BNY and Microsoft Test AI Agents Negotiating U.S. Treasury Bills
Key Facts
BNY and Microsoft said on November 17, 2026, that they had completed an initial test of AI agents negotiating a U.S. Treasury bill trade, testing the use of agents in financial-market transactions. BNY said agents developed by the two companies negotiated securely, while human traders reviewed and approved the proposed terms. The agents did not execute or record the trade on behalf of either company. Execution and recording took place through existing internal trading systems. BNY said the approach could make negotiations more consistent and allow traders to focus on higher-value work.
The two companies used BNY’s agent negotiation protocol, designed to verify participants and connect them securely. The protocol allows agents to exchange offers and responses, then verify that both parties understand the proposed terms in the same way. Microsoft Treasury gave its agent specific trading instructions and connected it to BNY through a secure gateway. Before negotiations began, BNY reviewed controls covering the counterparty, instrument, trade size, settlement terms and available pricing. Once terms were reached, they were independently compared with the protocol’s central record before traders at both companies reviewed and approved them.
The companies set limits on the agents’ authority before negotiations began. BNY said controls for each participant are based on predefined rules and AI security checks, keeping agents within a narrow scope based on approved pricing inputs. Human trading teams retain final decision-making and action authority, while agents cannot execute trades on behalf of either company. BNY’s global markets chief told Bloomberg that the scope of what agents can do and the parties they can engage with is set in advance. The agents can negotiate back and forth within those limits, she added.
The process records participants, exchanged messages, control checks and human decisions, BNY said, allowing the negotiation to be reviewed and reconstructed. The record provides a way to track how agents moved from offers to terms approved by traders. BNY says agent-based negotiation could occupy a space between electronic trading and high-touch dealing. Bloomberg quoted the bank’s global markets chief as saying the company is not proposing the disappearance of high-touch trading. Giving computers the ability to deliberate and seek common ground is the new aspect of the test, the executive added.
BNY places the test within a broader effort to adopt AI across the bank. In October 2026, the company said its enterprise platform Eliza had more than 300 AI solutions in production and that AI had helped write 40% of code during the second quarter of 2026. Its 2025 annual report also said the bank was focused in 2026 and beyond on expanding agentic solutions and more autonomous workflows, while keeping multi-agent systems supportive of human decision-making. As of September 30, 2026, BNY had $63.4 trillion in assets under custody and/or administration, while assets under management totaled $2.2 trillion. The scale of the bank’s business puts the negotiation test in the context of expanding technology use across its institutional operations.
Bloomberg described the trade as a proof of concept and said the key test is whether the approach can be scaled across different asset classes and in both calm and volatile markets. BNY plans to continue testing and expanding the protocol beyond the initial bilateral use to assess other negotiation workflows, without announcing a timetable for broader deployment. Future applications could include repurchase agreements and contractual negotiations, while areas BNY is also examining include foreign-exchange payments and collateral terms. These remain potential uses under study and are not part of the current test as announced. Further testing and expansion will determine the protocol’s suitability for other negotiation workflows and assets.