Hyundai Card and JPMorgan have simulated a pound-to-euro exchange transaction in which an artificial intelligence agent detected that funds were missing in a company account in Frankfurt, compared offers from three banks and requested the conversion. A second agent, operated by Kinexys, the blockchain unit of JPMorgan Payments, verified that the order was authorized and executed it in the test environment. The transaction was recorded automatically and without moving real money, according to JPMorgan.
A routine treasury task
The test reproduces a routine treasury task, such as maintaining a minimum balance in an operating account. In this case, the Hyundai Card agent periodically checked the balance of a Kinexys blockchain deposit account in Frankfurt. When it detected that it was below the minimum, it requested quotes to sell pounds from a designated account in London and cover the euro deficit. It chose the most competitive offer and sent the order to Kinexys.
Kinexys's technical report describes how the bank verified that the agent could act on behalf of Hyundai Card. The company approved a digital mandate linked to the agent's cryptographic identity. That mandate set out the permitted transactions and the conditions under which a person had to intervene. Upon receiving the order, Kinexys verified who sent it, on behalf of which company it was acting and whether the request complied with the rules on currencies, amounts and schedules.
The test included three situations. The pound-to-euro conversion was executed automatically because it met the conditions of the mandate. In a second scenario, Kinexys blocked an order in a currency that Hyundai Card had not authorized. In the third, the system held a request for an amount greater than the agent could manage on its own and only released it after the company's approval.
The system recorded the agent's objective, the order it sent, the bank's checks and the result in an immutable audit log. It also retained the version of the mandate that was in force at the time of each request. This way, a treasury manager can review why an order was executed and what permissions the agent had when it submitted it.
Payments from the car
The first phase was limited to a relationship between a bank and a client, two designated accounts and a one-way conversion, from pounds to euros. Although the agent compared quotes from three banks in the test scenario, the interaction between the two agents took place within JPMorgan's environment. The report identifies as an open issue that different entities can recognize and verify the permissions a company grants to its agent.
Crypto activity grows 24.6% in Spain and reaches $156.1 billion
Wee Kee Toh, global head of Kinexys Insights, believes that treasury can serve as a starting point for more complex transactions between agents and machines. For its part, Hyundai Card has pointed to payments from the car as a possible application of its collaboration with the bank, although it has not yet detailed what transactions could be carried out. In South Korea, Hyundai already allows the driver to pay for fuel or parking from the vehicle's screen with a registered card. An AI agent could go a step further and, if the driver asked it to find a compatible charger on their route and spend up to 25 euros, it could choose the station and initiate the payment within that limit.