Columns2026/07/29

What if the buyer wasn't human? ― Return policies in the age of AI agents that complete checkout

In the autumn of 2025, "Instant Checkout," a feature allowing direct product purchases within ChatGPT, launched in the United States. Google has AP2, and Shopify plans to introduce the Universal Commerce Protocol in 2026. Automated shopping experiences, where human intervention isn't required even for payment, are already a reality.

But what if the purchaser wasn't a human, but an AI? Would there be a difference in return policies between "human-initiated purchases" and "AI-initiated purchases"? This article explores this single question from three perspectives: legal frameworks, the implementation of major protocols, and the practical operations of e-commerce businesses.

What you’ll learn in this article:

→ That no legal framework exists as of July 2026 that invalidates return policies for AI purchases.
→ That terms and conditions stating "no returns for AI purchases" are more likely to be invalidated by mandatory laws.
→ Where the true divergence between human and AI purchases lies.
→ What Shopify Plus merchants should prepare for in their terms and payment flows.

Premise: There is no inherent "unconditional return" for online shopping

In Japanese e-commerce (mail order), there is no cooling-off period similar to door-to-door sales. The Consumer Affairs Agency explicitly states that "there is no unconditional cancellation (cooling-off) system for mail order."

Whether an item can be returned depends on the return policy (special return agreement) set by each site. If there is no special agreement stated in the advertisement, items can be returned within 8 days of product arrival, with the return shipping cost borne by the purchaser (Article 15-3 of the Specified Commercial Transactions Act). If "no returns" is clearly and visibly displayed, then no returns are allowed accordingly.

In essence, even for human purchases, there is no inherent unconditional right to return. Therefore, a simple distinction like "humans could return, but AI cannot" is inherently difficult to draw under the Japanese system.

Implementation Reality: AI purchases still have the merchant as the "seller"

Next, let's examine the protocols currently in operation. This also supports the idea that "there is no difference."

OpenAI and Stripe's ACP (Agentic Commerce Protocol) initiates payments using a token linked to the cart total for each merchant, and orders are passed to the merchant's backend. Approval, charging, fulfillment, and returns are all designed to be handled by the merchant, as before.
Shopify's agent checkout also explicitly states that "merchants will always remain the merchant of record." Returns and refunds operate under existing policies.

AI agents are merely intermediaries that process orders; they do not change the counterparty to the contract. As of July 2026, no mechanism exists that automatically invalidates return policies simply because an AI made the purchase.

Can AI even bypass 3D Secure?

I mentioned "completing the payment," but many people would stumble here. The one-time code sent to email or SMS for 3D Secure. If AI cannot enter that code, the purchase cannot be completed in the first place.

The actual design does not involve AI "bypassing" that number. It's designed to avoid the one-time code from the outset. There are two main mechanisms:

Avoid repeated code entry with tokens and pre-authorization. Visa Intelligent Commerce and Mastercard's Agentic Token issue network tokens scoped specifically for agents, rather than actual card numbers. Once a human completes identity verification (Visa uses passkeys) the first time, subsequent transactions use that token and a Merchant Initiated Transaction flag to lean towards frictionless (additional authentication-free) processing. For everyday small purchases, the code entry screen itself is designed not to appear.
If a challenge appears, revert to human intervention. If a card company requests a step-up (additional authentication) for conditions like high transaction amounts or unregistered categories, AI cannot solve CAPTCHAs, biometrics, or email codes on its own. This leads to escalation to human intervention. In the ACP specification, it's being discussed (GitHub issue #55) that 3DS challenges would be displayed in an iframe on the agent's screen, and humans would enter the code or biometric authentication, with the agent then relaying only the result.

In other words, the scenario of "AI automatically reading and bypassing email numbers" deviates from standard design. What actually happens is that AI completes payments within the scope of the permission initially granted by a human. As long as there are authentication barriers, purchases with "zero human involvement" are still closer to the exception.

This point also connects to the issue of returns. The clearer the initial human authentication remains, the easier it is to explain the order as reflecting the user's intent. Conversely, if a high-value transaction bypasses step-up, the room to argue "it was an unintended purchase" widens later on.

Whose "intent" is an order placed by AI anyway?

Before delving into the differences, let's establish the starting point of the contract. Under Japanese civil law, only persons (natural persons and legal entities) have legal capacity. An AI agent itself cannot be an "agent" under Article 99 of the Civil Code.

Thus, the prevailing academic and practical view positions AI as a "tool for expressing intent" or an "assistant in performance." It is understood that when a user activates an AI and deploys it for a transaction, there is a comprehensive intent to accept the AI's output as their own declaration of intent. Therefore, an order placed by AI is validly established as the user's own declaration of intent. The US UETA (Uniform Electronic Transactions Act) also explicitly states that contracts made by "electronic agents" are binding on the user, taking a similar approach.

An AI's order is attributed to the individual, and the contract is valid. The remaining two questions are: "How to cancel if there's a mistake?" and "How much can be written into the terms and conditions?"

Points of Divergence

Can an "AI purchase error" be cancelled?

For example, consider an incident like this: A user asks, "Find the cheapest flight ticket for a conference in Paris, Texas." However, the AI mistakes "Paris" for Paris, France, and automatically completes a non-refundable, first-class payment of 2 million yen.

Can this be canceled under Article 95 of the Civil Code for mistake (error) or the Electronic Consumer Contract Act's relief for operational errors? The Electronic Consumer Contract Act allows cancellation due to mistake, regardless of the consumer's gross negligence, if the business has not provided a confirmation screen for the application details. However, AI placing orders directly via API typically bypasses human-facing confirmation screens. Whether cancellation can be claimed under this structure is a point of contention among experts. Furthermore, if the act of "entrusting everything to AI itself is gross negligence," cancellation becomes even more difficult. A definitive interpretation is not yet established.

Will "no returns for AI purchases" be enforceable?

There's also the opposite concern. Can an e-commerce business write into its terms and conditions, "All orders placed via AI or automated bots are deemed B2B (business-to-business transactions), and consumer return rights do not apply," and refuse returns?

Here, mandatory provisions of law come into play. According to expert interpretations, as long as an individual is purchasing for personal use, the transaction does not automatically change from B2C to B2B simply because the purchase method involves AI. Provisions that completely exclude consumer protection based solely on the access method are highly likely to be judged invalid under Article 10 of the Consumer Contract Act as "unilaterally harming the interests of consumers." In other words, the phrase "no returns for AI purchases" will not serve as much of a shield as businesses might hope.

When bulk, automated purchasing constitutes a "business purpose"

The exception is when the substance, not just the means, changes. If AI is used for bulk purchasing and automated procurement (reselling) to sweep up inventory, it could be judged as a "purchase for business purposes." Consumer protection safeguards personal consumption, so if a purchase is deemed commercial, it falls outside the scope of flexible return policies. A single item purchased by a person and hundreds of items automatically ordered by an AI would be treated differently on the same site. What matters more is whether the "substance of the purchase is B2C," rather than "whether it's AI or not."

Are terms and payment flows preemptive?

Overseas, preparations are underway. In March 2026, Target revised its terms to treat transactions by authorized AI agents as "transactions authorized by you," placing responsibility for all actions, including returns, on the user. Mastercard's Agent Pay incorporates Human-in-the-Loop, requesting approval on smartphones for high-value or undefined category transactions, and Visa's Trusted Agent Protocol identifies legitimate agents with cryptographic keys. Having these preparations in place makes a significant difference in the initial response when problems arise.

Comparing human purchases with AI purchases that complete payment, here's how it looks:

Issue	Human Purchases	AI Purchases (Payment Completed)
Applicability of return policy	Depends on policy (no unconditional returns)	Same, depends on policy. No institutional invalidation
Attribution of order intent	Intent of the individual	Attributed to the individual under the "expression tool" theory. Contract is valid
Remedy for purchase errors	May be subject to operational error relief/mistake	Bypasses confirmation screens, applicability under discussion
"No returns for AI purchases" clause	—	Exclusion clause for B2B treatment risks invalidation (Consumer Contract Act Article 10)
Commercial determination	Consumer if for personal use	More likely to be deemed for business purposes with bulk/automated purchasing

As you can see, the return policy itself applies in both cases. What is uncertain is where AI purchases deviate from the implicit assumptions of the "buyer" that the policy was originally written for.

How Other Countries Respond — Japan, US, and EU

Key jurisdictions have different answers to the same question.

Consideration	Japan	United States	EU
Attribution of AI orders	Attributed to the principal under the "expression tool" theory / "performance assistant" theory	Binds the user as an electronic agent (UETA §14)	Requires human intent, but intent is recognized in the act of deployment
Cancellation of purchase errors	Applicability of mistake and Electronic Consumer Contract Act is under discussion	Statutory right of cancellation under UETA §10(2). Cannot be excluded by special agreement	14-day right of withdrawal fully applies, no need to prove mistake
Returns/Withdrawals	Depends on policy + 8 days under Specified Commercial Transactions Act	Depends on policy	14-day unconditional right of withdrawal

The US and EU stand out. UETA §10(2) in the US establishes an unconditional right to cancel automated transactions if the system does not have an error correction mechanism, and this cannot be excluded by terms of service. The EU's Consumer Rights Directive (CRD) 14-day right of withdrawal fully applies to AI purchases, allowing cancellation without proving mistake. Japan lacks such robust safety nets, with return policies and interpretation left to scholars and practitioners.

Our assessment — Setting up terms and payment flows preemptively

As supporters of Shopify Plus merchants, we view this issue not so much as one of consumer protection, but rather as a matter of designing terms and conditions and payment flows.

Returns in Japanese e-commerce are inherently policy-dependent, and the system doesn't automatically protect consumers. However, drawing a blanket line saying "no returns for AI purchases" could lead to the invalidation of that very clause by mandatory provisions of law. Therefore, the remaining option for businesses is not to exclude AI orders, but to decide in advance how to accept them. Our assessment is that stores that proactively implement the following four points will have a significantly different initial response duringトラブル (trouble) compared to those that do not:

Explicitly state in the terms and conditions that orders from authorized AI agents are considered "orders by the user themselves," along with provisions for return and chargeback liability (Target model).
For orders exceeding a certain amount, incorporate human confirmation by inserting a summary notification and temporary hold (cooling-off period) before completion (the philosophy of error correction required by the Electronic Contract Act and UETA §10(2)).
Provide machine-readable terms and conditions. The movement towards the Legal Context Protocol (led by the American Arbitration Association AAA-ICDR, etc.), which places governing law and cancellation provisions in /.well-known/legal-context.json, is continuous with Shopify's agents.md and /llms.txt.
Have a clear distinction at the order stage between bulk, automated AI purchases and individual single-item purchases.

In terms of identification, Shopify has enabled /agents.md and /llms.txt across all stores, and AP2 and ACP are moving towards leaving signed mandates (evidence of delegation). Agent orders that comply with the protocol can be audited later. The state of "not knowing whether an AI or a human made the purchase" is, at least for compliant traffic, moving towards resolution.

Four points that are still undetermined

To be honest, the judiciary has not yet provided definitive answers to these questions.

There is no established interpretation regarding whether AI malfunctions can be canceled under mistake or the operational error relief provisions of the Electronic Consumer Contract Act. No relevant court precedents have been found either.
The conclusion that "terms rejecting returns for AI-initiated orders by treating them as B2B are invalid" is an expert's assessment and not a definitive judgment by Japanese courts.
We could not identify any Japanese e-commerce terms and conditions that explicitly permit or disclaim liability for AI agent purchases (there are examples from Target and Shopify overseas).
The Legal Context Protocol is a newly introduced standard, and its widespread adoption is yet to come. Administrative bodies are also still in the guideline stage, such as AI promotion laws and METI checklists, and have not yet enacted laws on strict liability.

Even with these uncertainties, we believe it is safer for companies to preemptively establish their own terms and payment flows rather than waiting for regulations to solidify.

Return policies are effective. What shifts is the assumption of the "buyer."

It is not accurate to say that "return policies do not apply when AI makes a purchase." Return policies are effective even when AI makes a purchase, and a policy that states "no returns for AI purchases" could even be invalidated. What is actually called into question is how to remedy purchase errors, how far terms and conditions can go, and whether the substance of the purchase leans toward a business purpose. These are the key points.

Even if the buyer changes from a human to an AI, return policies remain effective. What is being re-evaluated is for whom those terms and conditions are written as the "buyer." In the next one to two years, stores that have fully incorporated AI agent orders into their terms and payment flows are likely to navigate potential return-related confusion more smoothly.

References and Sources

The information in this article is as of July 2026. Specifications related to agentic commerce, company terms, and standards are continuously updated and may change in the future. As of the time of writing, no laws, administrative interpretations, or court precedents in Japanese consumer law directly addressing AI purchases could be confirmed. Descriptions regarding legal interpretation are based on expert and industry organization assessments and do not constitute definitive judicial rulings.