JPMorgan's Kinexys and Hyundai Card test AI agent in simulated FX trade
For the treasury and payments innovation lead at a regional bank
JPMorgan's blockchain unit Kinexys and South Korea's Hyundai Card completed an initial proof-of-concept in which a Hyundai Card AI agent started a simulated foreign-exchange transaction on Kinexys' blockchain infrastructure without direct human involvement, according to a CNBC report cited by ETCIO on Nov. 18.1
The simulated trade was designed to help Hyundai Card maintain euro liquidity in its blockchain-based deposit account, according to the ETCIO account of the CNBC report.1 Kinexys built a framework that confines AI agents to authorised parameters and automatically rejects transactions that exceed their permissions, ETCIO reported; Wee Kee Toh, Global Head of Kinexys Insights at JPMorgan, said of the design: "You are going to draw a box around the agent."1
As an example, an agent authorised to execute transactions worth up to $5 million could not complete any payment above that threshold, according to the report.1
The companies are targeting the first half of 2027 for a phase that tests agent-to-agent interaction, where agents negotiate, agree terms and make payments, Toh told CNBC, according to ETCIO.1
Toh said developing the technology was not particularly difficult, and that the greater hurdle was accountability and managing risk when AI systems move money without direct human supervision, per the report.1
JPMorgan has not announced a commercial deployment, and the companies are evaluating whether the work could extend to other businesses within Hyundai Motor Group, according to ETCIO.1
Limits on amount and permissions can be enforced by the system, but the sources do not say how accountability is assigned when two agents agree terms with each other. Banks and corporate clients considering agent-initiated payments will need to settle that question in their own controls and contracts.
Kinexys and Hyundai Card are targeting the first half of 2027 for the phase testing autonomous agent-to-agent interactions.1
Our take. JPMorgan says building the agents was the easy part, and deciding who answers for them was the hard part. By the first half of 2027, your corporate clients will ask you to answer it.
Action. Corporate banking legal: Draft per-agent dollar limits, kill-switch rights and liability terms for client agent-initiated FX payments before clients ask.
What our records say
JPMorgan Chase built a firm-wide generative-AI platform, LLM Suite, with agents on top and a technology budget of nearly $20 billion a year, and it reports roughly 250,000 employees with access and almost 1,000 AI use cases.12 The bank said it expects to deploy agents that run autonomously for hours before year-end.1 Large language models screen payment validations, reducing false positives and improving queue management, and account validation rejection rates fell 15–20%.3
From our records
- JPMorgan Chasebank · latest Card-number block reported for outside AI agents (Nov 2026) · vendors Mitek, Kasisto, ChatGPT Enterprise · enforcement 5 actions since 2016; latest CFPB civil action (lawsuit, dismissed with prejudice) (Dec 2024)