Deutsche Bank and Ant Digital Technologies published a joint white paper on using AI agents in corporate payments and financial operations.
German banking group Deutsche Bank and technology company Ant Digital Technologies released a study titled Agentic Payments in Wholesale Finance. The paper describes how autonomous AI agents could be used in corporate liquidity management, cross-border settlement, and trade finance, and proposes an architecture for controlling transactions carried out by artificial intelligence (AI).
The authors examine the shift from using AI for data analysis to systems that can independently assess financial tasks, choose how to handle them, and initiate transactions. They name limited autonomy for AI agents as the main condition for adopting such technology in the institutional sector, with mandatory separation of permissions, authorization mechanisms, and the ability to audit every action.
The paper identifies 3 main applications for agentic payment systems:
- Liquidity and treasury management. AI agents can forecast liquidity shortfalls or surpluses, assess available instruments, and execute permitted transactions, including internal transfers, investments, external financing, and foreign exchange deals.
- Cross-border payments. Autonomous systems can choose the best transaction routes based on exchange rates, available payment infrastructure, execution time, fees, regulatory requirements, and settlement risks.
- Trade finance. AI agents can analyze trade documents, spot data discrepancies, handle exceptions, and generate payment instructions once the set conditions are met and the necessary approvals are obtained.
As a basis for these solutions, Deutsche Bank and Ant Digital Technologies propose a model called bounded agentic execution, under which AI agents carry out transactions only within predefined permissions.
The proposed architecture builds on the International Monetary Fund’s (IMF) 3-layer model for agentic finance infrastructure:
- The first layer handles task interpretation, decision-making, and process coordination.
- The second verifies permissions and enforces the established rules.
- The third executes the financial transactions.
Separating identification, authorization, and execution in this way prevents an agent from expanding its own permissions. An AI agent can propose a transaction, but its execution has to pass through independent control and authorization mechanisms. The study also calls for Know Your Agent (KYA) protocols to identify AI agents and link them to the owners responsible for them.
Sabih Behzad, Head of Digital Assets and Currencies Transformation at Deutsche Bank, said the main value of agentic solutions lies in better decision-making, coordination, and control of operations, in addition to payment automation. Zhuoqun Bian, President of Ant Digital Technologies, said agents need verifiable identities, clearly defined accountability, and separation between decision-making and transaction execution.
The paper presents a concept for agentic payments and the risk management mechanisms they require, but it doesn’t include quantitative estimates of the economic impact, testing results, or timelines for production deployment.
Earlier, Ant International and Deutsche Bank agreed to jointly develop and deploy comprehensive solutions for cross-border payments and global liquidity management. Ant International also took part in testing Agent Pay for Machines (AP4M), an automated settlement platform developed by Mastercard, while Deutsche Bank joined the TransferMate fintech initiative, which combines a global payment network with local tools for collecting, transferring, and holding funds.
