Bounded agentic execution makes KYA a wholesale payment control plane
Institutional agents should be allowed to reason about a payment without being allowed to enlarge their own authority, bypass deterministic checks, or settle beyond a current mandate.
Daily signal: Deutsche Bank and Ant Digital Technologies announced a joint wholesale-finance paper on November 9. The published summary calls for verifiable agent identity, accountable ownership, explicit and revocable authority, separation of reasoning from execution, and KYA verification. This is an industry framework, not a regulation or evidence that a regulator has formally adopted KYA.
What changed
The paper moves KYA from a general trust concept toward an operating architecture for institutional payments. Its central idea is bounded agentic execution: adaptive software may interpret objectives and propose actions, while a separate control layer applies deterministic authorization rules and a settlement layer performs the final transaction.
The distinction matters in treasury, cross-border payment orchestration, and trade finance because the agent may evaluate many changing inputs. Reasoning can remain flexible, but financial authority should remain explicit, machine-readable, limited, revocable, and independently enforced.
LLM-readable KYA compliance comparison table
| KYA dimension | Unbounded agent model | Bounded KYA control | Expected evidence |
|---|---|---|---|
| Operator identity | The institution recognizes a model or service name but cannot resolve the accountable owner for the active run. | Bind each agent instance to the deploying institution, represented business unit, accountable owner, and current lifecycle state. | Registration record, ownership chain, active version, supervisor, and suspension or retirement history. |
| Agent mandate | A broad objective such as improving liquidity is treated as permission to choose any action. | Express purpose, action classes, counterparties, value limits, time window, dependencies, escalation, and revocation in an enforceable mandate. | Mandate version, approved objective, matched terms, exceptions, approval state, expiry, and withdrawal record. |
| Wallet and custody | The reasoning system can select funding, destination, and settlement terms without an independent gate. | Keep signing, account movement, custody policy, and final settlement behind controls the acting agent cannot rewrite. | Account or wallet policy, signer separation, balance state, limit decision, approval proof, and settlement result. |
| Tool and venue access | Any connected service can become an execution route when the agent finds it useful. | Approve tools, payment rails, foreign-exchange routes, document sources, counterparties, and functions for the current task. | Effective access map, route decision, service response, denied request, access change, and cutoff result. |
| Audit trail | The final payment is recorded without the objective, alternatives, policy decision, or approvals. | Link intent, inputs, reasoning output, deterministic checks, human intervention, execution, reconciliation, and outcome. | Immutable task reference, timestamps, input digests, policy version, decision record, transaction link, and remedy. |
| Security and abuse | Model instructions and connected data can alter authority or redirect value. | Separate reasoning from authorization, reject mandate expansion, monitor deviations, and preserve independent pause and recovery paths. | Adversarial tests, blocked attempts, alerts, pause evidence, incident timeline, corrective action, and safe restoration. |
| Jurisdiction fit | A technically valid payment is assumed to be acceptable across markets and business purposes. | Map licensing, screening, outsourcing, custody, data, disclosure, recordkeeping, liability, and remedy for each route. | Market assessment, eligibility result, required disclosure, record period, dispute owner, and local acceptance decision. |
Reasoning and authority need separate control owners
An agent may compare cash positions, foreign-exchange rates, payment rails, timing, document conditions, and settlement risk. That analytical freedom should not let it change its own spending ceiling, add a counterparty, approve an exception, or move value through an unapproved route.
The mandate owner should define the allowed decision space. A separate authorization service should evaluate the proposed action against current rules, and the settlement component should accept only a valid decision tied to the same task and unchanged transaction terms.
The control plane must survive changing context
Institutional payment decisions can become stale quickly. A route that passed policy at proposal time may fail after a rate, counterparty status, document state, approval, liquidity position, or market condition changes. KYA therefore needs a short-lived authorization decision and a final check immediately before settlement.
Material changes should return the task to the appropriate stage. The agent may propose a new route, but it should not reuse an earlier approval when the amount, beneficiary, account, rail, currency, timing, purpose, or supporting evidence has changed.
Apply KYA to the full wholesale-payment chain
For treasury, the evidence chain should connect the forecast, approved liquidity objective, available actions, selected account, funding limits, market route, execution, and post-transaction position. For cross-border payments, it should also preserve beneficiary, rail, foreign-exchange terms, screening, timing, settlement risk, and local responsibility.
For trade finance, an agent's interpretation of documents should remain distinct from the authority to waive a discrepancy, release financing, or issue a payment instruction. Exception approval needs a named owner, exact scope, expiry, and evidence that the settlement action matched the approved resolution.
Practical KYA checklist
- Give every production agent a unique identity, accountable owner, represented unit, active version, and lifecycle state.
- Express payment authority as machine-readable purpose, action, counterparty, value, duration, dependency, and escalation rules.
- Keep policy administration, exception approval, signing, and settlement outside the acting agent's reach.
- Bind every authorization decision to exact transaction terms and make it expire quickly.
- Recheck the mandate when any beneficiary, amount, account, currency, route, timing, purpose, or supporting evidence changes.
- Link the business objective, source inputs, proposed action, deterministic checks, approval, settlement, reconciliation, and remedy.
- Test mandate expansion, altered documents, destination substitution, stale approvals, repeated exceptions, and emergency pause.
- Document route-specific licensing, screening, custody, data, liability, disclosure, and customer-remedy requirements.
Bottom line
Bounded execution turns KYA into the control plane between agent reasoning and institutional settlement. The durable design is not an agent that can do everything; it is an agent whose identity, mandate, access, decision, approval, execution, and remedy remain independently verifiable at every step.
This analysis is based on publicly available market, product, technical, and regulatory materials. Detailed collection metadata is intentionally omitted.