Skip to content
Crypto.world

The Ledger

Hyundai Card and JPMorgan test AI treasury agents

A proof of concept used separate AI agents to check account balances, compare foreign-exchange quotes and request a currency conversion and transfer. The test ran in a controlled environment, moved no real value and used bank-side checks to approve, escalate or deny requests.

Hyundai Card and JPMorgan test AI treasury agents

AI agents handle a simulated treasury flow

A proof of concept by Hyundai Card and Kinexys used separate AI agents to monitor account balances, compare foreign-exchange quotes and request funding transactions. The Kinexys report describes the test, which ChosunBiz also reported on Nov. 18, 2026.

The exercise modeled treasury funding between two named test accounts, a Frankfurt euro account and a London sterling account. The Kinexys report says it ran in a test environment integrated with back-end systems, not as a live product, and no real value was transacted.

Balance checks and currency quotes

Hyundai Card’s scenario tasked its treasury agent with keeping a Frankfurt euro account above a minimum balance. When the balance fell short, the agent was to sell pounds from a named London account. It checked the balance on a defined cycle, requested quotes from three banks, selected the most competitive and instructed the winning counterparty to execute.

A separate agent operated by Kinexys handled execution after the Hyundai agent submitted the payment request. The report describes this division between the agents’ roles as part of the transaction flow: the Hyundai agent planned and selected an option, while the Kinexys agent processed the request.

Identity, authority and outcomes

The client registered the agent’s public cryptographic key with Kinexys alongside an agent identifier and the legal entity it represented. The agent used its private key to authenticate and sign messages. Authority was set out in a separate machine-executable mandate: the agent drafted a proposed version, but the principal reviewed and approved it before Kinexys bound it to that agent’s public key.

The mandate listed permitted actions, escalation conditions and guardrails. Before release, Kinexys checked the agent’s credential, the principal it represented, whether the request fit the mandate, whether it appeared coherent and unusual, and whether the action was recorded in an audit trail. Requests within the mandate and below the escalation threshold executed automatically; requests above it were escalated for principal approval, while requests outside approved parameters were denied.

The report says the proof of concept exercised all three outcomes. In one denial test, Kinexys blocked a transaction request using a currency pair not included in the mandate. A separate escalation test held a funding amount above the autonomous threshold until the principal approved it.

Limits of the test and next steps

The proof of concept covered one funding objective, one currency pair and direction, and two named accounts. The report says it tested one agent-to-agent interaction in a single bank-client relationship, not a multi-bank, multi-client environment. The agent also operated within the bank’s agent-as-a-service environment.

Kinexys says it plans to expand the proof of concept with Hyundai to more complex multi-agent, multi-party scenarios, including connected vehicles and fleet management. The paper identifies shared standards for verifiable agent identity, portable mandates and interoperable controls as next steps, and says a shared trust-anchor model remains an open industry question.

Sources

  1. Original source Kinexys by J.P. Morgan jpmorgan.com

Other pages read

  1. ChosunBiz biz.chosun.com

Related stories

Explore

More articles