Book a demo
Topics · GOVERNANCE-AGENTS

Asset Mgmt Agent Governance Challenges | Moring

The governing challenges for asset management agents are autonomy without accountability, permission inheritance that outruns design, drift between creator intent and production behavior, and oversight programs designed for model outputs rather than agent actions. Each challenge maps to a specific control: named ownership, scoped identities, runtime monitoring and consequence-tiered review, aligned with SR 26-2, SR 11-7 and EU AI Act expectations.

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

Short answer

The governing challenges for asset management agents are autonomy without accountability, permission inheritance that outruns design, drift between creator intent and production behavior, and oversight programs designed for model outputs rather than agent actions. Each challenge maps to a specific control: named ownership, scoped identities, runtime monitoring and consequence-tiered review, aligned with SR 26-2, SR 11-7 and EU AI Act expectations.

How it works

The first challenge is autonomy without accountability. IBM's governance analysis observes that agents make decisions on the fly and adapt to changing conditions, unlike conventional software following rule-based programming, and that the characteristics making agentic AI powerful also make it difficult to govern. PwC's governance research on the three shifts for trusted autonomy frames the answer: treat AI agents like a digital workforce, with clear ownership, authority and oversight, and build trust from continuous monitoring, risk-based controls and human accountability at scale. In asset management, where an agent's recommendation can influence a portfolio decision, accountability cannot be deferred to the model vendor.

The second challenge is permission inheritance. Sweet's governance checklist identifies the main enterprise challenges as arising from autonomy: agents can drift from creator intent, inherit excessive permissions, or make unpredictable decisions unless monitored as they act. Zenity's CISO guidance adds that agents inherit delegated identities and API access across SaaS, cloud and endpoint environments and scale faster than governance programs, creating blind spots across identity, data access and runtime behavior. For a firm holding client assets, an agent that accumulates trading-system access while designed for research summarization is not a hypothetical; it is the default trajectory absent deliberate provisioning.

The third challenge is governance designed for the wrong object. Palo Alto's governance guidance notes that most AI governance programs were designed for model outputs, not autonomous actions, with oversight often stopping at training validation or pilot controls. Credo AI's framework research makes the same point structurally: the object of concern moves from output to action, and the relevant questions become operational, which actions the agent took, which it declined, what authorized them, what data and systems were touched, and whether the effects are reversible. Point-in-time governance artifacts such as model cards cannot answer those questions; runtime records must.

The fourth challenge is consequence-tiering. Robeco's agentic AI oversight work for asset management sets the standard that should rise with the consequence of the AI-assisted action: internal summarization does not need the same controls as investment recommendations, manager selection or client-facing communication. The governance program therefore has to classify workflows by consequence and apply controls proportionally, which requires an inventory of use cases before any enforcement discussion. A uniform control regime either over-governs low-risk work, or, more dangerously, under-governs the workflows with client and regulatory impact.

The final challenge is pace and scale. McKinsey data cited across the research shows adoption expanding while governance maturity lags, with only a small fraction of organizations considering their AI adoption mature. The EU AI Act's traceability and documentation obligations for high-risk systems, and SR 26-2 and SR 11-7 model risk expectations, are fixed reference points; the firms that manage these challenges treat them as the design constraints for every agent deployed, rather than a compliance exercise conducted after deployments have already multiplied.

Sources:

Frequently asked questions

Why is autonomy without accountability the first governance challenge?

IBM's analysis observes that agents make decisions on the fly and adapt to changing conditions, unlike conventional rule-based software, and that these same characteristics make agentic AI difficult to govern. PwC's trusted-autonomy research answers with a workforce analogy: clear ownership, authority and oversight, with trust built from continuous monitoring, risk-based controls and human accountability at scale. In asset management, where agent output can influence portfolio decisions, accountability cannot be deferred to a model vendor or an engineering team.

How does permission inheritance create risk in an asset management context?

Sweet's governance checklist identifies drift from creator intent, inherited excessive permissions and unpredictable decisions as the main autonomy-driven challenges, while Zenity adds that agents inherit delegated identities and API access across SaaS, cloud and endpoint environments and scale faster than governance programs. For a firm holding client assets, an agent accumulating trading-system access while designed for research summarization is the default trajectory absent deliberate provisioning, which is why scoped identities and access reviews precede any deployment.

Why do governance programs built for model outputs fail agents?

Palo Alto's guidance notes most AI governance programs were designed for model outputs, with oversight stopping at training validation or pilot controls. Credo AI's framework research sharpens the point: the object of concern moves from output to action, so the relevant questions become which actions the agent took, which it declined, what authorized them, what was touched and whether effects are reversible. Point-in-time artifacts such as model cards cannot answer those questions; only runtime records can.

How should controls scale with consequence?

Robeco's agentic AI oversight work for asset management sets the standard that should rise with the consequence of the AI-assisted action: internal summarization does not need the same controls as investment recommendations, manager selection or client-facing communication. The program must therefore inventory use cases, classify them by consequence, and apply controls proportionally. Under SR 26-2 and the EU AI Act, that tiering is also what makes the evidence trail proportionate rather than either over-governed or under-governed.

Sources

  1. pwc.com
  2. sweet.security
  3. ibm.com
  4. airlockdigital.com
  5. paloaltonetworks.com
  6. venn.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu