Book a demo
Topics · GOVERNANCE-AGENTS

Asset Management AI approval gate Challenges | Moring

Asset managers deploying approval gates face challenges on several fronts: agent behavior that outruns governance frameworks, regulatory uncertainty, shadow AI use, approver bottlenecks, and evidence gaps. Each challenge has a practical response, from runtime enforcement to risk-tiered routing, and none justifies delaying the control itself.

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

Short answer

Asset managers deploying approval gates face challenges on several fronts: agent behavior that outruns governance frameworks, regulatory uncertainty, shadow AI use, approver bottlenecks, and evidence gaps. Each challenge has a practical response, from runtime enforcement to risk-tiered routing, and none justifies delaying the control itself.

How it works

The first challenge is pace. Arcesium's framework observes that AI agents are advancing faster than many investment firms' governance frameworks, which means controls designed for yesterday's assistants must govern today's systems that call tools, use identities, retrieve data, and make decisions reaching beyond the original prompt. A gate designed around static prompts will miss the actions that matter.

The second challenge is regulatory uncertainty. Baker Tilly finds that uncertainty continues to slow AI adoption in asset management and that firms must map existing obligations onto new tools before those tools deepen into operations. Teams sometimes delay gates while waiting for clearer guidance, which inverts the risk: uncontrolled agent actions create the very record gaps that future scrutiny would penalize.

The third challenge is shadow AI. Governance research across the industry, including Domo's and Bitsight's material, stresses visibility into unapproved tool use as a core control problem. In an asset management firm, analysts experiment with agents on personal accounts and unsanctioned APIs, and a gate cannot authorize what it never sees, so discovery and nonpunitive reporting channels must accompany enforcement.

The fourth challenge is approver load. A poorly tiered gate sends every action to a committee, and the queue becomes the bottleneck that tempts staff to route around the control. Solytics recommends classifying systems by operational risk before setting approval gates and testing requirements, and assigning decision owners to every gate so unresolved risks cannot pass silently; both recommendations exist to keep the queue small and the decisions real.

The fifth challenge is evidence quality. Robeco's research asks who remains accountable when agentic systems act, and an approval log without context, limits, or rationale answers nothing during diligence. Boards and investors increasingly ask for traceability, and Bitsight's framing of governance emphasizes accountability structures alongside policies, so the gate must capture who approved what under which constraints, not merely a binary allow or deny.

None of these challenges argues for waiting. Each has a known response: runtime enforcement rather than prompt-time checks, risk-based tiering to protect approver attention, discovery to surface shadow use, and evidence capture designed for external readers. Firms that treat these as design inputs deploy gates that hold up in operation rather than demonstrations.

Frequently asked questions

What is the hardest operational challenge with approval gates?

Approver load is usually the hardest. If the gate routes every agent action to a human queue, approvals become a bottleneck and staff start looking for workarounds, which undermines the control. The response is risk-based tiering: classify workflows by operational consequence before configuring gates, as solytics-partners recommends, so only high-consequence actions such as trades, data purchases, or client-facing outputs require a named approver. Assigning a decision owner to each gate also prevents unresolved risks from passing silently while keeping routine work flowing.

How does regulatory uncertainty affect gate deployment in asset management?

Baker Tilly reports that regulatory uncertainty slows AI adoption and that firms need to understand how existing obligations apply before new tools move deeper into operations. Waiting for perfect guidance is itself a risk, because ungoverned agent actions create evidence gaps that later scrutiny would penalize. The practical approach is to anchor gates in obligations that already exist, extending model risk management principles in the spirit of SR 11-7 to AI systems, and to design gates so permissions and evidence requirements can be tuned as new supervisory expectations for AI arrive.

Why is shadow AI a problem for gate effectiveness?

A gate can only authorize actions that pass through it, and shadow AI use means agents and tools operating outside that path entirely. Industry governance material, including Domo and Bitsight, treats visibility into unapproved AI use as a foundational control problem, and asset management is exposed because analysts readily experiment with unsanctioned tools on research tasks. Discovery scans, declared exception channels, and nonpunitive reporting bring that activity into view so it can be gated, rather than driving it further underground where neither compliance nor security can see it.

What makes approval evidence inadequate during diligence?

An approval log that records only allow or deny answers little. Diligence and supervisory review ask who authorized a consequential action, under which limits, with what rationale, and whether exceptions followed a documented path, reflecting Robeco's question of who remains accountable when systems act. Evidence should capture the approver of record, the policy version and limits in force, the context presented, and the outcome of any escalation. Bitsight frames governance as accountability structures as well as policies, and gates that capture this context turn routine operation into usable audit material.

Sources

  1. robeco.com
  2. bitsight.com
  3. servicenow.com
  4. databricks.com
  5. bvp.com
  6. domo.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu