Book a demo
Topics · GOVERNANCE-AGENTS

Banking Agent Identity How It Works | Moring

Agent identity in banking works by registering every digital assistant as a unique, credentialed identity distinct from the humans it serves, then enforcing least-privilege access at runtime. Delegation tokens carry actor claims that bind each action to both the user and the agent, while admission-control mechanisms validate identity, capability scope, delegation chain, and policy compliance before any system state changes.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Agent identity in banking works by registering every digital assistant as a unique, credentialed identity distinct from the humans it serves, then enforcing least-privilege access at runtime. Delegation tokens carry actor claims that bind each action to both the user and the agent, while admission-control mechanisms validate identity, capability scope, delegation chain, and policy compliance before any system state changes.

How it works

The foundation is a design shift in identity and access management: agents are treated as first-class identities, not as extensions of the humans they act on behalf of. Guidance on agent identity management systems describes tenets for IAM platforms that manage agents across enterprise and consumer-facing use cases, starting from fundamentally sound identity capabilities as a precursor to deploying agents at scale. In a bank, this means each digital assistant is registered with its own credentials and lifecycle, giving centralized management of the agent, its owner, and its resource permissions from a single console across customer and workforce deployments.

Delegation, not impersonation, is the second mechanism. Using OAuth 2.0 token exchange, an agent swaps a user's token for a delegation token carrying an act claim that explicitly identifies both the user and the agent. That actor claim provides a clear audit trail of whose identity and permissions the agent exercised at any moment, which is exactly what banking attribution reviews require: for every consequential action, one that moves money, changes a customer record, or commits the bank to a decision, the bank must show which human, agent, or service performed it.

Admission control formalizes the boundary between intent and execution. A formal specification for governing autonomous agents in institutional environments places a cryptographic admission check between agent intent and system state mutation: before any action reaches execution, the system validates identity, capability scope, delegation chain, and policy compliance simultaneously. The protocol layers on top of existing access models and zero-trust architectures rather than replacing them, adding cryptographic identity, capability-based authorization, deterministic risk evaluation, verifiable chained delegation, transitive revocation, and immutable auditing. Those properties map directly onto what a bank needs to prove when an agent-initiated payment or credit action is questioned later.

The banking context sharpens the stakes. Agentic systems are entering payments in ways that could move transaction initiation from explicit human instructions toward agent-mediated decision making, touching authorization, liquidity management, settlement, compliance, and operational resilience. The central challenge identified is the interaction between probabilistic, adaptive decision making and the deterministic trust infrastructure that payment systems assume. Identity is the bridge: if every agent action is cryptographically bound to a delegated human authority, the bank can extend automation without losing the attribution that payment rails depend on.

Identity fabric thinking extends this to the enterprise level. Financial services identity guidance describes managing identity and consent across agents and humans as fundamental to creating trust, positioning identity as the connective tissue between intelligent automation and autonomous intelligence. Practically, that means consistent identity issuance, consent capture, and access policy across hundreds of agents, rather than bespoke identity handling per deployment, which is how shadow agents and unattributed actions accumulate.

Supervisory alignment is direct. The EU AI Act's requirements for high-risk systems, including traceability, logging, and accountability, presuppose that the bank can attribute every action to a specific actor with defined authority; agent identity infrastructure is what makes that technically possible. SR 11-7 model governance extends to agent-initiated decisions once those decisions are attributable, FFIEC examination themes on access control and identity management apply squarely to agent credentials and least-privilege enforcement, and DORA resilience testing can exercise identity revocation, including transitive revocation of delegated authority, as part of operational resilience.

Frequently asked questions

Sources

  1. nist.gov
  2. pingidentity.com
  3. elibrary.imf.org
  4. backbase.com
  5. arxiv.org
  6. developer.pingidentity.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - ithandbook.ffiec.gov
  10. - artificialintelligenceact.eu
  11. - eur-lex.europa.eu