An agent asks a user to approve a $120 order. The biometric check passes. Checkout then times out. Did the user authorize the purchase? Yes. Did the issuer authorize the payment? Unknown. Did the merchant accept the order? Unknown. Should the agent try again? Not until it can distinguish those facts.
That distinction is the control plane for agentic payments. A single success flag cannot represent user intent, payment execution, merchant acceptance, and delivery. A production system needs three linked receipts, each issued by the actor that can actually prove its state.
Product facts rechecked on 14 November 2026. The receipt model is an engineering synthesis, not an OpenAI protocol specification.
Approval is one boundary, not the whole transaction
OpenAI's ChatGPT Wallet announcement says the launch flow lets a user review the items, total, and selected card, then approve each purchase with a security check. The payment network supplies a tokenized credential instead of exposing the user's actual card details to the agent or business.
The legal and operational boundaries are more important than the token itself. The Wallet terms say approval authorizes a specific merchant, items or services, amount, card, recipient, delivery details, and other displayed terms. They also say that submitting an approved order does not guarantee merchant acceptance. The issuer decides whether to authorize the transaction; the merchant decides whether to accept and fulfill it.
OpenAI's help article makes the recovery rule explicit: ready to check out means approval is complete, not that the merchant placed the order. If the result is unclear, check for merchant confirmation before trying again.
These statements define three different claims:
- The user authorized a bounded purchase.
- A payment attempt reached a known or unknown financial state.
- The merchant accepted and fulfilled an order.
Collapsing them turns a timeout into a duplicate-charge risk and a payment success into a false delivery claim.
Receipt 1: bind intent to one canonical purchase
The intent receipt should be created at approval. It is not a copy of the chat transcript. It is a canonical, immutable statement of what the user approved.
At minimum, record:
{
"intent_id": "int_01",
"principal": "user_42",
"merchant": "merchant_17",
"items_hash": "sha256:...",
"maximum_total": {"amount": 12000, "currency": "USD"},
"recurring": false,
"delivery_hash": "sha256:...",
"approved_at": "2026-11-14T02:10:00Z",
"expires_at": "2026-11-14T02:25:00Z",
"approval_method": "device_biometric"
}
The important field is the binding, not the label. A changed merchant, item, quantity, price ceiling, recipient, address, or recurring-payment term creates a new intent. The agent may explain the change, but it cannot stretch an old approval to cover it.
This is the payment-specific extension of an agent action contract: authorization must bind to the exact normalized action, not merely to a conversational goal.
Receipt 2: preserve payment state, including unknown
The payment receipt belongs to the execution boundary. It should link the intent to one logical attempt and retain the provider, network, or issuer evidence available to the system.
Useful fields include:
payment_idandintent_id;- a stable idempotency or correlation key;
- credential reference, never the full sensitive credential;
- amount and currency actually submitted;
- provider or issuer reference;
- state such as
not_submitted,submitted,authorized,declined,reversed, orunknown; - last observation time and reconciliation path.
Unknown is a first-class result. A network timeout after submission does not prove failure. Generating a new logical payment because the agent did not receive a response can create a second charge. Recovery should query the existing attempt by stable evidence, wait for a callback, or escalate. A retry is safe only when the downstream contract deduplicates the same logical action or the previous attempt is conclusively closed.
The payment receipt also prevents a different mistake: interpreting authorization as settlement. Card authorization, capture, settlement, reversal, and dispute are different financial states. Keep the vocabulary aligned with the actual processor rather than inventing a generic paid flag.
Receipt 3: make the merchant close the business loop
The merchant is the seller and merchant of record under the Wallet terms. OpenAI's merchant guidance says the business handles the order, fulfillment, customer support, and fraud decisions. It also says the customer consents only to the specific transaction and the payment credential must not be reused for another purchase.
The fulfillment receipt therefore needs merchant evidence:
- merchant order ID and accepted purchase snapshot;
- acceptance, rejection, cancellation, or back-order state;
- fulfillment units and promised delivery method;
- shipment, reservation, download, or service-completion evidence;
- refund or return references;
- support owner and last state transition.
For physical goods, delivery may close days after payment. For a digital service, payment can succeed while entitlement creation fails. For a reservation, the merchant may authorize the card yet reject inventory. The fulfillment receipt keeps those failures visible instead of letting a financial event impersonate a completed user task.
Use a state machine, not a completion sentence
A minimal state model looks like this:
PROPOSED
-> INTENT_APPROVED
-> PAYMENT_SUBMITTED
-> PAYMENT_AUTHORIZED | PAYMENT_DECLINED | PAYMENT_UNKNOWN
-> ORDER_ACCEPTED | ORDER_REJECTED | ORDER_UNKNOWN
-> FULFILLED | CANCELLED | REFUNDED | DISPUTED
Three invariants matter:
- No material purchase change can reuse an earlier intent receipt.
- No unknown payment can trigger a fresh logical payment before reconciliation.
- No agent can report the user task complete without merchant-side fulfillment evidence appropriate to that task.
The states can live in one database. Their evidence and ownership should remain separate. That separation makes crash recovery possible: after restart, the agent reads the durable state and performs the next allowed query instead of reconstructing reality from its last natural-language message.
Measure four funnels, not one success rate
Teams often want a single checkout conversion number. That aggregate hides the control failure they need to fix.
Track at least four denominators:
| Funnel | Example measure | What it diagnoses |
|---|---|---|
| Intent | approved / presented | user understanding, changed details, expired approvals |
| Payment | authorized / submitted | issuer, credential, processor, and risk outcomes |
| Order | accepted / payment-authorized | merchant checkout and inventory acceptance |
| Fulfillment | completed / accepted orders | delivery, entitlement, cancellation, and support |
Also track payment-unknown duration, reconciliation closure, duplicate attempts prevented, approval-to-payload mismatches, and paid-but-unfulfilled cases. These metrics have different objects and windows. Summing them into one agent success score destroys the evidence needed for remediation.
Launch with failure drills
Before enabling real purchases, rehearse the transitions that a happy-path demo skips:
- change price, merchant, quantity, address, or recurring terms after approval;
- cut the connection immediately before and after payment submission;
- deliver duplicate callbacks and out-of-order merchant events;
- authorize payment while the merchant rejects inventory;
- accept the order while fulfillment or entitlement creation fails;
- cancel, partially refund, and dispute an order;
- restart the agent at every unknown state.
The pass condition is not that the model explains the problem. It is that deterministic controls refuse stale intent, deduplicate execution, reconcile unknowns, and produce a user-visible status that matches the responsible system of record.
Current product boundary
At launch, ChatGPT Wallet is rolling out gradually through dots to eligible US ChatGPT Pro users aged 18 or older with eligible US-issued credit cards and supported devices. The help page says it is not available in Chat, Work, or standalone Codex threads. OpenAI has not published merchant acceptance rates, fraud false-positive rates, payment-unknown rates, or long-run dispute outcomes.
Those limits matter. The official materials prove a concrete separation between approval, payment processing, merchant acceptance, and fulfillment. They do not prove that the three-receipt schema above is implemented as written or that early reliability generalizes to every merchant and payment rail.
Conclusion
Tokenization protects card data. Biometric approval protects a bounded act of intent. Neither proves that money reached a final state, the merchant accepted the order, or the user received what they asked for.
Agentic payment systems should preserve three receipts and three responsible parties: the user's intent, the financial attempt, and the merchant's fulfillment. When an outcome is unknown, the agent's next job is reconciliation, not optimism and not a blind retry.