ChatGPT o3 Handles Simulated Liquidity Choices, but Consistency Slips in Harder Cases
The simulated wholesale-payments exercise defines a narrow starting point for delegated cash management: routine payment priorities, with human oversight for potentially anomalous patterns.
Listen to this story
The audio brief
Story brief
3 key pointsA controlled BIS–Bank of Canada experiment found ChatGPT o3 could prioritize payments and preserve liquidity in routine simulated treasury scenarios without domain-specific training. Its reliability weakened as probabilistic inflows and competing obligations made decisions harder, and anomalous patterns were escalated to people. The result supports narrowly scoped agent deployments rather than unsupervised treasury...
- 01
The o3 agent delayed two smaller payments to preserve cash for a potentially imminent large, urgent payment.
- 02
Consistency declined as scenarios introduced probabilistic inflows and more competing priorities.
- 03
In finance-leader polling, 63% cited productivity, 45% automated reporting, and 35% forecasting and planning as likely technology impact areas.
AI agents are moving closer to decisions that determine which payments get made and when. In a simulated wholesale payment system, an agent using ChatGPT’s o3 reasoning model handled routine intraday liquidity choices autonomously, as treasury teams and banks pursue targeted operational deployments.
A controlled test of payment priorities
The Bank for International Settlements and Bank of Canada paper put the o3-based agent through stylized cash-manager scenarios without domain-specific training. The assignment was not simply to describe liquidity risk: it required the system to choose among pending payments as available cash and payment urgency changed.
In one scenario, the agent delayed two smaller payments when a large urgent payment might arrive shortly afterward, preserving liquidity in a way the paper characterized as consistent with prudential cash-management practice. It completed routine prioritization exercises on its own, but referred potentially anomalous payment patterns for human oversight.
The result is a meaningful but bounded demonstration. The agent adjusted when scenarios added probabilistic inflows and competing priorities, yet its consistency fell somewhat as the trade-offs grew more complex.
In polling of finance leaders from 26 countries and more than 60 industries, 63% expected technology to affect employee productivity most over the next 12 months.
Forty-five percent named automated reporting as an area expected to be affected by technology.
Only 6% said they did not expect AI to have a material impact.
The response is targeted, not wholesale
The same polling put financial forecasting and planning among the expected areas of change, cited by 35% of respondents. Sara Castelhano, J.P. Morgan Payments’ head of U.K., Europe and Canada Treasury Services, said AI and tokenization are shifting from pilots to targeted rollouts in forecasting, payment operations and controls.
Castelhano tied scale to reliable data, clear ownership and governance. Those conditions make deployment operationally accountable, rather than leaving a model to operate without defined responsibility.
Banks are assigning agents different slices of the workflow
- Goldman Sachs is developing Anthropic Claude-powered autonomous agents for trade accounting and client onboarding.
- Lloyds Banking Group has committed to enterprise-wide agentic AI deployment in 2026 and expects £100 million in value from automating fraud investigations and complex complaints.
- Lloyds’ plan reserves people for the most nuanced escalations while routine cases are diverted to AI.
Sources
- pymnts.comAI Agents Help Treasurers Move Faster | PYMNTS.com
