JPMorgan tests an AI agent that shops three banks for the best currency deal on its own
Built with Hyundai Card, the agent checks balances, picks the best of three quotes and sends the order to a separate Kinexys agent to execute, in a test environment for now; CNBC reports a wider phase is targeted for the first half of 2027.
By The Graze · November 18, 2026 · 5 min read

Payments
J.P. Morgan and Hyundai Card have completed a proof of concept in which AI agents handled a treasury funding request, including checking balances, comparing foreign-exchange quotes and instructing a payment. The official Kinexys paper is blunt about the limits: the test "runs in a test environment that is integrated with back-end systems; it is not a live product, and no real value is transacted." A broader phase, involving multiple agents, is targeted for the first half of 2027, Wee Kee Toh of Kinexys told CNBC.
What was tested
The official scenario is a currency conversion to fund an operating account. Hyundai Card deploys a treasury-management agent whose standing objective is to keep a minimum balance in a Frankfurt euro account, funding any shortfall by selling pounds from a named London account. The agent checks the balance on a defined cycle. When the balance falls below the target, it requests quotes from a panel of three banks, selects the most competitive and instructs the winning counterparty to execute.
Planning and quote comparison sit with the Hyundai Card agent. Execution sits with a separate agent operated by Kinexys. "This separation between the agent's decision-making and the authority to execute provides an infrastructure-level control," the paper says. Hyundai Card's own account is consistent with that split: its agent checked the balances and compared bank FX quotes, then instructed the Kinexys agent to convert and remit. The Kinexys agent verified the client's agent's identity and authority, then executed requests inside the mandate, or asked a staff member for approval or refused those outside it.
On every request, the bank runs five checks before any money would move: whether the agent is genuine, who it is acting for, whether the request is inside the mandate (including the currency pair, size, limits and operating window), whether the request "makes sense" given the agent's stated objective, and whether the goal, plan, action, checks and outcome have been written to an immutable audit trail. Only when all five pass is the instruction released.
The paper is careful about what this does and does not show. The scope is a single funding objective, a single currency pair and direction, and two named accounts. The client agent was deployed inside the bank's own "agent-as-a-service" environment, so the test did not first have to solve cross-institutional network connectivity. "The POC tests only a single agent-to-agent interaction based on a single bank-client relationship, and not a multi-agent multi-bank multi-client environment."
What comes next, and what is still missing
Toh told CNBC that the next step is to test what happens "when autonomous agents move beyond execution and begin interacting with each other in an ecosystem," and that this is "the biggest part of what we want to get as an expansion of the phase one scope." The companies are targeting the first half of 2027 for that more complex phase, CNBC reported. The official paper is less specific on timing. It says Kinexys plans to expand the proof of concept with Hyundai "to explore more complex multi-agent multi-party scenarios," including connected vehicles and fleet management. Hyundai Card said it plans to extend the co-operation into new areas, including in-vehicle payments.
There is no commercial timetable. "Technology-wise, it's really not the biggest hurdle," Toh told CNBC, pointing instead to accountability and to the risks of allowing agents to execute without direct human oversight. "You are going to draw a box around the agent. The agent will never be able to do anything out of the box." The $5 million figure that has circulated is Toh's example of a ceiling, not a limit disclosed for this test.
A mandate approved with one institution, the paper notes, is "not currently meaningful to another." The binding constraint, it argues, is agreement on common standards, not technical feasibility. The proof of concept "does not establish a definitive legal or liability framework."
| Item | What can be verified | Source |
|---|---|---|
| The test | A proof of concept in which a Hyundai Card AI agent requested FX funding and a Kinexys agent executed, escalated or denied the request under a bank-enforced mandate | Kinexys paper, "Trust in the Agentic Economy," 2026; Hyundai Card newsroom, 18 November 2026 |
| Real money | "Not a live product, and no real value is transacted." Hyundai Card described simulated transactions between virtual overseas affiliates | Official Kinexys paper; Hyundai Card newsroom |
| The scenario | Maintain a minimum balance in a Frankfurt euro account by selling pounds from a named London account; quotes from a panel of three banks | Official Kinexys paper |
| Three outcomes | Automatic execution inside the mandate and below a threshold; escalation above the threshold; denial outside the mandate. All three were exercised | Official Kinexys paper |
| Next phase | Targeted for the first half of 2027, involving multiple AI agents, Toh told CNBC. The official paper does not give that date | CNBC, 17-18 November 2026; official paper on the expansion plan |
| Commercial use | No timetable set | CNBC, quoting Toh |
Compiled by The Graze from Kinexys by J.P. Morgan's 2026 paper "Trust in the Agentic Economy," Hyundai Card's 18 November 2026 newsroom statement, and CNBC's 17-18 November 2026 report. The 2027 date is CNBC's, citing Toh.
Our take
The sentence that matters in this announcement is not that AI agents can "pay each other with no human in the loop." That is a useful shorthand for a pitch, and a poor description of what was actually run. What J.P. Morgan and Hyundai Card have shown is narrower, and more interesting for being so. In a boxed test environment, with no real value moving, a client agent can be given a written mandate, a bank agent can check it at machine speed, and the result can be an automatic payment, a hold for a person, or a refusal. The humans have not left the system. They have moved upstream, to the drafting and approving of the box, and downstream, to the exceptions.
That is a serious design. It is also not yet a market. A single currency pair, two named accounts, one bank-client relationship, and an agent running inside the bank's own environment is a laboratory, not a network. The paper itself says a mandate at one institution is not meaningful at another, and that the hard problem is standards, liability and the law, not software. Toh said the same thing in plainer language. Anyone telling you that treasury is about to be handed to unsupervised agents is selling past the evidence. Anyone telling you that a simulated FX ticket is a nothingburger is selling short of it. The honest position is the one the bank already printed: this is buildable today, it transacted no real value, and the next, harder test — agents dealing with one another, some time in 2027 if the CNBC timetable holds — is the one that will show whether the box still holds when the other side has a mind of its own.
Sources
- Kinexys by J.P. Morgan, Trust in the Agentic Economy: Governing AI agents operating under delegated authority, 2026
- Hyundai Card newsroom, Hyundai Card successfully completes an 'AI-agent-based financial transaction PoC' with Kinexys by J.P. Morgan, 18 November 2026
- CNBC, JPMorgan is working on AI agent transactions that don't involve humans, 18 November 2026







