TD treasury executive urges stronger controls for agentic AI adoption

Corporate treasury departments face a control challenge as agentic artificial intelligence moves from analysis into execution, according to Tom Gregory, head of treasury management for merchant and government banking at TD Bank U.S. A TD survey cited by Gregory found that nearly 80% of treasury functions still use manual or partly manual workflows, a gap he said could complicate attempts to automate payment, liquidity and risk processes.

Gregory, writing in a Finextra opinion post, said many treasury teams are being asked to work in real time while still relying on fragmented systems and human-led processes. He argued that the rise of agentic AI requires operating models and control frameworks to change before automated tools are given broader authority.

Agentic AI refers to systems that can take actions within defined objectives and constraints, rather than only producing analysis or recommendations. In treasury, Gregory said that could affect payment release, cash movement, liquidity decisions and the identification of risk events as they occur.

Execution changes the control problem

Traditional AI tools have already been used to improve visibility and reduce manual work in treasury, Gregory said. Agentic systems create a different issue because they may participate directly in execution, placing them closer to cash management and risk control.

According to Gregory, many treasury frameworks assume that a person starts and approves an action inside a defined workflow. That assumption becomes less reliable when a system can act under pre-set thresholds, especially where data remains split across systems and processes are still developing.

He said some organisations are testing AI faster than they are updating governance. In that setting, the central control question becomes how treasury teams can oversee actions made by systems operating within defined parameters.

Guardrails before scale

Gregory said treasury leaders should decide where automation is allowed to operate, which activities remain subject to human judgment and which controls must be in place before wider deployment. He pointed to thresholds for payment release, liquidity transfers and exception handling as areas where rules, approval structures and escalation paths should be defined in advance.

He compared the approach to fraud prevention, where technology and controls are supported by verification steps, clear processes and intervention when activity falls outside expected patterns. Gregory said AI adoption requires the same emphasis on transparency, accountability and oversight.

In his view, the objective is to use automation in areas where authority and limits are clear, rather than to remove human decision-making across treasury. He said governance, accountability and operating processes need to be aligned with the technology so that AI deployment does not create unmanaged risk.

Gregory identified cash positioning, payment activity and decision support as areas where treasury teams can test agentic systems under controlled conditions. These use cases, he said, may allow organisations to measure benefits while observing how automated systems behave inside guardrails before applying them to more complex or higher-risk decisions.

He said firms that set controls early are more likely to gain value as automation spreads across treasury operations. The broader issue, according to Gregory, is whether treasury departments can govern decisions that systems may increasingly execute on their behalf while preserving accountability and trust.

This story draws on original reporting from Finextra Research.