Emil Chetty, who leads data and AI product at Wise, explains a payment-agent authorization design in this 15-minute AI Engineer talk.
Approve the job once
- Repeated tool approvals turn the customer into a queue of decisions while the assistant waits.
- A payment needs proof that this customer authorized this recipient and these terms.
- Turn the request into a verified proposal, show it to the customer, then lock the approved version.
Keep planning flexible and authority fixed
- Begin with limited read-only exploration; trusted software resolves suppliers, bank details, and amounts.
- Give the customer a readable agreement and the services a matching machine-readable record.
- Provision permissions from the intersection of that record, Wise policy, and the customer’s current rights.
- Let payment and reconciliation workers choose their steps while payment services enforce the agreement.
Handle retries and changed terms separately
- Restarted workers follow the existing payment job and its recorded state to avoid duplicate payments.
- A changed or expired quote requires a revised proposal and another customer decision.
- An injected instruction to pay a new account fails the service check, even if the agent believes it.
Keep the trust boundary in software
- Record who approved the job, what authority was granted, and what actually happened.
- Keep model tests and execution controls alongside the customer-authorization record.
- Caveat: Chetty presents Wise’s design with a hypothetical supplier-payment example; he provides no measured rollout or adversarial-test results.
“What you approve is ultimately what we enforce.” — Emil Chetty, 10:18