Asset Mgmt Model Risk Conflict Resolution | Moring
Conflict resolution in AI model risk management gives asset management firms a defined path when investment, technology, risk, and compliance teams disagree over a model's fate. Effective structures pre-assign decision rights by tier, require documented positions, route deadlocks to a model risk committee with authority, and record outcomes, preventing silent workarounds that leave models running ungoverned.
Short answer
Conflict resolution in AI model risk management gives asset management firms a defined path when investment, technology, risk, and compliance teams disagree over a model's fate. Effective structures pre-assign decision rights by tier, require documented positions, route deadlocks to a model risk committee with authority, and record outcomes, preventing silent workarounds that leave models running ungoverned.
How it works
Disagreement is normal where models carry commercial weight. In an asset management firm, an investment team may view a new machine learning signal as alpha-bearing and resist validation delay; technology may flag the vendor hosting it as opaque; risk may demand additional testing; compliance may question whether its outputs reach client communications. Research on AI risk management as a function of AI governance argues these tensions resolve only inside a holistic structure with clear accountabilities, because each function is individually correct within its mandate and the conflict is structural, not personal.
The resolution mechanism starts with pre-assigned decision rights. The tier rubric should state, before any dispute arises, which decisions belong to whom: the model owner decides within tier tolerance, the validation function blocks deployment until standards are met, and the model risk committee decides when parties deadlock or when a model crosses tiers. United States expectations from SR 11-7 through SR 26-2 presuppose exactly this structure, requiring first-line ownership and second-line independent challenge with documented oversight, so firms can borrow the supervisory architecture rather than invent one.
Documented positions convert conflict into record. Each party states its requirement, the evidence behind it, and the control it protects, in writing attached to the model's inventory entry. This practice has two effects: it forces specificity, since a demand without evidence is easy to withdraw, and it creates the trail supervisors and internal auditors later sample. The Wharton-affiliated AIRS working group emphasized risk categorization for financial firms, and categorization does much of the work here, since disputes framed as tier questions resolve by rubric rather than by seniority.
Deadlock handling requires a committee with real authority. The model risk committee should hold the power to approve, condition, or reject a model, including conditional approvals with defined remediation and dates. Academic review of AI model governance in financial institutions finds governance weaknesses concentrated exactly where authority is ambiguous, so the committee charter matters: members must include independent second-line voices, and investment representation must be strong enough that decisions are not experienced as vetoes by a distant function. Robeco's asset management research frames the test investors apply: who remains accountable when tools act, and the committee's decision record is the answer.
Resolution must be fast enough that workarounds lose their appeal. The characteristic failure is silent adoption: a team denied approval runs the model anyway inside a research environment, and it drifts into production use without inventory or validation. Fast, conditional paths are the countermeasure, granting bounded usage with explicit expiry and monitoring obligations rather than binary approve or block. Practitioner model risk guidance for financial institutions describes conditional deployment as standard practice, and it converts would-be workarounds into governed pilots with evidence.
Finally, resolution outcomes feed the framework. Every significant dispute should end with a rubric update, a validation standard clarification, or a monitoring threshold adjustment, so the same argument does not recur with the next model. Firms that treat conflicts as calibration events accumulate institutional knowledge in the framework itself; firms that resolve ad hoc pay for the identical disagreement repeatedly, and the accumulated friction eventually shows up as ungoverned models, which is the outcome the entire structure exists to prevent.
Frequently asked questions
How should an asset management firm resolve disputes between investment teams and risk functions over AI models?
Resolve them through pre-assigned decision rights rather than negotiation by seniority. The tier rubric should state before any dispute which decisions belong to the model owner within tolerance, to the validation function before deployment, and to the model risk committee on deadlock. United States expectations from SR 11-7 through SR 26-2 already require first-line ownership with second-line independent challenge, so the supervisory structure doubles as the conflict architecture. Each party states its position in writing with evidence, disputes framed as tier questions resolve by rubric, and the committee holds authority to approve, condition, or reject.
What prevents teams from running ungoverned models after a model risk conflict?
Silent adoption is the characteristic failure, so the countermeasure is speed with boundaries. Conditional approvals grant bounded usage with explicit expiry, monitoring obligations, and defined remediation dates, which removes the incentive to work around a binary approve or block decision. Practitioner model risk guidance for financial institutions describes conditional deployment as standard. Inventory discovery tooling catches unregistered models operating anyway, and repeated workarounds become committee business with consequences. Robeco's asset management research frames why the firm must care: investors ask who remains accountable when tools act, and ungoverned models make that question unanswerable.
Where should final authority over contested AI models sit in asset management?
Final authority belongs to the model risk committee, and its charter must make that authority real. The committee should include independent second-line members, hold power to approve, condition, or reject models, and record decisions with rationale. Academic review of AI model governance in financial institutions finds weaknesses concentrated where authority is ambiguous, so vague consensus language is the risk, not strong committees. Investment representation should be substantial enough that decisions are not experienced as external vetoes, and every decision must name an accountable executive, which is exactly the accountability investors and supervisors under SR 26-2 lineage expectations look for.
How do model risk conflicts improve the framework over time?
Treat each significant dispute as a calibration event. When the outcome reveals a rubric gap, an ambiguous validation standard, or a poorly set monitoring threshold, the resolution should end with a framework update so the same argument does not recur with the next model. Documented positions attached to inventory entries build a precedent library that shortens future debates. Firms that operate this way accumulate institutional knowledge inside the framework itself, while firms resolving ad hoc pay repeatedly for identical disagreements, and the friction eventually produces ungoverned models, which is precisely the outcome the conflict resolution structure exists to prevent. Each resolution should update the model risk rubric, so asset management teams do not relitigate the same dispute with the next model.