The Bank for International Settlements published a working paper assessing whether a generative artificial intelligence agent can perform high-level intraday liquidity management and payment prioritisation for participants in wholesale payment systems. Prompt-based simulations using ChatGPT’s reasoning model indicate that, even without domain-specific training, a general-purpose large language model can replicate key prudential cash-management practices by preserving liquidity buffers while limiting settlement delays. The paper tests stylised real-time gross settlement (RTGS) scenarios covering precautionary buffering, prioritising queued payments under uncertain inflows and possible urgent obligations, and selecting initial liquidity in a multi-period liquidity–delay trade-off with explicit allocation, delay and borrowing costs. Repeated-run robustness checks suggest strong consistency in the simplest case (10 out of 10 repetitions) and slightly reduced consistency as constraints and contingencies increase (9 out of 10 and 8 out of 10 in more complex cases). In a separate interactive exercise using ChatGPT agent mode to complete a structured questionnaire, the agent executed routine choices autonomously and tended to defer to human oversight when faced with potentially anomalous payment instructions. The authors discuss regulatory and policy safeguards that central banks and supervisors may need to consider for AI-driven payment operations, including testing and reliability, transparency and accountability, human oversight, cyber resilience, and systemic risks such as synchronized behaviour and third-party concentration. Suggested next steps include exploring multi-agent dynamics, more realistic RTGS constraints, and system-wide stability considerations.

Original source View in tracker