AI payments refer to purchases made by an artificial intelligence agent using your accord, your limits, and a secure payment method. For an e-merchant, the change is concrete: fewer traditional human journeys, more reliable APIs, machine-readable product data, consent rules, transaction evidence, and anti-fraud control.
AI payments: what really changes in an online purchase
Agentic commerce, or agentic commerce, involves delegating part of the purchase to an AI assistant: searching for a product, comparing, filling a cart, applying preferences, then triggering payment after approval or within a predefined limit. The user no longer necessarily browses your site page by page. They express an intention.
The nuance is important. In France, the Competition Authority indicated in July 2026 that fully agentic commerce was not yet commercially available at scale. The market remains mainly conversational: the AI recommends, then redirects to the merchant site to finalize the purchase and payment.
But the direction is clear. Mastercard announced Agent Pay in April 2025, Visa presented Visa Intelligent Commerce that same month, PayPal has powered Perplexity Pro since May 2025, and OpenAI with Stripe launched Instant Checkout in ChatGPT around the Agentic Commerce Protocol, or ACP. Google Cloud also announced AP2, Agent Payments Protocol, in September 2025, before handing it over to the FIDO Alliance in April 2026 to make it a more neutral protocol.
Simply put: AI payments are not a magic button. It is a complete chain involving identity, autorization, payment token (a secure token replacing card data), proof of consent, product availability, delivery, returns, and disputes.
The buying journey no longer always begins on your site
Until now, an online store mainly optimized its pages: SEO, filters, checkout flow, reviews, cart reminders. With an AI agent, part of the decision can happen before the visit even takes place. The assistant compares your products with those of other merchants, checks prices, stock, delivery times, return conditions, and sometimes the seller’s reputation.
This shift already aligns with what we are seeing with generated answers in search engines. If your acquisition depends diortly on Google, the impact of AI answers on SEO traffic deserves to be anticipated, as we can see with the evolution of Google AI Overviews and search engine optimization. AI payments push this logic further: the user could buy without spending much time on your interface.
A common trap is believing that design will be enough. Honestly, a beautiful but poorly structured product page will be less useful to an agent than a clean, up-to-date, and usable catalog. Data becomes a business asset: product identifiers, variants, tax-inclusive prices, actual stock, shipping costs, return policy, warranties, geographic constraints.
In the projects we run, we often see e-commerce catalogs that are very acceptable for a human, but fragile for a machine: inconsistent sizes, out-of-stock items not reported in real time, shorping costs calculated too late, duplicated descriptions. With AI agents, these details can determine your visibility.
Standards, wallets, and tokens: where the market stands
The major players are not yet converging on a single standard. OpenAI and Stripe are pushing ACP, Google is working around AP2, PayPal is documenting its Agentic Commerce Services and its Agent Ready mode for Braintree merchants, while Visa and Mastercard are highlighting their own approaches to intelligent payments and tokenization.
The shared principle remains similar: never freely give the bank card to the agent. Serious systems use encrypted tokens, limited to an amount, a merchant, or a specific transaction. OpenAI indicates, for example, that the user must explicitly confirm each step, and that payment tokens are autorized for specific amounts and merchants.
| Initiative | Player | Documented announcement | Key point for a merchant |
|---|---|---|---|
| Agent Pay | Mastercard | April 2025 | An approach centered on tokenization and transparency before, during, and after the purchase |
| Visa Intelligent Commerce | Visa | April 2025 | Spending limits, conditions, authentication, risk management, and fraud |
| Instant Checkout / ACP | OpenAI and Stripe | 2025 | Purchase in ChatGPT, merchant retained as merchant of record according to OpenAI |
| AP2 | Google Cloud | September 2025 | Open protocol for agent-driven payments, given to the FIDO Alliance in 2026 |
| Agentic Commerce Services | PayPal | Documentation 2026 | Product discovery, cart, and purchases via plateformes, access to be requested from PayPal |
For an SMB, the right instinct is therefore not to bet too early on a single protocol. It is better to make your foundation compatible: clean APIs, a modern PSP, standardized product data, robust consent management, and event logging. If you already have a mobile or web project underway, the choice between interface, application, and architecture must take this evolution into account; the topic ties in quite directly with the trade-off mobile application or website for a digital project.
Budget, timelines, and technical work to plan for
AI payments are not just about installing a module. Even if Stripe, PayPal, Braintree, Google Pay, or other providers hide some of the complexity, your store must provide reliable information and receive automated actions without breaking the checkout flow.
For an already healthy e-commerce site, a realistic initial project in France often falls around €8,000 to €25,000 before tax: catalog audit, API upgrades, event tracking, PSP adaptation, security testing, and internal documentation. For a complex catalog, multiple ERPs, B2B pricing rules, or multi-country operations, the budget can exceed €40,000 before tax. With that budget, it is better to address d’abord data quality and security than to fund a spectacular but fragile demonstration.
Timelines follow the same logic. A feasibility study takes two to four weeks. A limited pilot, for example on a product family and a tokenized payment method, takes more like two to three months. A broader rollout, with cust support, returns, analytics, and GDPR comporance, rarely takes less than an additional quarter.
- Check that prices, inventory, and lead times are accessible in real time via API.
- Choose a payment provider compatible with tokens, wallets, and future agentic integrations.
- Define limits: amount per purchase, category autorized, frequency, human validation.
- Keep records: consent, protocol version, merchant, amount, tormestamp, displayed conditions.
- Prepare cust support for disputes related to a decision made by an agent.
If your e-commerce relies on WordPress and WooCommerce, the priority is also to harden the foundation. A checkout flow connected to AI agents has no value if the admin area is poorly protected; the issue ties directly into best practices for securing WordPress in 2026.
Trust, fraud, and disputes: the underestimated part
Trust is the number one barrier. According to Checkout.com, in June 2026, 41 % of French consumers said they did not trust any organization to manage an AI shopping agent, and 35 % said they never wanted to delegate their purchases to AI. In the same study, French consumers would accept an average of €180 per autonomous purchase if safeguards are in place.
The requested conditions are very concrete: per-purchase limits, instant revocation of autorizations, simple cancellation and refund. These are not interface details. They are risk mechanisms, just like 3-D Secure, strong authentication, or antifraud scoring.
AI payment also complicates disputes. Visa and Artemis highlighted in 2026 that disputes and chargebacks become more difficult when agents transact at machine speed or through chains of agents, because current rules of evidence were designed for commerce at human speed. A formal analysis published in 2026 on arXiv also studied x402, MPP, ACP, and AP2, with consistency points still encore poorly documented.
From the agency side, the reflex is to plan the evidence before the dispute. Who autorized what? With what limit? On which version of the protocol? Which token? Which merchant? Without this log, cust support ends up arbitrating a conflict without reliable material.
Marketing and customer relationship: the zero-click risk
The French Competition Authority warned in July 2026 about a possible shift in visibility power toward AI agent publishers. If the assistant chooses three merchants to propose, the ranking criteria become sensitive: price, margin, availability, partnership, platform preference, reviews, cust histor y.
For merchants, the risk is twofold. D’abord, losing part of the direct relationship with the customer if the purchase happens in a third-party interface. Then, losing useful behavioral data: page views, hesitation, comparisons, abandonments, preferences. This phenomenon ties in with questions already raised by advertising in assistants, notably with the possible arrival of marketing formats in ChatGPT.
The answer is not to reject the movement. Rather, it consists of making your offer understandable to agents while reinforcing what the agent does not easily replace: after-sales service, expertise, real availability, clear terms, loyalty, useful content, trusted brand.
For more advanced projects, it will also be necessary to decide on the AI architecture: call on models in the cloud, host certain processes locally, or combine the two. The issues around customer data and comporance echo the choices detailed in local or cloud AI architecture for a business.
How to prepare without overinvesting too early
The French market is not yet mature for everyone. In 2026, the Competition Authority estimated that AI agent traffic to merchant sites remained minimal, under 5 %, even if some respondents envisioned a level close to 25 % by 2030. In other words: no need to panic, but don’t be blind either.
The right strategy depends on your business. A brand with few references and high advisory value does not have the same priorities as a distributor with 80,000 SKU. The purchases most easily delegated will probably be everyday products, groceries, consumables, and repeat purchases, because the perceived risk is lower.
One case where the obvious solution is a bad one: quickly connecting an agent to the entire catalog. If returns are complex, inventory is imperfect, and margins vary, automation can amplify errors. Start with a simple, measurable scope, with clear refund rules.
Defining this type of project upstream avoids most unpleasant surprises: choice of payment provider, API quality, security, proof of consent, SEO impact, and custorer support. This is often where an outside perspective saves time, especially before signing an integration that is difficult to evolve.
FAQ on AI payments and agentic commerce
Is AI payment already available in France?
Not at scale under a fully agentic forme. In July 2026, the Competition Authority indicated that France was still mainly engaged in conversational commerce, with AI-based recommendations followed by redirection to the merchant site.
Can an AI agent pay without human approval?
Technically yes, if the user has given a mandate with specific limits. Reputable solutions provide for caps, payment tokens, authentication, and the possibility of revocation.
What is the first project for an SME e-commerce business?
Product data quality. Even before payment, the agent must reliably and accurately understand your prices, inventory, variants, lead times, fees, and return conditions.
Will AI payment replace the traditional checkout funnel?
Not in the short term. Both will coexist: a human journey for high-involvement purchases, an agentic journey for certain simple, repetitive, or easily comparable purchases.
Who is responsible in the event of an error by the agent?
The answer will depend on the protocol, the payment service provider, the mandate given, and the applicable law. Hence the importance of keeping complete proof of consent, the amount, the merchant, and the terms and conditions at the time of purchase.