JPMorgan Chase and South Korea's Hyundai Card have completed a proof of concept in which an autonomous AI agent executed a foreign exchange transaction with a bank agent and no human in the loop, and they plan a harder experiment next: multiple AI agents negotiating and paying each other inside an ecosystem. Wee Kee Toh, global head of Kinexys Insights at JPMorgan's blockchain unit, told CNBC the next phase, targeted for the first half of 2027, is about what happens when agents stop merely executing orders and start interacting with one another, and he called that expansion the biggest part of the planned scope.
The trial and what it proved
The initial test had a Hyundai Card AI agent request a currency trade to maintain euro liquidity in its Kinexys blockchain deposit account, and JPMorgan describes it as the first private blockchain-based autonomous test between a bank's agent and a client's agent for corporate treasury management. The transaction was simulated, and the bank has set no timeline for commercialization. The choice of partner matters as much as the technology: Hyundai Card sits inside Hyundai Motor Group, which holds a 78.04 percent stake through Hyundai Motor, Kia and Hyundai Commercial, and JPMorgan picked it partly because the group's broader operations offer more places to deploy agents, though no specific expansion projects have been announced.
The box around the agent
Toh's verdict on where the difficulty lives cuts against the usual assumption that the technology is the barrier. Building it was relatively straightforward, he said; what proved harder was governance and internal controls, meaning accountability when an AI agent executes financial transactions without direct human oversight. Kinexys answered with a constraint framework that confines each agent to predefined transaction parameters and automatically rejects anything outside them. His image for it is blunt: draw a box around the agent, and the agent can never do anything outside the box, so an agent authorized to transact up to $5 million simply cannot execute a larger trade. The design accepts capability limits in exchange for auditability, a trade regulators are likely to demand before any of this reaches production.
A bank already deep in blockchain rails
Neither side is new to this territory. Hyundai Card completed a proof of concept in July moving $20,000 in USDT stablecoin between Hyundai Motor's US and Mexican subsidiaries, with transfer and verification averaging seven minutes. Kinexys, meanwhile, said in August that it has processed more than $4 trillion in blockchain transactions since inception, about $7 billion a day, and rolled out its JPMD dollar deposit token to institutional clients last November. The agent trials add an autonomous decision layer on top of rails that already work at scale, which is what makes the governance question urgent rather than theoretical. A bank that spent last week explaining a software bug that slowed payments across the entire US system is now deliberately building systems designed to act without a human present, and the box it draws around those agents is the bet that structure beats supervision.