Wise: Approve the Job, Not Every Step

Emil Chetty, who leads data and AI product at Wise, told AI Engineer to stop making customers click approve on every agent step. No prediction market venue, odds or traded volume were disclosed for this talk and no price move followed it. That leaves no liquid market signal to test demand for the model. Not financial advice. Markets move fast.

Why it is news now is the control problem. Wise moved more than 243 billion in cross-border volume last year for nearly 19 million active customers across 90 local bank connections, and it holds more than 80 licenses worldwide. Chetty said regulators still expect the firm to know who authorized a payment, control what happens to money, and explain the outcome, even with a chatbot in front.

Coding agents taught the wrong lesson. Chetty argued teams made the human the API, watching the terminal and approving each tool call.

His alternative is short. Get the customer to agree to the job, let the agent choose the steps, and let services enforce the agreement. In his example Sarah the business owner says handle this week's approved supplier payments, reconcile results, and return exceptions. Trusted software resolves suppliers, bank details and amounts from records with read-only access, then shows Sarah a human promise and a machine-readable promise: three suppliers, 775 dollars plus fees, plus where the agent may decide and where it must return. She authenticates once, that draft becomes an immutable authorized job, and what she approved is what services enforce.

Execution keeps flexibility inside bounds. A provisioning platform checks the confirmed job, what Wise permits the agent to do, and Sarah's current rights, then grants limited rights to a payment worker and a reconciliation worker. The plan can change. The agreement cannot. If a Kubernetes pod restarts or her app crashes, workers resume the same job and share payment state instead of creating a duplicate. If a quote expires or terms change, the agent cannot rewrite its own agreement, it must surface a revised proposal for a new customer decision. Prompt injection that adds new payee details fails at the service check, Chetty said, because the boundary does not depend on the model spotting the scam.

The impact lands on finance operations, treasury and platform teams that budgeted for human-in-the-loop approvals. They have to replace per-step clicks with a single job authorization, an immutable job record, and service-side enforcement, or keep staffing the queue they promised to remove. For builders the pattern travels beyond payments, Chetty noted, to any regulated action where the question is not does this look safe but did this customer authorize this recipient on these terms. Europe's strong customer authentication and the U.S. GENIUS Act stablecoin rules were not written for software buyers, so a mandate proves intent was signed, not that it was wise, and the payment is only one action inside the automated procurement task.