تواصل معنا

Agentic Commerce: 3 Smart Steps Before AI Agents Buy

نونزيو جيوجليانو

11 نوفمبر 2026 • 10 دقائق للقراءة

Agentic commerce in a small shop: a hand holding a smartphone with blank product cards over a packing table

تحذير: بعض أجزاء المحتوى مترجمة تلقائياً وقد لا تكون دقيقة تماماً.

On the night of Sunday 20 September, Amazon started blocking Muse, the personal AI agent Meta had launched twelve days earlier, from shopping on Amazon.com. Amazon's complaint was that Meta had not told it Muse would shop there, that the agent did not identify itself while browsing and that it appeared to capture and store customer credentials. Meta, for its part, has said Muse “has no visibility into people's passwords or payment methods.” Since then, Muse's own shopping instructions have told it "Do not search Amazon."

OpenAI went the other way on 13 November and launched ChatGPT Wallet, which lets its agents pay at a merchant's existing checkout with a payment token instead of the customer's real card details. OpenAI's merchant documentation is plain about who carries the rest: “Your business handles the order, fulfillment, customer support, and fraud decisions.”

One store shut the door and one platform knocked on every door at once, and in both cases the decision sits with the merchant. That is agentic commerce seen from the seller's side this month. The platforms supply the rails, while the store still has to choose which agents get in, make its offer readable enough for a machine to act on and keep both the fraud decision and the proof.

still has to choose which agents get in, make its offer readable enough for a machine to act on and keep both the fraud decision and the proof.

What agentic commerce is, and what changed in November

Agentic commerce is buying and selling in which an AI agent searches, compares and completes a purchase on a person's behalf, within limits that person sets. Today the person still approves each payment, and with ChatGPT Wallet the user reviews the items, the total and the card before confirming with a security check such as Face ID or a fingerprint scan. It is the point where an AI agent stops answering and starts acting, the same line our guide to choosing an AI chatbot for business draws for customer service.

The definition is not new, but where the purchase happens has moved in the past year. OpenAI's first attempt, Instant Checkout, let people pay without leaving ChatGPT. In March 2026 the company changed course, writing that the first version "did not offer the level of flexibility that we aspire to provide." From then on it would let merchants "use their own checkout experiences" while it focused on product discovery.

ChatGPT Wallet, built in partnership with American Express, Mastercard and Visa, completes that turn by letting ChatGPT's agents, which OpenAI calls dots, pay on the merchant's own website. After the user approves, the payment network provides a payment token in place of the real card details, so neither the business nor the agent sees the card number. OpenAI's payments guide adds that Wallet “works out of the box with existing checkout pages and payment forms” and needs no Wallet-specific integration, which means the agent now arrives at your site, goes through your checkout and meets your fraud rules.

The launch is narrow, since Wallet is rolling out gradually to ChatGPT Pro subscribers in the US aged 18 and older with eligible US-issued cards, and OpenAI's standard product feed upload “currently targets the US.” For European stores, that is lead time rather than an exemption, because the work below is the same wherever the agent comes from, and Muse showed that agents do not wait for an official integration before they start shopping.

Several questions remain open, and we would not plan around guesses about them. OpenAI has not said whether or when Wallet will reach Europe, the announcement says nothing about merchant fees, no retailer has been named as accepting it, and nobody has yet settled how liability is split when an agent buys the wrong thing.

Let the right agents in, on purpose

Amazon's move against Muse was a business decision, and a legitimate one, because a store may refuse automated shoppers it cannot identify, and Amazon said publicly that such agents should “operate openly and respect service provider decisions.” The risk for most stores is the opposite case, an agent they would happily serve being turned away by a bot rule written years ago for scrapers, without anyone deciding anything.

A shop's glass door left half open, a reminder that letting AI agents in should be a deliberate decision

Telling the two apart is now possible because some agents sign their requests. ChatGPT uses Web Bot Auth, a protocol under development at the IETF and built on the HTTP Message Signatures standard, RFC 9421, to identify certain requests from its cloud browser. Each signed request carries a Signature-Agent header set to "https://chatgpt.com", and the public keys are published at chatgpt.com/.well-known/http-message-signatures-directory. According to OpenAI's agent access guide, Akamai, Cloudflare, HUMAN and Vercel verify these signatures automatically, and anyone else should verify the signature before applying an allowlist exception.

صياغة التميز في البرمجيات

دعنا نبني شيئاً استثنائياً معاً.
اعتمد على شركة Lasting Dynamics للحصول على جودة برمجيات لا مثيل لها.

اكتشف خدماتنا

A verified signature does not guarantee a successful visit, because rate limits, CAPTCHAs, headless-browser detection and custom firewall rules were written for a web where automated traffic was presumed hostile, and any of them can still stop a signed agent. OpenAI's announcement asks merchants to confirm their “general site security measures are not inadvertently blocking ChatGPT agents,” and in agentic commerce the firewall is the first salesperson an agent meets.

How often stores turn agents away is starting to be measured, and the first numbers come from a preprint published on arXiv in October 2026 that reports a pilot of a merchant-side readiness benchmark. The authors tested 555 merchants that two payment processors publicly name as clients, from a single location on 30 September. Neither of their two declared but unsigned agent profiles could read the home page at 105 sites (18.9%), counting every cause.

At three of the four large brands they assessed in full, in lodging, air travel and subscriptions, the declared profiles were refused at the offer page. The authors say the pilot shows the measurements can be taken but does not validate them or rank merchants, so read the numbers as a warning rather than a market statistic.

The agentic commerce policy worth writing down says which agents may browse, which may buy and which are blocked, and who reviews that list. Signed agents are a different case from agents that do not identify themselves, which was the core of Amazon's complaint about Muse, and a policy that treats them the same will either let too much in or keep too much out.

Make the offer readable enough for a machine to decide

An AI shopping agent buys only what it can verify, so when the total price, the delivery date or the return terms are vague, it cannot decide, and the sale goes to a store where it can. For a catalog team, agentic e-commerce therefore starts with product data rather than with a chatbot.

Products on a shelf with blank kraft price tags, the kind of offer details an AI shopping agent needs to read

OpenAI's product feed specification shows the minimum a machine expects, with one row per purchasable item or variant and required fields that run from item_id, title و description إلى price, التوافر, target_countries و store_country. Availability takes one of five values, from in_stock إلى unknown, and the price carries an ISO 4217 currency code, as in 79.99 USD. The spec notes that checkout has separate requirements, so the feed is a discovery tool, but it still makes a useful readiness test for agentic commerce, because it shows whether your catalog can describe itself without a human reading the page.

Most stores are still far from that, judging by the same benchmark. Among 330 merchants, nearly all retailers selling in the US and Canada, the all-in total with shipping and taxes could be determined at only 18. The check failed at 178, mostly because of tax. At the four large brands, the authors could not confirm that any buyer's instructions would be met, because each left the total or a refund term undetermined.

In practice, shipping cost, delivery date, taxes and return conditions need to exist as structured data the agent can read, not in a PDF, an image or a pop-up that appears after sign-in. Stock, too, has to follow the real inventory. None of this is new for a good product page, but a machine now notices when it is missing and leaves.

Stores that want to go further can expose actions as well as data. OpenAI's plugin guide describes a plugin as an MCP server plus skills and asks for product lookup and cart creation to work in guest mode, without sign-in, with a browser handoff URL (continue_url) whenever the plugin cannot complete a required step. The same guide notes that OpenAI is still working on how plugins and Wallet fit together. WebMCP, which lets a web page expose tools directly to agents, is in a Chrome origin trial and a W3C Community Group draft that is not on the standards track, so it is worth watching rather than building on yet.

ابتكار مستقبلك الرقمي

بدءاً من الفكرة إلى الإطلاق، نقوم بتصميم برامج قابلة للتطوير مصممة خصيصاً لتلبية احتياجات عملك.
شارك معنا لتسريع نموك.

تواصل معنا الآن

Keep the fraud decision, and the proof

OpenAI took the card number out of the agent's hands and left the fraud decision in yours. The merchant receives a network token instead, and the payments guide tells you to use the Payment Account Reference (PAR) to associate that token with the underlying account. It follows the same logic as the tokenization described in our piece on PCI DSS compliant software: the less real card data your systems hold, the less there is to protect.

The customer consents only to the specific transaction, so the credential must not be reused or saved for a separate purchase. Recurring charges are possible if your checkout supports recurring billing and the customer agrees to them.

Risk is where agentic payments differ most from the card payments your rules were trained on. OpenAI's documentation explains that risk systems tuned to detect bots rely on signals such as IP addresses and browser details, and warns that its cloud browser can set them off "by accident and cause purchases to fail." The Wallet announcement adds that they "may still flag legitimate purchases made by agents, leading to declines or cancellations." OpenAI is relaying user signals in the standard Forwarded HTTP header, with parameters that may include asn, لـ, ip-country, ip-prefix, locale, tz و user-agent, and its advice to merchants with their own risk systems is to “start observing the Forwarded header.”

The announcement names the fraud prevention partners OpenAI is working with, including Forter, Riskified, Sardine and Signifyd, and payment providers including Adyen, Checkout.com and Stripe, so the first call goes to whichever of them you already use.

In the European Union, the payments rulebook has not changed for agentic payments. A March 2026 analysis by Osborne Clarke concludes that agent-based payment models “remain subject to PSD2” and its rules on strong customer authentication, with no special regime. It expects disputes to “increasingly focus on whether the user authorised all elements of the transaction,” meaning the payee, the amount and the timing. The Payment Services Regulation meant to update those rules was still awaiting the European Parliament's second reading in November 2026, so PSD2 remains the rule to design for.

That is why the record matters as much as the rule. OpenAI says each Wallet approval "creates a clear record of the purchase the user authorized the agent to make," but that record lives on OpenAI's side. When a customer says they did not mean to buy something, the merchant needs its own. A useful log ties the purchase to the agent's verified signature, keeps a snapshot of the price, total, terms and stock the agent saw at that moment, and stores the approval or payment reference. The same log is your evidence in a chargeback and the data you need to tune risk rules written for people.

Buy the rails, build the integration

Nobody should build their own agent wallet or token scheme, because payments and tokens come from your payment provider, signature verification from your CDN or bot-management vendor and product feeds from your commerce platform. If your platform already exports feeds and your CDN already verifies signatures, much of what is sold as an agentic commerce platform is a set of settings you already pay for.

The work changes when the catalog lives in an ERP or PIM the platform cannot read, when prices and terms vary by customer or region, as with B2B contracts, or when the checkout and the risk rules are built in-house. In those cases, the integration is the real project. The feed has to be generated from the system of record, and the firewall needs a policy for signed agents. Availability and cart questions need a plugin or MCP server that can answer them, and the purchase log described above has to be wired into the order flow.

A store built on a modular stack, the approach our piece on composable commerce describes, usually has the APIs for this already. Whether you staff the work in-house or buy it as AI integration services, the effort sits in connecting systems you already own, which is the part of تطوير برمجيات الذكاء الاصطناعي that no platform ships in a box. Our view is that in agentic commerce the store that is easiest for an agent to verify will win the agent's sale, the way the fastest page used to win the human's.

A one-week test before Black Friday

A week is enough to see where a store stands before the Black Friday weekend, without starting a formal project. The quickest evidence is already in your CDN or firewall logs, where requests carrying the Signature-Agent header, or classified by your vendor as ChatGPT agent traffic, show what your rules actually did with signed agents. Opening a product page the way a signed agent would, or testing it with your CDN vendor's help, tells you whether the page loads or gets challenged.

البرامج التي تحقق النتائج

نحن نصمم ونبني منتجات رقمية عالية الجودة ومميزة.
الموثوقية والأداء والابتكار في كل خطوة.

اتصل بنا اليوم

For the data check, five products are enough if you ask of each whether a machine could read the all-in total and the return terms without a pop-up or an account. Your payment provider and fraud vendor can tell you in one call how they treat network-token transactions from agents and whether they already read the Forwarded header. The last task is to write down which agents you allow, which you block and which you will review, with a named owner for the list.

Do not expect a wave of agent orders on Black Friday, because Wallet is limited to US Pro subscribers and no retailer has been named as accepting it. Treat the week as a test, and its results as the brief for your agentic commerce work in the new year.

Could an AI agent read your prices, pass your firewall and leave you a record of what it bought? Ask us what we would check first.

نبذة عن المؤلف

This article was written by Nunzio Giugliano at Lasting Dynamics, an EU-based custom software development company that also builds AI software. It is written for the people who own an online sales channel and have to decide what to change, and the documentation, rules and figures it cites are dated and linked so you can check them before you build.

الأسئلة الشائعة

What is agentic commerce?

Agentic commerce is buying and selling in which an AI agent searches, compares and completes a purchase on a person's behalf, within limits that person sets. Today the person still approves each payment, and ChatGPT Wallet, for example, asks for a security check such as Face ID before every purchase.

Do I need to integrate ChatGPT Wallet?

No Wallet-specific integration is needed, because OpenAI says Wallet works with existing checkout pages and payment forms. Whether a purchase goes through still depends on your firewall letting signed agents in, your offer being readable by a machine and your risk rules accepting agent transactions.

Is ChatGPT Wallet available in Europe?

Not at launch. It is rolling out to ChatGPT Pro subscribers in the US aged 18 and older with eligible US-issued cards, and OpenAI has not announced a date for Europe. Any agent payment in the EU remains subject to PSD2 and its strong customer authentication rules.

Should I block AI agents?

It is a business decision, and Amazon's block of Meta's Muse shows it can be a legitimate one. Separate agents that sign their requests, such as ChatGPT's cloud browser, from agents that do not identify themselves, decide which may browse and which may buy, and make sure your existing bot rules do not take that decision for you.

Who is liable if an AI shopping agent buys the wrong thing?

That is still an open question, and no rule written for agents settles it yet. In the EU, disputes are expected to focus on whether the user authorized the payee, the amount and the timing, so a merchant needs its own record of what the agent saw and what was approved, whoever ends up bearing the cost.

رؤيتك، قانوننا

حوّل الأفكار الجريئة إلى تطبيقات قوية.
لنصنع معاً برمجيات تُحدث تأثيراً.

دعنا نتحدث

نونزيو جيوجليانو

نونزيو جيوجليانو هو أخصائي تسويق في شركة Lasting Dynamics، حيث يعمل على استراتيجيات تحسين محركات البحث (SEO) والمحتوى التقني والأبحاث في مجال الأعمال بين الشركات (B2B) حول مواضيع الذكاء الاصطناعي وبرامج المؤسسات والتحول الرقمي. ويركز عمله على توضيح الموضوعات التقنية المعقدة وجعلها مفيدة وسهلة الفهم لكل من الرؤساء التنفيذيين للتكنولوجيا (CTO) والمؤسسين وصانعي القرار والفرق التقنية.

الزبائن الأكاديمية
اتصل بنا
<?xml version="1.0"؟ <?xml version="1.0"؟