Asset Management AI Agent Identity Audit | Moring
An identity audit for AI agents verifies that every agent at an asset manager has a registered identity, a named owner, scoped permissions, and a record of what it actually did. Because agents authenticate programmatically, act at machine speed, and chain delegation across sub-agents and APIs, the audit must reconcile authorized access with observed behavior rather than simply sampling account lists.
Short answer
An identity audit for AI agents verifies that every agent at an asset manager has a registered identity, a named owner, scoped permissions, and a record of what it actually did. Because agents authenticate programmatically, act at machine speed, and chain delegation across sub-agents and APIs, the audit must reconcile authorized access with observed behavior rather than simply sampling account lists.
How it works
Auditing agent identities differs from auditing human access in ways asset managers need to plan for. Identity management research describes the underlying gap: agents use ephemeral credentials, act at machine speed without sessions, and create delegation chains across sub-agents and APIs that legacy identity models were never designed to track. A quarterly review of account lists cannot see any of that. The audit question is not whether accounts exist but whether each agent's recorded authorization matches its observed actions, and whether a named human stands behind each identity.
Start the audit with inventory reconciliation. Non-human identities, including APIs, bots, agents, and service accounts, already outnumber human users by roughly three to one across enterprises, according to identity management research, and they are frequently targeted in supply chain attacks. For an asset manager, that ratio is a warning about coverage: the agents governance believes are registered are almost certainly fewer than the credentials actually operating. The first audit test is therefore a discovery sweep, comparing the registered agent inventory against observed credentials, OAuth grants, API keys, and protocol connections, and documenting every gap with an owner and a remediation date.
The second test examines ownership and purpose. Checklist research on runtime agent governance ties each agent action to an owner, an approved scope, enforceable limits, and evidence auditors can verify, and warns that agents drift from their creator's intent. The audit should sample agent identities and verify three fields: a business owner in the function the agent serves, a technical owner accountable for credentials, and a recorded permission rationale tied to a business purpose. Agent identities missing any of these fields are findings, because without them, no reviewer can judge whether current access is appropriate.
The third, and most distinctive, test is intent-to-execution reconciliation. Identity governance research describes behavioral verification as the practice of reconciling what an agent is authorized to do with what it actually does, extending identity administration beyond provisioning and certification. In audit terms, this is substantive testing: pull a sample of agent actions, retrieve the permission mandate in force at the time, and check each action against it, including which systems were touched, which data was read or moved, and which approvals applied. An agent that stayed inside its grants but outside its intent is exactly the finding this test exists to surface.
Delegation chains deserve their own procedure. Agents spawn child agents and act through APIs, so an action may trace back through several identities before reaching a human decision. Auditing each link independently can miss the composite risk; the audit should reconstruct end-to-end chains for consequential actions and confirm that authority was properly delegated at each hop, with the originating authorization recorded. Identity research notes that these chains are precisely what legacy models cannot track, making manual reconstruction the current norm until governance tooling automates it.
The evidentiary output should map onto obligations the firm already carries. Model risk management under SR 11-7 and its successor SR 26-2 requires documented accountability for systems influencing decisions, and an agent identity audit produces exactly that: who owned what, what was authorized, what was observed, and what changed. Where European operations or vendors bring the EU AI Act into scope, the same records demonstrate deployer oversight. Asset managers should therefore persist identity audit artifacts into existing examination-ready repositories rather than leaving them in identity tooling exports.
Cadence and scope should reflect consequence. High-consequence agents, those touching valuations, order flow, or client communications, warrant continuous monitoring with periodic substantive sampling; low-consequence internal agents can sit on annual certification with discovery sweeps in between. The audit function also benefits from coordination: findings about drift, orphaned credentials, or unregistered agents should route to named owners with deadlines, mirroring how the firm handles model validation findings. An asset manager that audits agent identities this way converts a novel risk area into a familiar, examinable discipline.
Frequently asked questions
What fields does an asset management identity audit check for each AI agent?
Three core fields per sampled identity: a business owner in the function the agent serves, a technical owner accountable for its credentials, and a recorded permission rationale tied to a business purpose. Checklist research ties each agent action to an owner, an approved scope, enforceable limits, and verifiable evidence, and warns that agents drift from their creator's intent. Missing ownership or rationale fields are audit findings, because reviewers cannot judge whether access remains appropriate without them.
Why must an identity audit reconcile authorized access with observed behavior?
Because agents take variable actions under stable grants. Identity governance research calls this behavioral verification: reconciling what an agent is authorized to do with what it actually does, extending administration beyond provisioning and certification. For an asset manager, an agent scoped to read performance data that begins exporting it externally satisfies every static entitlement check while violating its purpose. Substantive sampling of actions against the permission mandate in force is the only test that surfaces such drift, and it mirrors how auditors already test control operation.
How should delegation chains across sub-agents be audited?
End to end rather than link by link. Agents act through child agents and APIs, so a consequential action may pass through several identities before reaching a human decision, and identity research notes these chains are what legacy identity models cannot track. For each sampled consequential action, reconstruct the full chain, confirm authority was delegated properly at every hop, and verify the originating human authorization is recorded. Composite risk can hide between links that each look individually acceptable.
What should trigger an out-of-cycle agent identity audit at an asset manager?
Events that change the access landscape: a new agent deployment touching portfolios, valuations, or client communications; a vendor or data provider change; a detected drift incident; or termination of a project whose credentials may linger. Identity research warns agents persist after the business context that justified their privileges has changed, so project wind-downs and mandate closures are natural triggers. A discovery sweep comparing registered agents against observed credentials should also follow any merger, reorganization, or platform migration.