Book a demo
Topics · REGULATORY

Asset Management Model Risk Vendor Selection | Moring

Vendor selection in asset management model risk means evaluating AI providers on governability, not just performance: model transparency and documentation, drift disclosure and version stability, data handling, audit evidence, and contractual support for oversight. Research warns that externally hosted models can change behind stable APIs, and governance guidance treats vendor models as inventory items with owners.

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

Short answer

Vendor selection in asset management model risk means evaluating AI providers on governability, not just performance: model transparency and documentation, drift disclosure and version stability, data handling, audit evidence, and contractual support for oversight. Research warns that externally hosted models can change behind stable APIs, and governance guidance treats vendor models as inventory items with owners.

How it works

Most asset managers will not train frontier models; they will buy them. That makes vendor selection a model risk decision of the first order. AI-MRM research frames the discipline as identifying, measuring, monitoring, and controlling risk across the model lifecycle, and a vendor model enters that lifecycle at procurement, where the cheapest controls are applied: contractual, architectural, and evidentiary.

What should the evaluation cover? Model governance research ties ownership, validation, monitoring, change management, and retirement into a lifecycle, and vendor models must fit each link. Transparency and documentation enable validation; version disclosure and change notice terms enable monitoring; data handling terms determine what client and portfolio information leaves the firm; and audit evidence provisions determine whether attestation is possible at all.

The hidden risk is silent change. Lifecycle research warns that the model which performed brilliantly in testing can degrade silently in production, and vendor-hosted models can alter behavior behind a stable API. Evaluation should therefore ask how a vendor communicates model updates, whether versions can be pinned, and what output-distribution evidence the vendor will provide between reviews.

Governability extends to the plumbing. Risk commentary describes the Model Context Protocol as a common denominator for AI data connections and model governance, and gateway research shows vendors' agents and tools can be routed through controlled layers with authentication, logging, and policy enforcement. A vendor whose systems support gateway-mediated access with least-privilege scopes is materially easier to govern than one demanding direct integration.

Regulatory expectations apply to purchased models as much as built ones. 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, so vendor agents require firm-side controls. EU AI Act duties add provider and deployer role distinctions that asset managers with European business must map in contracts.

A practical selection standard: treat every vendor model and agent as an inventory item with a named internal owner, a tier, defined access scopes, and monitoring obligations, and score vendors on their ability to support those four. Asset management governance research recommends defining agent authority and extending existing compliance principles to AI, and procurement is where that extension either begins or quietly fails.

Reference checks and procurement evidence close the loop. Peer adoption, incident transparency, and responsiveness on model documentation are observable signals, and the contract should encode the governability findings: change notice periods, version pinning rights, evidence delivery, and exit provisions that preserve the firm's ability to replace the model without breaking workflows. Vendor selection done this way extends the model risk framework outward instead of leaving it at the firm's boundary.

Frequently asked questions

What should asset managers evaluate beyond model performance when selecting AI vendors?

Governability: documentation enabling validation, version disclosure and change notice terms, data handling limits, audit evidence provisions, and support for gateway-mediated access with least-privilege scopes. Vendor models must fit the ownership, validation, monitoring, change management, and retirement chain that asset management model governance research describes.

Why is silent vendor model change a specific risk?

Because externally hosted models can alter behavior behind a stable API, and lifecycle research warns that production models can degrade silently until damage surfaces. Asset management vendors should be evaluated on update communication, version pinning, and interim output-distribution evidence so model risk stays detectable between reviews.

How do SR 26-2 and EU AI Act duties apply to vendor models?

SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 and applies to purchased models as much as built ones, though commentary notes it regulates models rather than agents. EU AI Act provider and deployer role distinctions must be mapped in contracts where firms serve European clients.

Should vendor agents be treated differently from vendor models?

Yes, because agents act across systems rather than produce outputs alone. Treat them as inventory items with named internal owners, tiers, access scopes, and logging obligations, and score vendors on supporting those controls, consistent with asset management governance research on defining agent authority.

Sources

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