You approve the product, the amount and the payment. But who decided beforehand which offers you would see, which criteria would matter most and what could be ruled out? ChatGPT Wallet gives us a concrete reason to examine agentic shopping: the decisions we delegate to AI and what that division of responsibility means for brands.
OpenAI is developing the ability for agents to make purchases. For users, it promises convenience: less browsing, comparing offers and filling in forms. For businesses, it creates an additional task: ensuring that systems acting on a customer’s behalf can interpret an offer correctly and check whether it meets the requirements.
But the most interesting change may happen before payment. It concerns who narrows down the choice.
What has OpenAI announced about ChatGPT Wallet?
On 13 November 2026, OpenAI introduced ChatGPT Wallet, a wallet that enables agents to make payments on a user’s behalf. The solution was developed in collaboration with American Express, Chase, Mastercard and Visa. The initial phased rollout covers a limited group of adult ChatGPT Pro users in the US who hold eligible cards issued in the country.
Before a purchase, the user reviews the products, total amount and selected card, then approves the transaction and completes a security check. The agent receives substitute payment details in the form of a token, without access to the actual card details.
Importantly, OpenAI also describes a future direction – authorising an agent in advance to make purchases within defined limits, such as a budget, category or timeframe. This is a preview of future development, not a description of a feature generally available in the current rollout.
This news builds on a topic I explored in my article about shopping, advertising and personalisation in ChatGPT. This time, however, I’ll look at how control over a purchase is distributed.
What does agentic shopping actually mean?
For this analysis, I use an operational definition: agentic shopping is the extent to which an AI system can select and carry out actions leading to a purchase, within the bounds of the user’s task and authorisation.
This definition makes it possible to examine specific behaviours. Simply generating a list of products tells us little about a system’s autonomy. What matters is whether it can search for information, change its plan, choose tools, compare offers and respond to obstacles.
Anthropic describes a similar architectural distinction: in a pre-programmed workflow, code determines the next steps, whereas in an agentic system, the model dynamically directs the process and use of tools. This is useful conceptual context, though it does not describe the specific architecture of ChatGPT Wallet.
In shopping, the degree of autonomy can vary considerably. A user might specify an exact product and ask the agent only to place the order. Or they might describe a need, budget and deadline, leaving the agent to choose the product and retailer.
Both tasks end in a purchase. In the second, the system has much greater influence over what gets bought.
Payment approval is just one point of control
Imagine asking an agent to find a suitcase for a specific flight. The system presents one option, explains its advantages and asks you to approve the purchase.
You control the spend. But to assess your control over the choice, you need more information – which offers the system considered, which it rejected and why it judged the selected product suitable. It may have checked every requirement correctly. Or it may have treated the ambiguous phrase “cabin baggage” as sufficient proof that the suitcase meets the rules of a particular fare. These are very different outcomes, even if the payment approval screen looks identical.
Approving a transaction confirms that the user accepted the purchase presented. On its own, it does not prove that the earlier selection was complete and accurate.
This is our methodological conclusion, not a description of an observed Wallet error. It shows why research into AI-assisted shopping should also cover the decisions made before payment.
A framework for delegated shopping
Rather than assigning a single “autonomy level” to the entire process, I suggest breaking it down into decisions and actions. For each stage, establish what the person specifies, what the agent can do and when it should ask for further approval.
area | what does the user specify? | what should be checked in the agent’s behaviour? |
|---|---|---|
need and constraints | goal, budget, deadline, mandatory requirements | whether the system preserves the requirements and clarifies ambiguities |
search and comparison | the scope of the search, or permission for the agent to define it | which sources and offers it considers, and how it justifies exclusions |
product and retailer selection | the exact choice, or the criteria for making it | whether the recommendation meets the criteria and flags missing data |
actions in the shop | the permitted actions | whether the system selects the right variant, delivery option and terms |
payment and exceptions | approval for the spend and acceptable deviations | whether a change in price, product or terms requires a fresh decision |
To be clear, this is an analytical framework of my own, not an official OpenAI classification. Its strength is that it can describe different configurations. An agent might have considerable freedom to search, but no freedom to pay. Or it might purchase an exact product specified by the user, making few decisions about the offer itself.
For a brand, this means distinguishing between two situations: the customer has already chosen it and delegates the purchase, or they ask the system to choose from the brands available.
Example: a suitcase that must meet specific requirements
Consider this hypothetical instruction:
Find a suitcase for up to 150 euros that meets the under-seat baggage limit for my flight and will arrive before Friday. If you can’t confirm any of these requirements, ask me before buying
This task requires several types of information to be brought together, including the dimensions of the specific product variant, the carrier’s rules for the relevant flight and fare, the total price, availability and delivery date to the specified address.
Notice that the phrase “ideal for flying” does not establish whether the suitcase meets the requirements, and “dispatched within 24 hours” does not confirm a delivery date. What’s more, the price shown in a comparison may not include delivery.
That’s why I suggest assigning one of three statuses to each mandatory requirement:
confirmed – available evidence shows that the specific offer meets the requirement
not met – available evidence shows that the offer fails to meet the requirement
undetermined – there is not enough information to reach a conclusion
A lack of confirmation should not automatically be treated as either compliance or non-compliance. In our example, it is a reason to investigate further or ask the user.
For businesses, this offers practical guidance on product content. It is worth specifying which claims need evidence and which product variants they apply to. We explore this approach to structuring offers in more detail in Semantio’s article on preparing products and services for AI answer research.
What needs to be clear in a brand’s offer?
In its guide for commerce partners, OpenAI notes that information used in the shopping journey may come from the public web, private API integrations and product data feeds. The task may then be carried out through a browser or an integration that exposes specific actions.
The implication is that businesses need to check information for consistency across these sources. If the product description, specification and offer data give different dimensions or terms, the discrepancy needs to be resolved.
In practice, it is worth linking:
a product to a clearly defined variant
an attribute to its unit, scope and conditions of use
a claim to a current source of evidence
an offer to its retailer, market, price and availability
This is a practical application of the approach described in Brand semantics infrastructure. In shopping, the relationship between a product claim and the requirement a customer wants to meet becomes particularly important.
For Brand Semantics, this means extending work on a brand’s semantic strategy to ask: do the available details make it possible to verify the offer in a real decision-making situation?
A clear offer and a working purchase journey need separate tests
Correctly identifying a product does not guarantee that the task can be completed successfully.
web.dev explains that agents can use screenshots, HTML and a page’s accessibility tree. That makes form labels, element roles, clear interface states and links between buttons and the right products important. These are technical pointers for agent interaction with a website, not a guarantee that every purchase will succeed.
Payment is a separate layer. According to OpenAI’s payment documentation, Wallet works with existing payment forms without an integration dedicated solely to the wallet. At the same time, risk assessment systems may mistakenly block purchases made through an agent’s browser.
Businesses therefore need to diagnose these issues separately. Misreading product specifications, selecting the wrong variant and having a payment declined have different causes and require different responses.
How can you test whether delegated AI shopping changes brand choice?
From an AI visibility research perspective, it is useful to compare three ways of framing the same task:
exploration – present several offers that meet the requirements and explain the differences
selection – choose one offer, justify the choice and flag any uncertainties
action preparation – check the selected variant, cost and terms before approval
This is a proposed research design, not the result of a completed experiment. It can help establish whether a brand that appears in a comparison remains the recommendation when the system is asked to select one product, and whether the reasoning still relies on accurate information.
For the analysis, retain the full responses, visible sources and test conditions. Repeat comparisons while accounting for factors such as platform, mode, language, market, date and available user context.
You could also use a simple supporting metric: the share of recommendations for which every mandatory scenario requirement can be verified. The denominator would be the recommendations assessed, and the numerator those that passed verification. The result would apply only to the set of tasks and conditions tested.
This extends the approach to measuring brand visibility and representation in AI: alongside brand presence, we assess whether the brand is a sound match for the need.
Semantio is a separate research tool for analysing brand representation in AI responses. In this context, it can support research into presence, recommendations, errors and sources. Testing shop interactions, payment authorisation and order confirmation requires a separate study of task execution. Monitoring response outputs does not confirm what happens during a Wallet purchase.
What doesn’t ChatGPT Wallet mean?
The Wallet announcement does not justify claiming that autonomous shopping is already widespread, including in Poland. Nor does it show how the rollout will affect sales for individual brands.
The ability to pay does not, by itself, prove that the product was selected correctly.
Well-structured data and an accessible website do not guarantee a recommendation.
A correct recommendation is not yet a confirmed order.
The key research questions remain open:
Does broader delegation actually narrow the set of brands considered?
How often does the system flag missing data?
Does it ask for approval when the terms change?
How do results vary across markets and types of purchase?
ChatGPT Wallet gives these questions a concrete context. For brands, a useful starting point is the customer’s task: establish what needs to be checked, where the evidence can be found and whether the offer still makes sense when the system uses it to prepare an actual purchase.
And I’ll leave you – and myself – with those questions ;)

