A user approves a purchase. The merchant receives a submission, but the browser closes before it displays an order number. The assistant knows it had permission to try. It does not yet know whether the merchant placed the order.
OpenAI’s November 13 Wallet announcement matters because it gives payment by an agent a dedicated review step instead of making checkout another ordinary browser action. The purchase guide makes a useful distinction explicit: approval completes the user’s review; it does not confirm an order. This article uses that distinction to examine how an agent should represent a checkout outcome.
Start with the rollout that is documented
The release notes describe a gradual rollout in dots for US ChatGPT Pro subscribers aged 18 or older, with an eligible US-issued card and US billing address. The purchase guide says Wallet currently works in dot conversations and is unavailable in Chat, Work, and standalone Codex threads. The launch describes parameters for broader delegated purchasing as future work. Account, region, card, and merchant availability were not tested here.
The approval flow shows the merchant, items, total, and selected card and requests biometric verification. After approval, OpenAI describes a substitute payment credential handled through the payment network without exposing the user’s underlying card details to the merchant or agent. Its merchant guidance says existing checkout processes payment, while the merchant remains responsible for the order and fulfillment. These are separate systems with separate records.
Distinguish what the records actually establish
| Evidence | What it supports | What remains open |
|---|---|---|
| User approval record | The user consented to the reviewed purchase | Whether checkout was submitted or succeeded |
| Provider payment observation | An authorization, capture, decline, or another payment state, as reported | Which order it belongs to, or whether fulfillment happened |
| Merchant order confirmation | A merchant acknowledges a specific order | Whether the item has shipped or will arrive |
| Fulfillment evidence | A shipped, delivered, or otherwise completed item, as recorded | Any remaining return, dispute, cancellation, or service obligation |
A pending charge is especially easy for a model to overinterpret. It may be visible before an order can be found. An empty confirmation slot in a browser does not prove nothing happened either. The Help Center advises checking the merchant’s order status when a result is unclear. An application that models states explicitly can simply state that the purchase was approved and the merchant outcome is unresolved. That is more precise than a false choice between “bought” and “failed.”
The product guide also distinguishes saving a card from approving a purchase and cautions that an accepted recurring plan may authorize subsequent merchant charges on those terms. For an engineering review, bind the review request to the actual offer: seller, item, quantity, final amount and currency, applicable fulfillment terms, and any recurring commitment. This is a proposed control boundary; OpenAI’s documentation does not establish which of these fields are machine-compared for every merchant. If a material field changes, bring the changed proposal back into review.
Design the pause before you design the retry
A confirmed order should have a merchant identifier, and each retry of the same proposed checkout should be linked to a stable operation record. For a controlled integration, record the evidence needed to reconcile an interrupted attempt. Do not infer that a payment token ensures one merchant order. Payment authorization and merchant order creation are coordinated by a larger system, and their timing can differ.
The ContextOS execution contract separates confirmation, definite non-application, and unknown effects. That is an application design principle, not a description of Wallet’s internal implementation. If merchant reconciliation cannot determine what occurred, preserve the unknown result for review. A safely interrupted workflow may be less convenient than an immediate retry but prevents a second external action being mistaken for simple recovery.
Evaluate the checkout as a state machine
Use a synthetic store and fake payment provider; no real payment is needed. Predefine the effect log and what each response means. Then vary which record is available when the agent provides its summary.
- Change the shipping cost after review. Test whether the modified order requires a new decision.
- Return a simulated merchant confirmation late. Test whether the agent keeps the operation pending until it has evidence and later links the right receipt.
- Let the merchant accept an order while the payment UI loses its response. Verify that the same operation is reconciled without a silent duplicate.
- Include a routine success where all records arrive promptly. Evaluate how much friction the checks add to legitimate work.
Measure incorrectly claimed success, duplicate effects, unresolved time, valid completions, and user interventions per attempted order. Analyze low-friction successes alongside the uncommon ambiguous states; a high checkout completion rate can leave expensive mistakes hidden in its remainder. This is a proposed evaluation, with no claims about a measured Wallet failure rate.
The research opportunity is to make commerce agents trustworthy at the boundary where a conversation turns into an external commitment. A secure approval is an important starting point. A trustworthy result tells the user what was approved, what was observed, and what still needs to be resolved.