A groundbreaking study from the Bank for International Settlements (BIS) demonstrates that generative artificial intelligence (AI) agents can perform critical liquidity management functions in central banks and high-value payments systems traditionally managed by humans.
The research, carried out with ChatGPT’s o1 reasoning model in agent mode, simulated real scenarios where AI had to balance liquidity costs and risks of delay in multi-million dollar transactions.
The experiment design three scenarios that replicate real challenges in RTGS or real-time settlement systems (Fedwire, TARGET2, Lynx, etc.), the heart of the traditional financial system.
In the first scenario, the AI had only $10 of liquidity and two outstanding payments of $1 each. Faced with the possibility of an urgent order for $10, he decided to freeze everything. His own explanation made it clear why he made the decision: “I delay small payments now to preserve liquidity and be able to attend to the urgent transaction if it arrives.”
The second scenario introduced greater complexity with the probability of receiving external funds (90%) and execute urgent payments (50%). In this case, the AI processed only lower-risk transactions, demonstrating dynamic prioritization capabilities.
Tests showed that even varying probabilities from 50% to 0.1% or scaling amounts up to billions of dollars, the AI maintained its precautionary approach. However, in complex situations its consistency decreased slightly, with occasional variations in decisions.
AI is already a better treasurer than most humans, says BIS
The study proposes developing “AI assistants” for routine tasksreserving human roles for supervision and strategic decisions. The researchers project that similar systems could be tested in regulatory sandbox environments before real implementations.
“The results suggest that specific AI solutions could reduce operational costs and improve operational efficiency and safety,” the BIS report states. But he warns of limitations: the models depend on historical data and can fail in the face of extreme events or “black swans” outside of their trained experience.
The study compares this approach with traditional reinforcement learning. The authors highlight that, unlike traditional reinforcement learning (which requires thousands of simulations), generative AI achieved “excellent results with zero specific training.”
So for that level of effectiveness, the report’s authors believe that AI could save millions in tied up liquidity and dramatically reduce payment queues in RTGS systems.
Although the BIS report focuses on traditional financial systems, its findings are not surprising in the world of digital assets. This is because decentralized finance (DeFi) applications have been managing liquidity 100% automatically with pools of automatic market makers (AMM) for years. flash loans and algorithms that rebalance in seconds.
What the BIS celebrates as innovation, Uniswap, Aave and Curve have already been doing since 2020 with billions of dollars at stake, as CriptoNoticias has been reporting.