Your AI Bought What?
Imagine telling an AI assistant:
“Book me a decent hotel near the conference. Nothing fancy. Keep it under $250.”
It finds one for $239. Success.
Then come the resort fee, parking charge, nonrefundable cancellation policy, and a room overlooking the industrial air-conditioning equipment. The agent followed your instructions. Sort of.
That “sort of” is where banks are headed next.
Six large banks recently published principles for payments made or helped along by AI agents. Their basic point is sensible: customers should know when an agent is acting, what permission it has, how to stop it, and who is responsible when things go sideways.
These are voluntary principles, not new regulations. The detailed plumbing hasn’t been built yet. But the banks have identified the right problem.
A normal payment record can tell us which card was used, how much was charged, and whether the transaction was approved. It may not tell us what the customer actually asked the AI to do.
Was the agent allowed to spend $250 before fees or after them? Could it choose a nonrefundable option? Did the customer approve that merchant? Had the permission expired? Did the agent make the purchase, or merely help the customer make it?
Fraud and dispute teams will eventually need those answers. "The transaction was authorized” won’t be terribly comforting if nobody can show what was authorized.
Computers are excellent at saving 4,000 lines of technical activity while misplacing the one sentence a human cares about.
Banks and credit unions don’t need an “agentic commerce transformation program” tomorrow morning. They do need to ask their payment, digital banking, and fraud providers a plain question:
If a customer disputes a purchase made by an AI agent, can you show us the agent, the customer’s instruction, the limits on its authority, and the moment the payment was approved?
If the answer is only, “The agent used valid credentials,” that proves the wallet opened. It doesn’t prove the customer said yes to what came out of it.
Comments