Book a demo
Topics · REGULATORY

Asset Management Model Risk Stop Control | Moring

Stop controls are the family of mechanisms an asset manager uses to halt model or agent activity: kill switches for immediate halts, rate limits for bounding, permission revocations for narrowing, and rollback for restoration. Research on governance enforcement stresses that agents keep shipping without guardrails when no mechanism can stop them; stop controls are that mechanism.

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

Short answer

Stop controls are the family of mechanisms an asset manager uses to halt model or agent activity: kill switches for immediate halts, rate limits for bounding, permission revocations for narrowing, and rollback for restoration. Research on governance enforcement stresses that agents keep shipping without guardrails when no mechanism can stop them; stop controls are that mechanism.

How it works

Stop control is the umbrella term for every mechanism that halts or bounds an AI system in an asset manager's estate: the kill switch that freezes an agent, the rate limit that caps a runaway loop, the permission revocation that narrows an over-scoped integration, the rollback that restores a validated version. They belong to one family because they share a purpose: making containment faster than escalation.

The need is well documented. Governance research observes that many programs fail at enforcement: policy documents are written, guardrails are discussed, checklists are drafted, and agents still ship without any of it because there was no mechanism to stop it. Separately, agent-focused risk research notes that unmonitored models quietly decay while chained agents compound risk by acting without human review, so the interval for intervention is short.

Stop controls should be tiered like everything else in model risk. AI-MRM guidance extends SR 11-7 principles through drift detection, model tiering, and attestation, and an asset manager can map stop authority to tiers: a research summarization agent may be halted by its product owner, while a model feeding valuation or execution requires dual authorization and immediate notification of risk and compliance functions.

Each stop control should produce evidence automatically. Governance research emphasizes self-generating evidence and living inventories over policy prose, so every halt, revocation, or limit breach is logged with the actor, reason, affected model version, and restoration path. This log doubles as the record supervisors and boards will ask for, and it feeds the quarterly review cycle with concrete containment data.

The regulatory frame makes stop controls more than prudence. SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 in the first model risk overhaul in fifteen years, and commentary notes it regulates models rather than agents, leaving stop paths for agents to firm design. EU AI Act human oversight expectations for firms with European operations point in the same direction: intervention must be possible and demonstrable.

Design guidance from platform research rounds out the picture: centralized layers that track models, score risk, and enforce policies give stop controls a single place to live. An asset manager that consolidates kill switches, limits, revocations, and rollback into one governed control plane can stop any model or agent within minutes, with a complete record, which is the standard the rest of the framework quietly assumes.

A consolidated control plane also simplifies the human side. Stop authority, escalation contacts, and drill schedules live in one place, so a new model owner inherits the same intervention procedures as every other owner. Platform research describing centralized layers that track models, score risk, and enforce policies supports this consolidation, and it prevents the common failure where stop capability exists in one team's tooling but not in the workflows that need it most.

Frequently asked questions

What counts as a stop control in model risk?

Kill switches for immediate halts, rate limits for bounding runaway or anomalous activity, permission revocations for narrowing access, and rollback for restoring a validated version. They share one purpose: making containment faster than escalation when a model or agent misbehaves in an asset management workflow.

How should stop authority be assigned?

Map stop authority to model tiers: a research summarization agent may be halted by its product owner, while a model feeding valuation or execution requires dual authorization with immediate notification to risk and compliance. AI-MRM tiering research supports scaling intervention rights to decision impact.

Why must stop controls generate evidence automatically?

Because model risk governance research emphasizes self-generating evidence over policy documents, and containment records are what boards, supervisors, and post-incident reviews request. Every halt, revocation, or limit breach in an asset management workflow should log the actor, reason, model version, and restoration path automatically.

Do regulators expect stop controls for agents?

Effectively yes. SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 but regulates models rather than agents, so stop paths are firm-designed, and EU AI Act human oversight expectations apply where firms operate in Europe. Demonstrable intervention capability is the converging expectation.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu