JPMorgan and Hyundai Card Complete First AI-Agent FX Trade Trial
JPMorgan and Hyundai Card ran a simulated FX transaction between autonomous AI agents on the Kinexys blockchain. Agent-to-agent negotiation tests are targeted for H1 2027, with no commercial timeline set.

Why it matters
- JPMorgan and Hyundai Card completed a proof-of-concept in which AI agents executed a simulated FX transaction on a private blockchain, reported by CNBC on November 17, 2026.
- A more complex multi-agent testing phase, including agent-to-agent negotiation, is targeted for the first half of 2027.
- JPMorgan calls it the first private blockchain-based test between autonomous bank and client agents for corporate treasury management.
- Kinexys's framework confines each agent to predefined limits; an agent authorized up to $5 million is blocked above that threshold.
- JPMorgan has set no timeline for commercializing the technology.
By James Calloway
JPMorgan Chase and South Korea's Hyundai Card have completed a proof-of-concept in which AI agents executed a simulated foreign-exchange transaction on a private blockchain without direct human intervention. CNBC reported the milestone on November 17, 2026, and the两家 institutions — the two banks — have set their sights on a more demanding phase: multiple AI agents negotiating and agreeing with each other directly, targeted for the first half of 2027.
The trial is small in scope but significant in kind. JPMorgan describes it as the first private blockchain-based test between autonomous bank and client agents for corporate treasury management. In plain terms, a bank's software agent and a customer's software agent transacted with each other, on JPMorgan's Kinexys blockchain, with no human clicking approve.
What actually happened in the trial?
The scenario addressed a routine but high-volume corporate problem: treasury liquidity. A Hyundai Card AI agent requested a foreign-exchange transaction to maintain euro liquidity in its Kinexys blockchain deposit account. Kinexys is JPMorgan's blockchain unit, formerly known as Onyx, which runs deposit accounts and payment rails on distributed-ledger technology (a shared, tamper-evident database maintained across multiple nodes).
The JPMorgan side processed the request and executed the simulated FX transaction. The exercise covered execution only. Nobody's agent negotiated terms, and no real money moved — the transaction was a simulation.
The takeaway: the first phase proved bounded execution, meaning agents can complete a predefined transaction type inside strict limits. It did not prove, and was not designed to prove, that agents can handle open-ended negotiation.
What does the 2027 phase change?
The next phase, targeted for the first half of 2027, would add exactly what the first phase omitted: negotiation and agreement between multiple AI agents before payment. Wee Kee Toh, global head of Kinexys Insights at Kinexys by J.P. Morgan, framed the shift as a move beyond isolated execution into an interacting ecosystem.
"What happens when agents are negotiating, agreeing directly with each other and then making payments? What does that actually look like?" Toh asked. He identified agent-to-agent interaction as the biggest intended expansion beyond the first phase's scope.
That expansion raises the stakes considerably. If agents only execute pre-authorized instructions, accountability is straightforward: the authorizing human owns the outcome. If agents negotiate terms with each other before paying, the questions of who authorized what, and who is liable when an agreed deal goes wrong, become materially harder to answer.
The takeaway: 2027 is a testing milestone, not a product launch — JPMorgan has set no commercialization timeline.
Why governance, not code, is the bottleneck
Toh's most candid remark concerns where the real difficulty lies. Building the agent technology, he said, was relatively straightforward compared with establishing governance and internal controls. The harder problems are accountability and the risks of letting AI agents execute financial transactions without direct human oversight.
This admission deserves weight. Banks routinely deploy automation, but autonomous agents that initiate and complete transactions sit in a different risk category from scripts that follow fixed workflows. A payments agent that misreads a counterparty instruction, or that another party's agent manipulates, moves money before any human review occurs.
Kinexys has built a framework to bound that risk. Key elements reported from the trial:
- Each agent is confined to predefined transaction parameters.
- Requests outside authorized limits are automatically rejected.
- Controls enforce the agent's authority at the moment a transaction executes.
Toh described the design in blunt terms: "You are going to draw a box around the agent. The agent will never be able to do anything out of the box." He gave a concrete illustration: an agent authorized to transact up to $5 million would be blocked from executing anything above that threshold.
The skeptical read is equally concrete. A hard cap prevents size violations, but it does not by itself prevent a bad transaction within the limit — a payment to the wrong counterparty, at the wrong price, within a $5 million allowance. Negotiation between agents, the very capability targeted for 2027, is precisely where within-limit errors become possible.
The takeaway: the hard part is not making agents work; it is proving, to regulators and to counterparties, who answers when they work incorrectly.
Could this spread across Hyundai Motor Group?
The partners are also exploring whether to extend the collaboration to other businesses within Hyundai Motor Group, the sprawling automotive conglomerate that includes Hyundai Card. Toh said the group's broader operations were one reason JPMorgan selected Hyundai Card as its partner.
No specific additional projects have been announced. The statement signals intent, not commitment, and readers should treat it accordingly.
The takeaway: the partnership has room to grow, but nothing beyond the completed pilot and the 2027 test phase is on a confirmed schedule.
Why this matters beyond one pilot
Corporate treasury is a sensible first arena for agentic AI in banking. Liquidity management involves frequent, rule-bound, high-value transactions — exactly the profile where deterministic limits and automated execution plausibly add efficiency without requiring judgment. Kinexys already runs blockchain deposit accounts, so the plumbing for machine-to-machine settlement exists.
The unresolved questions are institutional, not technical, and Toh said as much. If autonomous agents negotiate and pay each other by 2027 in tests, regulators, auditors and the banks themselves will need answers on liability, authority and dispute resolution that no framework in the pilot yet provides. JPMorgan has not set a commercialization date, which is the honest position: the technology works in a box, and everything commercially interesting happens outside it.
What to watch next is whether the first-half 2027 trials demonstrate not just that agents can strike deals, but that a rejected, disputed or erroneous agent-to-agent agreement has a clear owner. Until that question has an answer, this remains a laboratory exercise — an instructive one, but a laboratory exercise nonetheless.
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles
Related articles
- AI Agents Break Financial Rules 90% of the Time Under Pressure
- JPMorganChase taps OpenAI for Codex and cybersecurity
- JPMorganChase Widens Claude Access Beyond Engineers, With Hard Limits
- JPMorgan, Amazon and Kohl's Are Blocking AI Agents. Experts Say They Can't Win That Fight
- BNY Runs 125+ AI Use Cases in Production on OpenAI-Powered Platform