OpenAI announced ChatGPT Wallet on November 13, 2026. The initial rollout lets eligible US ChatGPT Pro subscribers use dots to complete purchases, while requiring a user security check for every purchase. For QA teams, the news is less about a new checkout button than a new authorization boundary: an agent can prepare a transaction, but the final purchase must remain accurately reviewed, explicitly approved, and traceable.
OpenAI says users review the items, total, and selected card before approving with a security check. After approval, the payment network supplies a payment token instead of exposing the actual card details to the agent or merchant. Availability is limited and gradual, so treat any integration as a controlled rollout—not a promise that every customer, card, site, or risk system will behave identically.
What changed with ChatGPT Wallet
- Each current purchase requires the user to review and approve it.
- The merchant receives a payment token, rather than the user’s real card credentials.
- OpenAI says existing checkout flows can work with Wallet tokens, but fraud and risk systems may still decline or cancel legitimate agent-led orders.
- The first rollout is for eligible US ChatGPT Pro subscribers, age 18+, using eligible US-issued cards through dots.
Why this matters for QA engineers
Agentic checkout creates a three-party contract between the shopper, agent, and merchant. A green order confirmation alone is not enough evidence. Your test suite should prove that the approved basket, amount, identity, authorization state, payment result, and resulting order all refer to the same transaction—and that the system fails safely when they do not.
A focused ChatGPT Wallet QA checklist
- Approval integrity: Change price, quantity, shipping, seller, or tax after the review screen and assert that the existing approval cannot be reused.
- Token boundaries: Confirm logs, analytics, support exports, and browser telemetry never contain full card data. Store only approved transaction references.
- Idempotency: Retry a timeout, reconnect a client, and replay a webhook. Exactly one successful approval must create at most one order and one capture.
- Risk decisions: Test legitimate agent requests that are approved, declined, challenged, or later canceled. Each outcome needs a customer-safe explanation and a reconcilable backend status.
- Access and bot controls: Verify signed agent requests are accepted where intended while unsigned, expired, altered, and unauthorized requests remain blocked.
- Audit trail: Correlate the reviewed basket, explicit approval, tokenized authorization, order ID, fulfillment status, and refund or dispute path without exposing sensitive payment data.
Use an evidence-first test matrix
| Scenario | Expected result | Evidence to retain |
|---|---|---|
| Basket changes after review | New approval is required | Old and new basket hashes; approval timestamps |
| Payment callback retries | One order and one capture maximum | Idempotency key; processor and order events |
| Risk-system decline | No fulfillment; clear recoverable state | Masked decision code; order state transition |
| Unauthorized agent request | Request rejected before checkout | Authentication failure event; no order record |
Start with a narrow pilot
Begin with a staging catalog and low-risk products. Freeze representative price, tax, inventory, and promotion fixtures; inject approved failure paths; and reconcile processor events with your order system. Keep agent access, payment approval, and fulfillment as separate checkpoints. That separation makes it possible to locate a defect without mistaking a payment decline for an agent failure—or vice versa.
OpenAI also advises businesses to make product data accurate, avoid unintentionally blocking agents, and work with risk providers on legitimate agent transactions. Those are production-readiness tasks, but they are also concrete regression-test requirements.
Bottom line
ChatGPT Wallet QA should center on a simple rule: an agent may assist with checkout, but a testable user authorization and a privacy-safe evidence trail must govern the purchase. Validate the boundaries now, while the rollout is limited, before agent-enabled commerce reaches a broader customer path.