Book a demo
Topics · GOVERNANCE-AGENTS

Asset Management Agent Governance Why Is It Needed | Moring

AI agent governance is needed because agents no longer just answer questions: they retrieve data, call tools, and act on business systems with delegated authority. For asset managers, ungoverned agents touch research, valuations, operations, and client communications, where errors carry fiduciary and regulatory consequences. Policy alone does not constrain behavior; governance turns enterprise rules into runtime controls with named accountability.

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

Short answer

AI agent governance is needed because agents no longer just answer questions: they retrieve data, call tools, and act on business systems with delegated authority. For asset managers, ungoverned agents touch research, valuations, operations, and client communications, where errors carry fiduciary and regulatory consequences. Policy alone does not constrain behavior; governance turns enterprise rules into runtime controls with named accountability.

How it works

The case for AI agent governance begins with a structural shift. Earlier AI systems generated insights, and human operators decided what to do next. Today's agents retrieve data, call APIs, and trigger real workflow actions on their own, which means the challenge has moved from trusting what an AI says to deciding what it should be allowed to do. For an asset manager, that shift lands on sensitive ground: agents increasingly participate in research workflows, trade support, valuation processes, and client communications, all areas where a firm's fiduciary obligations and supervisory expectations apply to machine actions exactly as they apply to human ones.

The scale of adoption explains the urgency. Enterprise research on the agentic opportunity describes multi-trillion-dollar potential across use cases, while a separate survey of organizational maturity found only one percent of organizations consider their AI adoption mature. Adoption is racing ahead of control, and governance research describes the resulting gap concretely: organizations create policy documents, discuss guardrails, and draft checklists, yet agents still ship without any of it because there was no mechanism to stop them. An asset manager with enthusiastic teams and no runtime enforcement discovers agent sprawl after the fact, when actions have already been taken on systems that matter.

Why is ordinary software governance insufficient? Because agents operate with delegated authority and can affect multiple systems at once, a risk profile that differs from traditional applications. An agent with excessive permissions could expose sensitive client information, alter operational settings, or act across systems in combinations no single human role would combine. Governance research on delegated authority argues the discipline must extend beyond model alignment, compliance, or monitoring to establish explicit oversight and accountability for agent behavior at runtime, with clear boundaries on what agents can access and perform.

The workforce framing makes the need concrete for asset management. Workforce governance research argues agents are becoming digital coworkers that need enough access to be useful but not so much autonomy that they create unmanaged risk, and that each should carry a verified identity, a defined role, task-specific permissions, and auditable records. In an asset management firm, the analogue is exact: a junior analyst operates within supervision, an entitlement system, and a mandate; an agent performing analyst-like work should carry the same structure. Without it, accountability for machine actions dissolves, and accountability is precisely what fiduciary supervision cannot tolerate.

Regulatory posture is the third driver. Asset managers already operate model risk disciplines shaped by SR 11-7 and continued under SR 26-2, which require documented validation, independent review, and clear accountability for systems influencing decisions. Agents that act on valuations, client reporting, or operations fit that risk character, and supervisors will reasonably expect the same evidentiary standard. Where European operations or vendors bring the EU AI Act into scope, deployer duties add explicit transparency and human oversight requirements. Governance is needed, in short, because the firm's existing obligations already reach the behavior agents exhibit; the only question is whether evidence exists.

There is also a practical, non-regulatory reason: ungoverned agents are simply more expensive and less reliable to operate. Runtime enforcement research notes that strong controls cover data access, prompts, model choices, tool calls, identities, and APIs, and that encoding them as permissions, thresholds, execution limits, and monitoring lets teams adopt agents quickly within boundaries rather than oscillating between prohibition and free-for-all. Firms that skip governance tend to rediscover it piecemeal after incidents, paying more for the same controls while operating blind in the interim.

The honest summary is that governance is needed not because agents are new but because they act. Cloud adoption guidance recommends a centralized, enforceable baseline for every agent, aligned with existing identity, data governance, and security practices, precisely because baseline policies applied per team do not hold as adoption grows. An asset manager that installs that baseline, registers its agents, and enforces boundaries at runtime gets the benefits of agentic automation on investment and operations work while keeping the control posture its clients, its board, and its supervisors already recognize.

Frequently asked questions

Why can an asset manager not rely on its existing software governance for AI agents?

Because agents operate with delegated authority and can affect multiple systems at once, a risk profile governance research identifies as distinct from traditional applications. An agent can combine data access, tool execution, and external communication in ways no single human role would, and it acts without continuous human oversight. Existing software governance reviews applications statically; agents need runtime boundaries, verified identities, and auditable records of each delegated action, particularly where they touch investment, operational, or client-facing systems.

Is agent governance necessary if the firm's agents only do internal research tasks?

Internal tasks still warrant governance, for three reasons. First, inventories routinely reveal that agents have spread beyond their intended scope, since employees deploy them and platforms auto-provision them. Second, research agents consume proprietary information whose handling rules matter as much as the task itself. Third, governance research emphasizes that controls should be encoded as permissions, thresholds, and execution limits so adoption can proceed safely; light-touch controls on internal agents are part of a gradient, not an exemption, and they establish the structure before higher-consequence agents arrive.

What business risk does an asset manager take by delaying agent governance?

The primary risks are accountability gaps and incident discovery after the fact. Survey research cited in governance analysis found only one percent of organizations consider their AI adoption mature, and traceability surveys show very few firms can fully trace AI output. An asset manager delaying governance operates agents that touch valuations, client reporting, or operations without verifiable authorization trails, meaning an error, a data leak, or an unauthorized action surfaces without evidence of what was approved, by whom, and under which policy, exactly when documentation matters most.

How does the need for governance connect to the firm's model risk obligations?

Directly, because the risk character is comparable. Model risk management under SR 11-7 and its successor SR 26-2 requires documented validation, independent review, and clear accountability for systems that influence decisions. Agents acting on valuations, client reporting, or operational workflows influence decisions in the same sense, so supervisors can reasonably expect equivalent evidence. Governance supplies that evidence through registration, action logging, and runtime enforcement; without it, the firm faces its model risk obligations with agents it cannot document, which is a supervisory exposure rather than a technical one.

Sources

  1. pwc.com
  2. sweet.security
  3. dataiku.com
  4. akamai.com
  5. learn.microsoft.com
  6. paloaltonetworks.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu