Deutsche Bank and Ant Digital Technologies have published a joint white paper, Agentic Payments in Wholesale Finance, examining how AI agents could operate inside institutional finance. Released in Hong Kong on November 9, the paper asks how much authority a bank should delegate to an agent while keeping its decisions secure, controlled and auditable.
Highlights#
The paper names three areas where agents could deliver real operational value. In treasury and liquidity management, agents would forecast shortfalls and surpluses, then execute compliant actions across internal transfers, investments, financing and foreign exchange. In cross-border payment orchestration, they would weigh FX rates, rails and timing to optimise cost, speed and settlement risk. In trade finance, they would interpret documentation, flag discrepancies and generate payment instructions once conditions and approvals are met.
Crucially, the authors reject unrestricted autonomy. They advocate "bounded agentic execution" and adopt a three-layer reference architecture proposed by the IMF: an intent and orchestration layer for reasoning, a control and authorisation layer that enforces deterministic rule-based limits, and a settlement layer that executes transactions. Reasoning and execution are deliberately separated.
The safeguards section reads like a checklist for agent commerce generally: explicit, bounded and revocable authorisation limits, robust agent identity binding to an accountable owner, and Know Your Agent (KYA) verification protocols. Ant Digital's Anvita Flow platform appears as the implementation case study, showing how institutions can orchestrate payment workflows with controls over agent authority held tight.
Why it matters#
When two institutions of this scale converge on revocable authority, bound identity and separated execution as the baseline for agent payments, they are describing the same trust architecture escrow and reputation layers provide for autonomous agents everywhere.