Hong Kong – Deutsche Bank and Ant Digital Technologies have published a joint white paper arguing that AI agents in wholesale finance should operate within strict limits rather than with unrestricted autonomy.
The paper examines how autonomous AI could be applied in institutional finance and which governance frameworks it would require. The two firms announced the publication on 9 November.
The paper arrives as AI moves beyond generating insights towards systems that can interpret objectives, weigh alternatives and act independently. Consequently, it poses a central question for wholesale institutions: how much authority should be delegated to agents, and how can their decisions remain secure, controlled and fully auditable?
Sabih Behzad, Head of Digital Assets and Currencies Transformation at Deutsche Bank, addresses the broader potential of agentic AI in his foreword, “The potential value lies not simply in executing payments automatically, but in improving how those decisions are made, coordinated, and controlled,” he stated.
Zhuoqun Bian, President of Ant Digital Technologies, takes a security-first view of adoption in her foreword, “Realising this potential in wholesale finance depends on foundations that are neither optional nor purely technical. Agents need verifiable identity, bound to an accountable owner. Their authority must be explicit, machine-readable, bounded, and revocable. Decision-making, control, and execution must be separated, so that an agent may propose but cannot extend its own authority,” she stated.
The paper identifies three areas where agentic payments could deliver operational efficiencies. The first is treasury and liquidity management, where agents could forecast liquidity shortfalls and surpluses, assess response options and execute compliant actions across internal transfers, investments, external financing and foreign exchange.
The second is cross-border payments, where agents could weigh variables such as exchange rates, payment rails and execution timing. In doing so, they could optimise transaction costs, speed, availability, regulatory compliance and settlement risk.
The third is trade finance, where agents could interpret complex documentation, identify data discrepancies and coordinate exception handling. They could also generate financing or payment instructions once conditions are met and approvals obtained.
However, the paper also highlights the governance, accountability and risk management challenges that accompany greater AI autonomy. It therefore advocates “bounded agentic execution” rather than unrestricted autonomy.
To support the move from early experimentation to institutional adoption, the paper draws on a three-layer reference architecture proposed by the International Monetary Fund (IMF). The first layer, intent and orchestration, oversees agent reasoning, decision-making and workflow orchestration.
The second layer, control and authorisation, enforces deterministic, rule-based constraints and keeps actions within predefined authorisation boundaries. The third, settlement, carries out the final execution of financial transactions.
At the centre of the design is a deliberate separation between AI-driven reasoning and transaction execution. The safeguards outlined include explicit, bounded and revocable authorisation limits, robust agent identity binding and Know Your Agent (KYA) verification protocols.
The paper also presents Anvita Flow, an agentic platform developed by Ant Digital Technologies, as an implementation case study. According to the paper, the platform illustrates how institutions can orchestrate payment workflows in real-world operating environments while maintaining controls over agent authority and execution.
