AI news ·
AI agents ace intraday cash management in RTGS without special training, BIS finds
BIS trials show LLM agents can handle routine RTGS cash flows, balancing speed and liquidity like a seasoned operator. Pilot with guardrails, track KPIs, scale on results.
AI agents are ready for routine cash management
New research from the Bank for International Settlements (BIS) suggests generative AI agents can handle core cash-management tasks in real-time gross settlement (RTGS) systems without specialized training.
Using prompt-based experiments with ChatGPT's reasoning model, researchers simulated payment flows with liquidity shocks and competing priorities. The agent maintained precautionary buffers, reprioritized payments under tight constraints, and balanced settlement speed against liquidity usage-much like a seasoned treasury operator.
The takeaway: routine cash-management work can be automated with general-purpose large language models. That could reduce operating costs and improve intraday liquidity efficiency for institutions running high-volume payment operations. See the BIS summary for context: BIS working paper.
Why this matters for management
- Lower handling costs on repetitive intraday decisions (queuing, throttling, netting windows).
- Faster settlement with fewer liquidity drains during peak periods.
- Consistent execution under stress, with clear rules and audit trails.
- 24/7 coverage that supports human teams during off-hours and end-of-day crunch.
- Better use of skilled staff for exceptions, forecasting, and strategic liquidity planning.
What to pilot in the next 90 days
- Start with a contained use case: RTGS queue re-prioritization or buffer advisories for high-value payments.
- Define KPIs before you test: average queue time, percentage settled within window, peak liquidity drawn, and alert precision.
- Prototype with prompt-only agents. No deep integration at first-run in shadow mode against historical and same-day data.
- Set guardrails: minimum buffer levels, escalation thresholds, and hard stops for exceptional payments.
- Review daily: compare agent recommendations vs. human actions, then refine prompts and rules.
- Validate regulator expectations: auditability, explainability, and clear operator override.
- Plan for production latency limits. The agent must respond within your RTGS and messaging SLAs.
Governance and risk controls
- Human-in-the-loop for high-value or high-risk decisions; automated for routine flows.
- Deterministic fallbacks: if confidence is low or data is incomplete, revert to predefined rules.
- Stress-test across adverse scenarios: payment surges, system delays, and liquidity shocks.
- Strict data boundaries and access controls; log every prompt, decision, and override.
- Monitor model drift and updates; re-approve prompts and thresholds after each change.
- Incident playbook: escalation paths, rollbacks, and communication protocols.
Build the capability
Stand up a small cross-functional squad: treasury ops, payments IT, risk, and compliance. Keep it simple-clear prompts, clear rules, and a tight feedback loop.
- Document decision policies and translate them into prompts plus guardrails.
- Instrument dashboards for KPIs and exceptions.
- Train operators on reviewing and overriding agent recommendations.
- If you need structured upskilling, explore focused AI courses by role: Complete AI Training - Courses by Job.
Bottom line
General-purpose LLMs can already handle routine intraday liquidity decisions under pressure. Treat them as automation teammates: start small, measure everything, and scale where the numbers prove out.