Asset Mgmt Agent Governance Guide | Moring
A guide to governing AI agents in asset management starts from one premise: governance is enforcement, not paperwork. Dataiku notes many programs fail because agents ship without any mechanism to stop them, and its guide pairs five program components with a ten-point production gate. This page adapts that structure, plus Microsoft's baseline recommendation and Sweet's runtime tests, for firms supervising investment-relevant agents.
Short answer
A guide to governing AI agents in asset management starts from one premise: governance is enforcement, not paperwork. Dataiku notes many programs fail because agents ship without any mechanism to stop them, and its guide pairs five program components with a ten-point production gate. This page adapts that structure, plus Microsoft's baseline recommendation and Sweet's runtime tests, for firms supervising investment-relevant agents.
How it works
A useful guide begins by naming the failure it prevents. Dataiku observes that many governance programs fail at enforcement: policy documents exist, guardrails are discussed, checklists are drafted, yet agents ship without any of it because there was no mechanism to stop them. Its guide therefore defines a production-grade program as five components: policy and authority boundaries, access and identity controls, development and evaluation practices, runtime guardrails, and monitoring with incident response. For an asset manager, those components map naturally onto familiar structures, investment policy committees, information security, model validation, operations, and compliance, which is why agent governance should be framed as an extension of existing supervisory machinery rather than a parallel bureaucracy.
The second element is the gate. Dataiku's ten-point deployment checklist functions as the boundary between an agent and production, and Microsoft's Cloud Adoption Framework supplies the floor beneath it: establish a centralized, enforceable governance and security baseline for all agents, aligned with existing identity, data governance, and security practices, with minimum requirements every agent must meet before operating. Together they convert admission from a conversation into a criteria check. In asset management terms, an agent that computes or transmits investment-relevant outputs gets admitted only with a verified identity, scoped permissions, a named owner, and a defined evidence plan.
The third element is runtime. Sweet's guidance states the operating test plainly: governance turns enterprise rules into runtime controls, checking whether each delegated action is expected, permitted, and reversible before an autonomous agent changes systems or data, with coverage extending to data access, prompts, model choices, tool calls, identities, and APIs because failures involve injected instructions, leakage, and over-permissioned credentials. Palo Alto Networks describes the same territory as the structured management of delegated authority, setting clear boundaries on what agents can access and perform at runtime, going beyond model alignment or monitoring alone. These runtime checks are what make the guide enforceable after deployment rather than only before.
The fourth element is inventory and its maintenance. Zenity's CISO checklist places discovery and inventory first, because agents embedded across SaaS, cloud, and endpoint environments inherit delegated identities and API access and scale faster than governance programs, creating blind spots across identity and access. Akamai's white paper adds that organizations must choose among policy-based, risk-based, trust-based adaptive, and hybrid framework models depending on regulatory context and risk tolerance. For an asset manager, the pragmatic recommendation is a risk-tiered model: discovery feeds a consequence classification aligned with SR 26-2 scoping, and oversight intensity follows the tier rather than applying uniform friction to every agent.
The guide closes where supervision begins. The EU AI Act's documentation and traceability duties for high-risk systems mean the evidence the program generates serves regulators as well as internal control, and model risk expectations in asset management similarly reward demonstrable oversight. The recurring admonition across sources is proportionality: agents need enough access and autonomy to be useful, framed inside boundaries that make every consequential action attributable, approvable, and reversible. A firm that implements the five components, holds the deployment gate, enforces runtime tests, and maintains inventory will find that examinations become reviews of standing evidence rather than reconstructions from memory.
Frequently asked questions
What are the five components of a production-grade agent governance program?
Dataiku's guide defines them as policy and authority boundaries, access and identity controls, development and evaluation practices, runtime guardrails, and monitoring with incident response. Their value is that they map onto structures asset management firms already operate, from investment policy committees to information security and model validation, so agent governance extends existing supervisory machinery rather than duplicating it. Paired with the ten-point deployment checklist that gates production, the components close the enforcement gap Dataiku identifies, where agents ship despite drafted policies because no mechanism stopped them.
How should an asset management firm structure its agent inventory?
Discovery comes first, following Zenity's CISO checklist, because agents across SaaS, cloud, and endpoint environments inherit identities and API access and scale faster than governance programs, producing blind spots across identity and access. Each inventoried agent should then receive a consequence classification, aligned with SR 26-2 scoping where actions touch investment decisions, and Akamai's framework menu of policy-based, risk-based, trust-based, and hybrid models suggests a risk-tiered approach. Oversight intensity follows the tier, so consequential agents receive deep review while low-risk agents run with monitoring, avoiding uniform friction that teams route around. That structure gives an asset management firm a defensible supervision answer.
What runtime tests make agent governance enforceable rather than aspirational?
Sweet's guidance states them directly: each delegated action is checked for whether it was expected, permitted, and reversible before the agent changes systems or data, with coverage across data access, prompts, model choices, tool calls, identities, and APIs, because failures involve injected instructions, leakage, and over-permissioned credentials. Palo Alto Networks frames the same discipline as structured management of delegated authority with clear runtime boundaries. For asset management, these tests catch an agent reaching beyond its mandate at the moment of the call, when correction is cheap, rather than in a monthly report assembled after harm.
How does this governance guide connect to regulatory obligations?
The EU AI Act imposes documentation and traceability duties on high-risk systems, which the program's evidence satisfies as a by-product of normal operation rather than through separate compliance assembly. Model risk expectations in asset management similarly reward demonstrable oversight of systems affecting investment outcomes. Dataiku's observation that enforcement, not policy drafting, is where programs fail is the operational link: gates, runtime checks, and inventory generate the records supervisors want to see. A firm following this guide enters examinations with standing evidence, while unstructured peers reconstruct history from memory under deadline pressure.