logo

BNY tests AI Treasury negotiation as Vietnam’s banking AI rules take effect

Add The Asian Banker on Google
Discover more trusted banking and financial services insights by adding The Asian Banker as a preferred source on Google.
BNY tests AI Treasury negotiation as Vietnam’s banking AI rules take effect
  • 84

Financial Technology Weekly: BNY and Microsoft completed an AI-negotiated US Treasury bill pilot, while Raiffeisen Bank International reported an 80% reduction in processing time and token consumption through redesigned AI workflows. Vietnam’s banking AI rules took effect, as regulators in the UK, Singapore and Malaysia outlined approaches to AI governance, cyber resilience and agent authorisation. JPMorgan Chase also disclosed a vendor cyberattack involving potentially exposed customer data.

BNY and Microsoft completed a pilot using AI agents to negotiate a US Treasury bill trade, retaining human approval and existing execution systems. Raiffeisen Bank International reported that shifting predictable tasks from AI agents to conventional automation reduced token consumption and processing time by 80% in tests.

Vietnam’s banking AI rules came into effect, while UK authorities outlined how firms could recover compromised AI models and agent memory. JPMorgan Chase’s disclosure of a vendor cyberattack highlighted a separate source of operational risk following an earlier software disruption.

Read more on the week’s key developments:

1. BNY and Microsoft test AI negotiation of a US Treasury trade

BNY and Microsoft announced on 17 November that they had completed a pilot in which AI agents negotiated a US Treasury bill transaction. BNY’s Agentic Negotiation Protocol verified participants and allowed the agents to exchange offers and counteroffers within predefined limits. Human traders at both firms reviewed and approved the agreed terms before the transaction was executed through existing systems.

The pilot automated negotiations between counterparties, not trade execution. Before seeking trader approval, the protocol checked that both agents had independently submitted matching terms to a central ledger. The agents could not execute the trade themselves. The pilot therefore introduced automated negotiation while retaining human approval and existing execution and booking processes.

2. Raiffeisen reports 80% reduction in processing time in AI workflow tests

Raiffeisen Bank International reported an 80% reduction in token consumption and processing time in tests of a payment enquiry workflow combining AI agents with rules-based automation. The result was disclosed in Appian’s 17 November announcement. Claus Mitterlehner, the bank’s head of smart automation, said the team initially assigned most tasks to agents, but revised the design after testing.

Predictable steps were transferred to conventional automation, with agents retained for tasks requiring judgement. Mitterlehner attributed the improvement to this redesign, although the release does not specify the benchmark configuration or absolute processing times. The findings indicate that task allocation can materially affect AI processing demands. However, the reported efficiency gains do not establish corresponding reductions in staffing requirements or total operating costs.

3. Vietnam’s banking AI safety rules take effect

The main provisions of the State Bank of Vietnam’s Circular 53/2026/TT-NHNN took effect on 15 November. Issued on 30 September, the circular establishes safety, risk management and deployment conditions for AI used in banking operations and customer financial services. Before production deployment, institutions must obtain internal approval backed by evidence of testing, risk classification and operational readiness.

Institutions must establish accuracy, reliability and performance criteria before deployment, alongside documented arrangements for monitoring, incident response and recovery. Deployment therefore requires institutions to demonstrate operational controls as well as model performance. The timetable is phased: systems already operating before 15 November generally have until 1 March 2027 to meet the circular’s requirements in Chapters II–IV, while specified provisions have later commencement dates.

4. UK regulators address AI model recovery in cyber guidance

The Bank of England, Prudential Regulation Authority and Financial Conduct Authority published AI cyber resilience practices on 19 November alongside a consultation on clarifying existing cyber risk requirements. The consultation proposes guidance rather than new rules and closes on 26 February 2027. The separate AI practices do not constitute formal supervisory expectations.

The AI document identifies recovery tasks that go beyond restarting an interrupted service. Firms may need to restore an earlier version of a fine-tuned model or agent state, or remove unsafe stored memory. It also recommends comparing model, tool-use and chat logs to identify abnormal behaviour. Recovery consequently involves validating the AI model’s information and state, not just restoring service availability.

5. JPMorgan discloses vendor cyberattack after separate software disruption

JPMorgan Chase disclosed on 19 November that service providers had suffered a cyberattack and that attackers were believed to have accessed marketing data that could include customer names and email addresses. Spokesman Joe Evangelisti said the data was held by vendors. The bank said the attack had caused no disruption to its services and that it would contact affected customers where appropriate.

The disclosure followed a separate disruption on 16 November, which JPMorgan attributed to a software bug in its own hardware. That incident delayed the bank’s transaction services and affected payments involving other banks for which JPMorgan acted as a counterparty. JPMorgan said the two events were unrelated. The incidents exposed different operational risks: the internal software failure disrupted transaction services, while the vendor attack involved the potential exposure of externally held customer information.

6. MAS develops a protocol for identifying and authorising AI agents

The Monetary Authority of Singapore announced a new phase of its Safeguards for Agentic Finance at Runtime programme on 19 November, including work on a common “know your agent” protocol. Managing director Chia Der Jiun said the protocol would identify agents, whom they represent and the authority and mandates granted to them. MAS also plans an Agentic Money Sandbox with the Future of Finance Institute for treasury and payment applications.

Chia identified a gap in current banking controls: some customers give agents their bank login credentials, leaving institutions unable to establish whether an agent is properly authorised or acting within its instructions. The proposed protocol would enable the exchange of information on an agent’s identity, authority and mandate, allowing financial institutions to establish dedicated secure channels for agent access. In the planned treasury tests, corporates would set the mandates, while participating banks would determine the controls for accepting instructions.

7. Four New Zealand banks join a network to share account fraud signals

ANZ New Zealand, ASB, BNZ and Westpac New Zealand joined the planned BioCatch Trust New Zealand network on 18 November. The network will share behavioural, device and selected account signals so a sending bank can assess risk indicators associated with a recipient account before releasing funds. BioCatch said implementation would take several months; the New Zealand network is not yet operational.

The network is designed to complement Confirmation of Payee, which checks whether the supplied name matches the receiving account but does not establish whether that account is being used legitimately. The new network would give the sending bank additional risk information from the receiving side, including in scams where customers have been persuaded to authorise payments. Banks would retain responsibility for deciding whether to suspend or investigate transactions.

8. UOB explores multilingual AI for customer calls and compliance reviews

UOB and A*STAR signed a memorandum of understanding on 18 November to explore applying the MERaLiON multilingual AI model to relationship managers’ customer conversations. The work will begin through UOB’s Project AURA and examine how AI could support live calls and subsequent compliance reviews. Group chief financial officer Leong Yung Chee said the collaboration would examine language and cultural nuances in banking conversations. The announcement sets out an exploratory project rather than a completed deployment.

The proposed workflow links language recognition to verification of customer communications. MERaLiON supports switching between languages within a conversation, while the application could prompt staff about product information and risk disclosures and review whether these were communicated clearly. The intended application extends beyond transcription to assessing whether customer-facing disclosures were delivered. UOB says relationship managers would remain accountable for their conversations and advice.

9. Bank Negara Malaysia retains existing rules for financial-sector AI

Bank Negara Malaysia said on 19 November that it would not introduce AI-specific regulation at this stage, publishing feedback on its August 2025 discussion paper after receiving 113 responses. AI-enabled financial activities remain subject to existing requirements, including governance, risk management, consumer protection and Shariah compliance. The central bank will expand its supervisory focus on AI adoption, governance and reliance on technology providers.

The decision maintains the technology-neutral approach outlined in the discussion paper. The practical requirement is to translate existing obligations into controls for each AI use case. BNM’s feedback calls for stronger safeguards for higher-risk applications, including testing, monitoring and human oversight, supported by a documented rationale and periodic reassessment. Its list of applicable policies includes technology risk management and outsourcing, bringing AI governance within existing operational risk and third-party oversight frameworks.

10. Coop Norge moves 2.4 million member accounts to a cloud core

Coop Norge completed the migration of all 2.4 million member accounts to Mambu’s cloud core banking platform, implementation partner Knowit announced on 17 November. The Norwegian group, owned by 56 cooperatives, now manages membership and voluntary deposits on the platform. Knowit acted as lead systems integrator.

The accounts cover a combination of membership administration and savings rather than the full range of banking services. Coop Oppdal, one of the cooperatives, says its savings arrangements cover purchase dividends, bonuses and ordinary deposits, with a member account created automatically on joining. The project therefore combines membership administration with deposit recordkeeping. Its scope is narrower than a bank-wide core migration involving lending and payments, focusing instead on member savings and loyalty administration.