Your Next Customer Is an AI Agent

Imagine a customer entering your online store - comparing prices, reading reviews, checking availability, and finally completing the purchase. Except it isn't a human at all. It's an AI agent acting on a person's behalf, navigating through your catalog in a split second and deciding independently what goes into the shopping cart. On August 4, 2026, precisely this scenario got a solid legal foundation: the US Court of Appeals for the Ninth Circuit in San Francisco, in the case Amazon.com Services LLC v. Perplexity AI, Inc. (No. 26-1444), lifted a preliminary injunction that had barred Perplexity's AI browser "Comet" from accessing Amazon. For online retailers worldwide - including the German Mittelstand - this is more than a footnote from the US legal system. It's an early signal of where e-commerce is heading over the next few years.

The ruling in brief: what actually happened

To understand the significance of the decision, it helps to look at the backstory. In March 2026, US District Judge Maxine Chesney in the Northern District of California had initially sided with Amazon: she issued a preliminary injunction barring Perplexity's Comet browser from Amazon. The reasoning at the time: Comet accessed Amazon's servers "without authorization" - a violation of the Computer Fraud and Abuse Act (CFAA), the central US federal law against unauthorized computer access, as well as its California counterpart, the Computer Data Access and Fraud Act (CDAFA).

Perplexity appealed - and on August 4, 2026, the Ninth Circuit ruled in its favor. The appeals court overturned the injunction with a notable line of reasoning: legally speaking, it isn't Perplexity's servers accessing Amazon's servers, but the users themselves. Comet, the court found, merely receives screenshots generated by the respective user's own browser and forwards instructions - without any direct server-to-server communication taking place between Perplexity and Amazon. Since the CFAA and CDAFA define "access" as an action a person takes against a computer system, the court held that this requirement was not met by Perplexity itself.

According to Reuters and several law firms specializing in technology law, such as Cooley and Wilson Sonsini, this is the first ruling by a US federal court of appeals ever on the question of whether AI agents acting on behalf of users may legally access third-party online platforms. A precedent, then - even if it formally applies only in the US.

Why the court ruled this way - and why it isn't a blanket permission slip

Important for context: the ruling is not a blanket authorization for AI agents to move freely across any platform they like. Amazon had argued, among other things, that Comet harmed the customer experience and posed cybersecurity risks. At this stage of the proceedings, however, the court found these arguments insufficiently weighty - and pointed to two counterarguments as well: a potential restriction of consumer choice, and a chilling effect on new AI technologies if every platform could use an anti-hacking law to decide which software its own users are allowed to run.

The core idea of the decision can be summed up like this: a law against unauthorized computer access is the wrong tool for regulating which software a user runs on their own behalf on the internet. Anyone who wants to control agent access needs to do so through contract law - not anti-hacking law.

That is precisely the key takeaway from law firms Cooley and Wilson Sonsini in their assessments of the ruling: platform operators will increasingly need to rely on their terms of service to regulate or restrict AI agent access, rather than relying on the CFAA and CDAFA. The case is also not finally settled: it returns to the trial court, and Amazon, according to German reporting by Handelsblatt and etailment, has already announced further legal steps. The story, in other words, is not yet fully told.

Does this apply to your store in Germany too?

Precision is called for here, because a shortened headline like "Amazon loses to AI agents" easily invites the wrong conclusions. The CFAA and CDAFA are US federal and state laws, respectively. They do not bind German and European online retailers directly. So if a Comet user or another AI shopping agent tries to shop in your store tomorrow, this US ruling does not legally intervene directly in your relationship with that agent. For German cases, the more relevant frameworks would be the UWG (Act Against Unfair Competition), general contract law, the GDPR - for example on the question of how an agent accesses a user's account on their behalf - and, going forward, the EU AI Act. No specific German court rulings on this question exist so far.

Still, it would be a mistake to dismiss the ruling as a purely American curiosity. It shows, in a representative way, how courts worldwide might think about this question going forward - and it offers a very concrete, practical insight that carries over directly to German stores: platform operators who want to retain control over AI agents cannot rely on an anti-hacking law, but must regulate this explicitly by contract - that is, in their own terms and conditions. That holds in the US, and it holds in principle similarly in Germany, where access rules for automated systems are likewise primarily enforced through contract law. In parallel, new EU AI labeling obligations have applied since early August 2026, along with expanded enforcement powers for the EU Commission against providers of so-called general-purpose AI systems - an additional regulatory framework that complements your compliance considerations without replacing the fundamental question of platform access.

The real point: your next customer really could be an agent

Setting the legal question aside, this case carries a much bigger message for your business: agentic commerce is no longer a distant vision of the future, but a court case with a real case name and a real docket number. Comet functions as an AI-powered browser and personal shopping assistant. Users give commands like "find the cheapest suitable monitor and order it" - and the agent navigates independently, selects products, and completes purchases. Classic touchpoints such as your product search, your ad placements, and your carefully curated recommendations are partially or entirely bypassed in the process. The agent doesn't read your beautiful homepage banner. It reads your product data.

In parallel, the payment infrastructure that enables exactly this kind of behavior at scale is being built right now. Several industry standards for agentic commerce emerged in 2025 and 2026: OpenAI and Stripe's Agentic Commerce Protocol with a so-called Shared Payment Token, Visa's Trusted Agent Protocol, which has run jointly with Cloudflare since October 2025, Mastercard's Agent Pay, announced in April 2025, with so-called Agentic Tokens, and Google's open protocol AP2, active since September 2025 with more than 60 launch partners. These standards create the technical foundation for agents to increasingly pay independently - without a human re-entering credit card details for every single purchase.

How quickly this will affect actual revenue can currently only be roughly estimated, and the available figures come predominantly from market research and SEO industry sources, not from official statistics. Mordor Intelligence puts the global market for agentic AI in retail and e-commerce at around $46.74 billion in 2025, with an estimate of $60.43 billion for 2026 and a forecast of $218.37 billion by 2031. Other industry reports say that currently only around 3 percent of all online transactions run through an AI agent, while at the same time around 89 percent of surveyed retailers are already actively preparing for agentic commerce, according to surveys. Figures on consumer usage also vary considerably by source - from 28 to 35 percent already actively using AI shopping assistants in individual studies, to forecasts of nearly 50 percent of all online shoppers by 2030. You should treat these orders of magnitude with appropriate caution - as a trend indicator, not as a robust metric for your own planning.

What you as a Mittelstand business can do right now

Even though the ruling triggers no direct legal obligation for you, it does open up several very practical areas for action. First: explicitly regulate AI agent access in your terms and conditions, if you want to limit or steer it. Don't rely on "surely they're not allowed to do that" holding up legally in a real dispute - that was exactly Amazon's misjudgment in this case. Contract law instead of anti-hacking law is the lesson from the US law firms, and this blueprint works for German store operators too.

Second, and at least as important in practice: make yourself visible to agents instead of hiding from them. Since agents like Comet partially bypass the classic customer journey and compare prices, availability, and product features directly, you should consistently maintain structured product data - Schema.org or JSON-LD markup with price, availability, GTIN, reviews, and clear unique selling points. Industry sources report that pages with comprehensive schema markup are cited noticeably more often in AI answers - a figure that comes from SEO vendor sources and is not independently verified, but the general direction holds: what an agent cannot read in a properly structured way, it also cannot recommend or buy.

Third, a technical check of your product catalogs is worthwhile. Server-side rendering, or complete HTML for product pages, makes it considerably easier for AI crawlers and agents to capture product data correctly - a point equally relevant for Shopware, Shopify, and custom shop systems across the German Mittelstand. If your catalog only becomes readable after full JavaScript rendering, you risk being simply invisible to agents - regardless of how good your prices or service actually are.

Fourth: keep an eye on the payment side. With protocols like Visa's Trusted Agent Protocol, Mastercard's Agent Pay, Google's AP2, and OpenAI and Stripe's Agentic Commerce Protocol, a new payment infrastructure for autonomous purchases is emerging right now. In the medium term, it's conceivable that European payment providers will offer comparable mechanisms via PSD2 or future PSD3 interfaces as well. For choosing your payment provider, this can already be a criterion worth keeping in mind today.

Fifth, and this concerns more the process side: check whether and how you can even recognize when a purchase is made by an AI agent rather than a human. That's relevant for fraud prevention, for handling returns, and for your customer service - and it fits with the EU rules in force since early August 2026, under which AI systems must identify themselves to users. The question of whether and how a store itself detects that it's dealing with an agent will likely gain importance over the coming months.

Sixth, and this is the crucial caveat: it would be wrong to derive a finished, all-time blueprint from this one ruling. The Ninth Circuit's decision concerns a preliminary injunction, not a final, legally binding decision on the merits. The case goes back to the lower court, and Amazon has already signaled it intends to continue the fight. Whether and how other US courts, European courts, or lawmakers will assess the question going forward remains open. There is also no official statement yet from Amazon Germany or German trade associations such as the HDE or the bevh on this specific ruling - German coverage in Handelsblatt, etailment, and Lebensmittelzeitung has so far mainly referenced the US sources.

For you as a store operator, that means: act with foresight, but avoid panic decisions based on a single, not-yet-final US court ruling. The underlying direction - AI agents are becoming a new, relevant customer group, and the rules of the game for that are only now being written - remains unaffected by this, and is the real reason to become active now.

Conclusion: a precedent, not a free pass - but a clear signal

The Ninth Circuit's ruling of August 4, 2026 is the first decision by a US federal court of appeals on the question of whether AI agents acting on behalf of users may legally access third-party platforms. It doesn't bind German online retailers directly, but it delivers a clear message: whoever wants to control AI agent access to their own platform needs a solid contractual basis for that, not anti-hacking arguments. And whoever wants to hold their own in the competition for customers who increasingly consist of software rather than people can no longer avoid structured, technically accessible product data. Your next customer no longer has to be a human clicking through your store. It can be an agent, deciding on a human's behalf in seconds - and the question is no longer whether that's coming, but how well you're already prepared for it today.