When AI Commerce Gets Terms and Conditions
The most revealing document in AI commerce this week was not a product announcement.
It was a set of legal terms.
On September 28, Stripe updated its terms for agentic commerce sellers. They specify who remains the merchant of record, who handles fraud, who fulfils the order, who manages returns, who must keep product information accurate, and what an AI commerce interface is allowed to control.
Elsewhere, six global banks published common principles for agentic commerce. Merchants organised around their own positions on liability, fees and data. In the UK, a real AI-assisted bank payment was completed under regulatory supervision.
Taken together, these developments point to a less glamorous but more consequential phase of AI commerce.
The industry is moving beyond proving that an agent can buy.
It is beginning to decide what happens when that purchase becomes a real commercial obligation.
Stripe Is Turning Agentic Commerce Into a Defined Commercial Relationship
On September 28, Stripe updated its service terms with a dedicated section for Agentic Commerce Seller Services.
The language matters because it makes several responsibilities explicit.
The seller remains the merchant of record. The seller is responsible for fulfilling orders, receipts, customer service, warranties, returns, refunds and disputes. Merchants also retain responsibility for transaction fraud and for ensuring the product information supplied into agentic interfaces is accurate, current and not misleading. Stripe's terms also state that an agentic interface provider can decide whether a merchant appears, whether its products are surfaced or recommended, and whether a transaction is enabled.
That is a much clearer commercial architecture than “an AI helped someone shop.”
Imagine an AI interface recommends a coffee machine, initiates checkout and the customer later discovers that the listed delivery date was wrong. Several systems participated in the sale, but participation does not make responsibility disappear. Stripe's terms explicitly leave core seller obligations with the merchant.
The significance is not Stripe's particular allocation of responsibility. It is that agentic commerce now needs an allocation at all.
Once an AI interface sits between buyer and seller, commerce acquires another participant with influence over discovery and execution. Contracts must define where that participant's role begins and ends.
Mental Model — Commercial Chain: AI can insert another link between customer and merchant without removing the legal and operational obligations already attached to the sale.
Adding an AI intermediary does not make the underlying commercial responsibilities autonomous.
Source: Stripe Agentic Commerce Seller Terms
Six Banks Are Trying to Establish Rules Before Volume Arrives
On September 22, ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest jointly published principles for trusted agentic commerce.
The five areas are revealing: transparency, safety, privacy and data, choice, and interoperability. The banks also identified identity, authorization, fraud prevention, liability management and customer protection as issues that need consistent treatment.
Banks rarely need to agree on principles for software that merely recommends products. They become much more interested when that software can influence how money moves.
Consider an agent instructed to renew a household service at the best available price. It could choose the provider, select a payment method and initiate the transaction. If the agent chooses a less protected payment route or acts outside the customer's intended mandate, the problem is no longer simply whether the AI made a poor recommendation. Someone must determine what protection applies and where the customer goes for recourse.
That explains why these principles arrive before agentic payments reach mass adoption.
Financial infrastructure works because participants understand the rules surrounding a transaction, not merely the technical method used to send money.
Agents disturb those assumptions by inserting automated decision-making between intent and payment.
Mental Model — Rules Before Scale: Payment systems scale when participants share expectations about control, protection and responsibility. Agentic commerce needs those expectations before autonomous transactions become routine.
A payment initiated correctly by software can still create an unclear commercial outcome if nobody knows which protections apply.
Source: Trusted Agentic Commerce Principles
Merchants Are Starting to Negotiate Their Side of Agentic Commerce
On September 24, the Merchant Advisory Group detailed work from its Agentic Commerce Working Group, which was created to establish a common merchant position on the emerging ecosystem.
Its concerns go well beyond whether AI generates more sales.
The group is examining liability allocation, fraud, fee stacking, merchant payment-routing rights, data ownership, agent identification and governance standards. Its position paper sets out seven foundational conditions for merchant participation.
That is an important change in the conversation.
An AI agent might find a product, route a shopper through an external interface and choose how the transaction reaches the merchant. Each additional participant can create economic consequences: another fee, another data relationship, another fraud surface or another party involved when something goes wrong.
Imagine a retailer receiving an agent-initiated order that travelled through an AI platform, a commerce interface and a payment provider. The retailer may get the sale, but still needs to know who owns the customer data, whether it can choose payment routing, what fees were added and who carries liability for an unauthorized transaction.
Agentic commerce cannot mature solely around consumer convenience.
Merchants are now asking what participation costs and what rights they retain.
Mental Model — Transaction Constitution: A new commerce ecosystem needs more than technical standards. Participants also need agreed rights over money, data, risk and control.
Merchant participation cannot be assumed simply because an AI agent can technically reach the checkout.
Source: Merchant Voice in Agentic Commerce
The UK Put an Agentic Payment Inside Regulatory Supervision
On September 22, GoCardless processed what it describes as the UK's first agentic account-to-account payment, developed through the Financial Conduct Authority's AI Live Testing programme.
The transaction was a recurring donation to Trussell. A donor interacted with an AI agent, selected a monthly amount and entered bank details so the agent could establish a Direct Debit mandate. The entire journey took place through the conversation.
The interesting part is not that AI can collect a payment.
It is that a real transaction was designed around payer control while being tested alongside a regulator.
Imagine replacing the charity donation with an electricity bill. A consumer might tell an agent to find a cheaper provider and switch automatically if savings exceed a certain amount. That combines recommendation, consent, recurring payment and an ongoing commercial relationship.
A demo can prove that workflow works.
A regulated market needs to determine whether consent remains clear, whether the mandate is valid, what evidence exists afterward and what happens if the customer disputes the action.
Agentic payments are therefore beginning to move from technology experiments into supervised financial processes.
That is where the harder work starts.
Mental Model — Regulated Autonomy: An agent can automate execution while the surrounding system preserves human consent, evidence and recourse.
A successful autonomous payment is not commercially complete until it can survive a dispute.
Source: UK Agentic Account-to-Account Payment
The System That Is Emerging
The next infrastructure layer in AI commerce is not another protocol for helping machines buy.
It is a commercial responsibility layer.
Traditional ecommerce hides a remarkable amount of institutional machinery behind a simple Buy button. Merchants know who fulfils the order. Payment networks define transaction rules. Banks provide protections. Contracts allocate risk. Regulators define obligations. Dispute systems exist when the transaction fails.
AI agents introduce a new participant without automatically rewriting any of that machinery.
That is why this week's developments fit together.
Stripe is documenting the responsibilities of sellers and agentic interfaces. Banks are defining common principles around protection and control. Merchants are organising around liability, economics and data rights. GoCardless is testing an agentic transaction inside a regulatory environment.
The shift is from technical permission to institutional legitimacy.
Power will not sit only with the company that builds the smartest shopping agent. It will also sit with whoever defines the contractual and operational rules through which that agent participates.
For business leaders, this creates a different readiness question.
Do not ask only whether your systems can accept an agentic transaction.
Ask who remains merchant of record, who owns the customer relationship, who chooses the payment route, which party handles fraud, what evidence proves consent, who resolves a dispute, and which system is responsible when the agent acts incorrectly.
Those questions sound administrative.
They are actually the architecture of the market.
Core Truth
Agentic commerce becomes real when responsibility can move through the transaction as clearly as the money does.
