Asset Mgmt Agent Governance Controls | Moring
The control set for agent governance in asset management has five layers: identity and access, runtime guardrails, monitoring and observability, human oversight, and incident response. PwC's workforce governance research and Sweet's runtime framework define the layers; SR 26-2, SR 11-7 and the EU AI Act define what the evidence from each layer must show.
Short answer
The control set for agent governance in asset management has five layers: identity and access, runtime guardrails, monitoring and observability, human oversight, and incident response. PwC's workforce governance research and Sweet's runtime framework define the layers; SR 26-2, SR 11-7 and the EU AI Act define what the evidence from each layer must show.
How it works
The first control layer is identity and access. Every agent holds a unique, traceable identity with task-scoped permissions, following the pattern PwC's workforce governance research specifies: a verified identity, a defined role, task-specific permissions and auditable records, with access governance reflecting industry risk, organizational role, agent-to-agent interaction and task scope. Airlock Digital adds least privilege, restricting agents to only the systems, data and actions their assigned tasks require. For an asset management firm, this layer is what makes every other control attributable, because a control that fires on an agent cannot be evidence unless the agent's identity and scope are known.
The second layer is runtime guardrails. Sweet's governance guidance encodes the pattern: permissions, thresholds, execution limits, live monitoring and automatic escalation, checking whether each delegated action is expected, permitted and reversible before an agent changes systems or data. This is the enforcement layer that Dataiku's research identifies as the most common governance failure point, where policy documents exist and agents ship anyway because no mechanism stopped them. Guardrails evaluate behavior at action time, not at deployment alone, which is what makes them controls rather than documentation.
The third layer is monitoring and observability. Atlan's guidance defines the target: the ability to see, understand and explain what agents are doing across systems with enough detail to debug issues, enforce guardrails and prove compliance, answering the question of why an agent acted before regulators, auditors or users have to ask. Lumenova's executive guide frames the same four questions: what did the agent do, why did it do it, what data and tools did it use, and was the action aligned with policy, risk thresholds and regulation. In asset management, observability is what connects an agent's behavior to a regulated deliverable such as a client report.
The fourth layer is human oversight, and it must be proportional. PwC's principle is that human oversight should increase as agent autonomy and consequence increase, and Sweet's guidance reserves human review for high-risk actions such as spend, production changes or data deletion. Airlock Digital's human-in-the-loop controls require approval before sensitive or high-impact actions. AvePoint's State of AI 2026 research found 95.5 percent of organizations that experienced an agent-related breach took at least one mitigating action, most commonly adding human-in-the-loop controls, which confirms this layer as both the most common remedy and the most commonly needed one.
The fifth layer is incident response. AvePoint's framework research defines audit and observability as the domain enabling oversight, and its breach data, 88.4 percent of organizations experiencing at least one agent-related security breach in the past twelve months, makes the response capability non-theoretical. A complete control system detects, contains, remediates and records, with the post-incident control change documented, because that sequence is what SR 11-7 validation and the EU AI Act's documentation expectations treat as evidence of an operating control system rather than a collection of settings.
Sources:
Frequently asked questions
What does the identity and access control layer require for agents?
Every agent holds a unique, traceable identity with task-scoped permissions, following PwC's workforce governance specification of a verified identity, a defined role, task-specific permissions and auditable records, with access governance reflecting industry risk, organizational role, agent-to-agent interaction and task scope. Airlock Digital adds least privilege, restricting agents to required systems, data and actions. This layer makes every other control attributable, because a control that fires cannot be evidence unless the agent's identity and scope are known.
Why are runtime guardrails the most common governance failure point?
Dataiku's research identifies enforcement as the failure point where policy documents exist and agents ship anyway because no mechanism stopped them. Sweet's guidance supplies the working pattern: permissions, thresholds, execution limits, live monitoring and automatic escalation, checking whether each delegated action is expected, permitted and reversible before an agent changes systems or data. Guardrails must evaluate behavior at action time, not at deployment alone, which is what distinguishes a control from documentation in SR 26-2 examination.
What questions must agent observability answer in a regulated firm?
Lumenova's executive guide frames four: what did the agent do, why did it do it, what data and tools did it use, and was the action aligned with policy, risk thresholds and regulation. Atlan's guidance adds the capability standard, seeing, understanding and explaining agent behavior with enough detail to debug, enforce guardrails and prove compliance before regulators or auditors have to ask. In asset management, that observability is what connects an agent's behavior to a regulated deliverable such as a client report.
How should human oversight be proportioned across the control set?
PwC's principle is that human oversight should increase as agent autonomy and consequence increase, and Sweet's guidance reserves human review for high-risk actions such as spend, production changes or data deletion. AvePoint's State of AI 2026 research found 95.5 percent of organizations that experienced an agent-related breach took at least one mitigating action, most commonly adding human-in-the-loop controls, confirming this layer as both the most common remedy and the most commonly needed one.