to $1.5 billion in ARR and through its 2024 IPO. Three days later The Information reported Harvey in talks to raise at least $500 million at a $15.5 billion valuation, up from $11 billion in March. Those are talks, not a signed round. Then on August 6, BARBRI published what ten leaders across nine firms told it in interviews. Firms grade their own technology rollouts a C. And this: “Firms know who has activated AI tools, but almost none of them know who has actually changed the way they work.” The billable hour came back as the primary barrier to adoption, again.

The read: Seat activation is the number your vendor reports to you, and at most firms it is the only number anybody has. When renewal comes and a partner asks what the spend bought, a login count is not an answer.

STILL WATCHING

  • Thomson Reuters v. ROSS Intelligence was argued in the Third Circuit on June 11 before Judges Restrepo, Montgomery-Reeves and Bove. Nine weeks, no opinion.

  • The Thomson model ships inside CoCounsel Legal this month. Ten days after Thomson Reuters published its benchmark table, nobody outside the company has verified it.

  • In Cole v. Hobby Town Unlimited (C.D. Ill.), the response to Judge Darrow's show-cause order came due Friday.

  • The hallucination tracker sits at 1,868 as of its August 8 update, 1,297 of them in the United States. Attorneys account for 724 and pro se filers 1,096. The attorney number moved by seven since Wednesday's issue.

  • The notetaker cases. The motion to dismiss in the consolidated Otter.ai privacy litigation has been under submission since May 20. Chamberlain v. Granola drew Judge Edward Chen in the Northern District of California.

QUICK HITS

  • A capable agentic model now runs on one consumer GPU. Meta released Muse Glimmer this morning: 30 billion parameters, Apache 2.0, free on Hugging Face, distilled from its larger Muse Spark model and, in Meta's words, “small enough to run on a Mac or PC with a single consumer GPU.” Every confidentiality objection that begins with “we can't send client data to a third party” now has a version of its answer sitting on the desk. Mark Zuckerberg, in an essay published the same day: “Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.”

  • Y Combinator's summer batch includes a law firm, not a tool. Erinys describes itself as “building the first AI-native plaintiff-side litigation law firm network.” Also in the batch: Perceptron ML, selling private models trained on a firm's own matters, claiming a grounding engine that “verifies every fact against a primary source before a model can cite it.” Both are company claims and neither is something you can buy yet. One of them intends to compete with you rather than sell to you.

  • Anthropic hired its first Head of Claude for Legal. Robert Mahari holds a joint JD-PhD in legal AI from MIT and Harvard Law and founded Akiva AI. Within three days the complaint surfaced that his law firm experience amounts to about four months of internships. Richard Tromans at Artificial Lawyer defended the hire: “do you need to be a partner at a top law firm to do those things? Or to manage the execution of those things? No, not at all.”

The credential your overnight job uses to sign into a records portal belongs to a person. Find out which one.

See you in the next one.