An AI agent can buy exactly what you requested and still leave you with an unresolved financial workflow. Your approval establishes permission, not whether the merchant accepted the order, whether a payment completed, or whether the receipt matches the approved cart. Once agents can spend money, those distinctions become product requirements. A checkout timeout is especially revealing: retrying might finish an interrupted purchase, or it might buy the same supplies twice. The right response depends on evidence about the transaction, not on how confidently the agent describes its progress.
Harshith's Newsletter / Reader access
Continue reading.
Subscribe free to Harshith's Newsletter to read the rest of this article and unlock every blog and newsletter edition.
Confirm your email to read. Free. Unsubscribe anytime.We remember this browser for 180 days, unless you sign out or clear site data.Already subscribed?
OpenAI’s ChatGPT Wallet announcement, examined in a November 14, 2026 retrieval, describes users reviewing items, total, and selected card before approving each purchase with a security check. After approval, a payment network supplies a token instead of exposing the user’s actual card details. The announcement describes a limited rollout through dots for ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. That is a concrete payment capability, but not evidence that every downstream purchasing problem is solved.
The useful design move is to separate four things: purchase intent, user authorization, transaction outcome, and reconciliation. This article develops that separation through an illustrative office-supply order, including a changed cart and an ambiguous payment. The state checklist and failure tests are application-design recommendations, not documented Wallet features. They give buyers better approval questions and builders a practical way to decide when an agent may proceed, when it must stop, and what evidence permits the workflow to close.
What Wallet changes, and what it does not
According to OpenAI’s Wallet description, the user adds a card once, reviews each purchase, and approves it through a security check such as Face ID or a fingerprint scan. Each approval creates a record of the purchase the user authorized. The network then provides substitute payment credentials, including a card number, security code, and expiration date. The merchant and agent do not receive the user’s actual card details.
That mechanism separates access to a payment instrument from exposure of its underlying credentials. It also creates an authorization checkpoint. Neither property, by itself, establishes that the merchant submitted the intended charge, created the correct order, or fulfilled it. A secure way to initiate payment is valuable precisely because it can sit inside a larger workflow, rather than being mistaken for that entire workflow.
OpenAI says users retain existing card rewards and protections for eligible disputes, and transact directly with businesses, which retain responsibility for pricing, fulfillment, returns, and customer service. It also warns that risk systems can flag legitimate agent purchases, producing declines or cancellations. These are distinct claims in the Wallet announcement: payment protection is a remedy framework, not a guarantee that an intended purchase occurred.
The examined announcement describes initial rollout through dots, with the age, location, plan, and card restrictions above. It does not establish a generally available Wallet API, detailed token-lifecycle guarantees, or a transaction-status interface that builders can assume exists. It presents parameter-based preapproval as a future expectation, not the current per-purchase approval flow. Design around the capability described in the examined materials, and verify integration details separately.
Four records answer four different questions
A chat transcript tends to compress everything into one narrative: asked, approved, bought. A purchasing system should preserve separate records, because each answers a different operational question. These need not be four separate databases. They do need distinct meanings, identifiers, and rules for updating them.
| Record | Question it answers | Useful contents | What it does not prove |
|---|---|---|---|
| Purchase intent | What does the requester want? | Permitted items, budget, quantity, destination, deadline, substitution rules | Permission to charge a particular cart |
| User authorization | What exact purchase may proceed? | Merchant, cart version, total, currency, payment reference, approver, approval time | Merchant acceptance or payment completion |
| Transaction outcome | What happened after submission? | Attempt identifier, order reference, observed payment status, evidence source, uncertainty | That the order matches intent or was delivered |
| Reconciliation | Do the records agree? | Approved request, merchant receipt, payment evidence, discrepancies and resolution | Physical delivery unless delivery evidence is included |
The central invariant is simple: an execution attempt must refer to the purchase version that was approved. If the cart changes, the earlier approval remains historical evidence, but should not silently authorize the new cart. Likewise, a merchant receipt is evidence of what the merchant recorded. It is not an instruction to revise the approved request until the two appear to match.
This is a more specific version of the handoff problem in AI automation for finance teams. The agent can reduce preparation and follow-up work, but someone still needs to own the boundary between a requested purchase, a committed transaction, and the financial record.
Illustrative worked example: office supplies
An operations manager asks an agent to replenish a small office. The intent record, SUP-042, permits two ten-ream cases of A4 paper and four boxes of black ballpoint pens, each containing twelve pens. The delivered total must not exceed $180. Only one-time purchases are allowed. The order must go to the saved office address and arrive by Friday. Substitutions, subscriptions, and split orders require a new decision.
Those rules make preparation possible without granting blanket spending authority. The agent can compare offers and assemble a cart. It cannot infer that anything under $180 is acceptable. A different paper size, delivery to a home address, or a recurring pen subscription could satisfy the price constraint while violating the request.
| Component | Quantity and price | Amount |
|---|---|---|
| A4 paper, ten-ream case | 2 × $48.00 | $96.00 |
| Black ballpoint pens, twelve-pack | 4 × $9.00 | $36.00 |
| Standard shipping | Delivery by Friday | $12.00 |
| Tax | Checkout calculation | $11.52 |
| Final total | USD, one-time purchase | $155.52 |
The recommended approval view shows the merchant, exact products and pack sizes, quantities, shipping commitment, destination, tax, final total, currency, and selected payment reference. It identifies SUP-042 and cart version A1. The manager approves this snapshot. The $180 budget explains why the proposal is eligible for review; the $155.52 snapshot identifies what the manager actually authorized.
Now suppose the merchant changes shipping to express delivery for $22.00. Holding the illustrative tax constant, the new total is $165.52. The purchase remains under budget, but its commercial terms changed. Under this recommended policy, the application marks A1 stale, creates A2, explains the $10 difference, and asks for renewed approval. It does not let the agent reinterpret the original budget as permission for a changed checkout.
Materiality should be defined before deployment. For this workflow, merchant, product, quantity, currency, destination, total, delivery commitment, and recurring-payment terms are approval-bound fields. Presentation-only changes are not. A blanket tolerance for small price changes is a different authorization policy and should be explicit, not improvised by the model. The examined Wallet announcement does not document how it handles these specific cart-change cases.
Suppose the manager approves A2 and the merchant accepts the order. The application records attempt ATT-01 against A2, merchant order CO-7318, and the available payment evidence. It then compares the merchant receipt with A2: two paper cases, four pen boxes, $22 shipping, $11.52 tax, $165.52 total, correct destination, no recurring terms. A matching receipt supports order reconciliation. Delivery by Friday remains a separate obligation to track.
An uncertain payment is not a failed payment
Change one event in that example. After submitting A2, the agent’s browser times out before displaying confirmation. The application knows it attempted submission. It does not know whether the merchant received the request, whether payment processing advanced, or whether an order exists. The honest status is outcome unknown, not purchase failed.
This distinction prevents the most dangerous recovery shortcut: submitting again because no success screen appeared. A second click may create another order. Even a fresh user approval does not establish that the first attempt failed. The manager might approve the same intended purchase twice while mistakenly believing the first approval produced nothing.
- 01
Preserve the unresolved attempt
Keep ATT-01 linked to A2, including submission time, merchant, total, and any available checkout or payment reference. Do not overwrite it with a generic error or replace it with a new attempt.
- 02
Check existing status before creating new spending
Inspect available merchant order history, confirmation messages, or an authenticated status endpoint if the merchant offers one. Compare payment-account evidence where available. A pending card entry alone does not establish that the merchant created the intended order.
- 03
Classify the evidence, including conflicts
If order CO-7318 appears with the correct cart, continue verification instead of retrying. If authoritative evidence establishes failure and no order commitment remains, evaluate a retry. If a pending payment exists but no order can be found, retain the unknown status and route it to investigation.
- 04
Retry only under a defined recovery rule
Where an integration documents idempotent submission, use its supported mechanism for the same logical purchase. Do not assume a browser checkout has that protection. Without reliable status or duplicate prevention, stop automatic retries and assign a human owner.
Idempotency means repeated submissions of the same logical operation should not create additional effects under the provider’s documented contract. A local attempt identifier helps correlate records, but cannot force an external merchant to deduplicate orders. It is also not a license to reuse an identifier for a materially different cart. The examined Wallet materials do not establish an idempotency contract.
The evidence record should let another person reconstruct the attempt without trusting the agent’s retrospective story. That extends the principle in Your AI agent needs a receipt: preserve what the agent was allowed to do, what it attempted, and what external evidence came back. An agent-generated summary can explain those records, but should not substitute for them.
A proposed transaction-state checklist
Treat this as a minimum application state model. Payment and order status should remain separate attributes: a payment authorization, a merchant order, a settled charge, and fulfillment describe different events. The workflow state expresses whether the application has enough evidence to proceed.
- Draft Intent is recorded; the cart is still being prepared. No purchase permission has been granted.
- Awaiting approval A versioned final snapshot is ready. Required commercial terms are visible, and the cart satisfies the intent constraints.
- Authorized A valid approval references that snapshot. Before submission, check that approval-bound fields still match and any application-defined approval expiry has not passed.
- Submitted An attempt is durably recorded before the external action. Prevent another worker from concurrently submitting the same logical purchase.
- Outcome unknown Submission may have taken effect, but evidence is incomplete or conflicting. Status checks are allowed; automatic new spending is blocked.
- Outcome established Available evidence supports acceptance, decline, cancellation, or another specific result. Preserve the source and distinguish order status from payment status.
- Reconciled or exception The approved snapshot, merchant record, and available payment evidence match, or a discrepancy has a named owner and resolution path. Fulfillment can remain open.
Do not make state transitions depend only on the model’s prose. Recommended enforcement belongs at the action boundary: reject stale approvals, block duplicate concurrent attempts, restrict recovery actions, and require evidence references for closure. The model can help interpret a receipt or explain a discrepancy; deterministic checks should enforce exact totals, currencies, quantities, and version links where structured data permits.
Store only the payment references needed for audit and support, not token credentials in general-purpose logs. Access to purchasing records should also reflect their sensitivity: addresses, item details, and employee purchasing patterns can be confidential even when actual card details are absent.
Test the boundaries, not just checkout success
The following is a proposed failure-test matrix, not a report of tests performed. Run these cases in an appropriate test environment before enabling spending. The useful assertion is not merely that the agent notices a problem. It is that the application prevents the wrong financial action and preserves enough evidence for recovery.
| Injected condition | Expected control | Evidence to inspect |
|---|---|---|
| Shipping price changes after approval | Invalidate the snapshot and require renewed approval | Old and new versions; no submission under stale approval |
| Same price, different paper size | Block the substitution under the example’s policy | Product identity and intent-constraint comparison |
| Timeout immediately after submission | Enter outcome unknown; check status before retrying | Preserved attempt and status-check results |
| Two workers submit simultaneously | Allow only one active execution path | Concurrency-control record and external attempt count |
| Pending card entry, no visible order | Keep payment and order status separate; escalate | Payment observation, merchant search, assigned owner |
| Receipt contains an extra pen box | Open a reconciliation exception | Approved quantity versus merchant quantity |
| Checkout adds recurring billing | Block execution and request a new decision | Changed terms and blocked action |
| Agent reports success without external evidence | Refuse reconciliation closure | Missing order or receipt reference |
Include a restart case too: interrupt the application after it records submission but before it records the response. On recovery, it should investigate the existing attempt, not resume from the last conversational instruction to buy. This is where durable workflow state matters more than an excellent prompt.
Limits: safeguards do not remove accountability
The dots product description says Custom Rules can allow actions, require approval, or block them, while built-in safety requirements continue to apply. It also describes action review and explicitly advises reviewing consequential work because dots can make mistakes. Those are useful governance layers, but the examined page does not establish how every purchasing constraint maps into Wallet enforcement.
The examined sources are vendor descriptions, not independent end-to-end reliability evaluations. They do not provide enough detail to quantify duplicate-order rates, approval-to-cart binding strength, or recovery performance. That limits implementation conclusions from these materials; it is not a claim that independent work exists nowhere.
A biometric security check addresses approval security. It does not establish organizational purchasing authority, accounting treatment, or whether the manager understood a pack-size substitution. Eligible card dispute protections may help resolve a problem, but they do not replace prevention or evidence. A reconciled order can still arrive late, damaged, or incomplete.
When the agent buys something outside the approved scope, investigate an authorization incident rather than treating it only as an expense discrepancy. When an AI Agent Succeeds the Wrong Way provides the next layer: preserve evidence, classify the boundary violation, and change the control that allowed it. A refund can resolve the money without resolving the underlying failure.
Use this next: choose the delegation boundary
Start with one narrow purchasing category, not general spending. Choose predictable items, explicit substitution rules, and a merchant whose order evidence is accessible. Then decide which of three delegation boundaries your application can actually support.
- Preparation only Use this when final terms or transaction status cannot be reliably inspected. Let the agent compare products and prepare a cart; keep checkout with the user.
- Approved execution with manual exceptions Use this when the final cart can be reviewed and external order evidence is available, but ambiguous outcomes still require human investigation. Make unknown status a normal, visible handoff.
- Controlled execution and recovery Use this only when the integration supports reliable status inspection, documented duplicate prevention where needed, durable attempts, and tested reconciliation. These are prerequisites to verify, not capabilities to infer from the Wallet announcement.
Before enabling a purchase, answer four questions: What exact version is approved? What changes invalidate that approval? How will an uncertain outcome be resolved without new spending? Who owns a mismatch between the approved cart, merchant receipt, and payment record? If any answer is vague, reduce the agent’s authority rather than increasing the prompt’s detail.
My judgment is that the most important design artifact here is not the approval button. It is the unresolved-attempt record. Approval opens the possibility of a financial commitment; that record keeps an interrupted workflow from inventing certainty or permission. Build the path from uncertainty to evidence before building the retry path.

