Deutsche Bank and Ant Digital Technologies have outlined a framework for using AI agents in wholesale payments without giving software unrestricted control over money movement, arguing that treasury, cross-border payments and trade finance should separate machine reasoning from final transaction authorization.
Their joint white paper, Agentic payments in wholesale finance, released on Nov. 9, proposes bounded agentic execution as a model for deploying AI in institutional financial workflows. Under that approach, agents could assess data, recommend actions and coordinate complex tasks, but their authority would be explicitly defined, limited and revocable before payments are settled.
The paper places the debate over AI in finance around operational control rather than the technology's ability to generate insights. AI agents are described as systems capable of interpreting objectives, evaluating choices and carrying out multi-step tasks. In payment operations, that capability raises a practical governance question: when an agent takes action, institutions need a clear record of who granted its authority, what it was permitted to do and whether the action complied with preset rules.
Behzad, Deutsche Bank's head of digital assets and currencies transformation, addresses that issue in the paper's foreword, focusing on decision-making and control in payment execution. Bian, president of Ant Digital Technologies, sets out several requirements for agent deployment, including a verifiable identity for each agent connected to an accountable owner, explicit authorization that can be withdrawn, and a separation between an agent proposing an action and the system granting permission to execute it.
Treasury agents could manage liquidity decisions
The first application concerns corporate and institutional treasury operations, where firms manage cash across accounts, currencies and jurisdictions. According to the paper, an AI agent could forecast liquidity shortages or surpluses, assess potential responses and carry out approved actions across internal fund transfers, investments, external financing and foreign-exchange operations.
That could reduce the manual work involved in monitoring balances and comparing funding options across a large organization. A treasury desk handling multiple entities and currencies may face a recurring choice between moving internal cash, borrowing externally, investing excess funds or exchanging currencies. The paper envisions agents handling the analytical and workflow-heavy parts of that process within rules set by the institution.
The distinction between identifying an option and executing it runs through the proposed design. An agent might determine that transferring liquidity between two internal accounts is the most efficient response to an expected shortfall, but the transfer would remain subject to pre-authorized limits and deterministic controls.
Cross-border payments put execution choices in focus
The second use case is cross-border payment orchestration, an area where payment costs and settlement outcomes can vary depending on foreign-exchange pricing, available payment rails, local operating hours and regulatory requirements.
Deutsche Bank and Ant Digital Technologies say agents could evaluate foreign-exchange rates, route availability and timing to select an execution path that balances transaction costs, speed, availability, compliance requirements and settlement risk. The proposed role goes beyond choosing the cheapest route. A low-cost route may be unsuitable if it creates delays, fails a compliance check or adds settlement uncertainty.
For banks and large companies, this would turn payment routing into a continuously assessed decision rather than a fixed operational process. Yet it also creates a stronger need for controls, since payment instructions may involve several currencies, financial institutions and legal jurisdictions. The paper's model seeks to keep the AI component focused on assessment and orchestration while placing final transaction execution behind rule-based authorization.
Trade finance agents would handle document exceptions
Trade finance is the third use case identified in the document. The process often requires banks and companies to examine shipping records, invoices, letters of credit and other documents before financing is released or payments are made.
The paper says agents could interpret trade documentation, identify inconsistencies in the data, coordinate the handling of exceptions and generate financing or payment instructions once contractual conditions have been met and approvals secured. Documentation discrepancies are a longstanding operational challenge in trade finance, where even small differences between records can delay a transaction.
An agent could help organize those workflows and direct exceptions to the appropriate teams, according to the paper. The proposed model does not remove approval steps: payment or financing instructions would follow only after applicable conditions and authorizations are satisfied.
Three layers separate reasoning from settlement
The document references a three-layer framework attributed to the International Monetary Fund: intent and orchestration, control and authorization, and settlement.
The intent and orchestration layer is where an AI agent would perform adaptive reasoning, make decisions and coordinate workflows. This is the part of the system designed to interpret objectives and examine available options.
Control and authorization would operate differently. The paper describes it as deterministic and rule-based, enforcing boundaries that have been established in advance. Those boundaries could include transaction limits, approved counterparties, permitted currencies, designated payment rails and escalation requirements.
Settlement forms the final layer, where the actual financial transaction is executed. By separating these functions, the framework aims to prevent an AI system's reasoning process from directly becoming an unrestricted instruction to move funds.
The paper also calls for identity binding, ensuring that every agent can be linked to a responsible entity or owner. Its proposed know your agent, or KYA, protocols would extend familiar identity and accountability concepts into automated workflows. Along with explicit and revocable authority limits, the measures are intended to create an auditable trail for agent activity.
Ant Digital Technologies' Anvita Flow platform is included as a case study. The platform is presented as an agentic-payment system designed to run workflows while maintaining controls over agent authority and execution.
The proposal reflects an emerging institutional view that AI agents may be most useful in payments when their autonomy is narrow enough to be governed. For wholesale finance, where a flawed instruction can involve substantial sums and multiple counterparties, the framework gives banks a way to automate decision-heavy processes without placing final settlement beyond predefined controls.
Explore how banking AI agents evolve into secure trading tools with Toobit’s AI copy trading guide for practical applications.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
