Book a demo
Topics · GOVERNANCE-AGENTS

Wealth management AI approval gate SR 26-2 | Moring

Senior compliance leaders in wealth management must navigate the SR 26-2 transition by distinguishing between traditional models and generative AI. Because SR 26-2 excludes generative and agentic AI from its model definition, firms can optimize their AI approval gate by applying specialized governance to these tools rather than treating them under legacy model risk management protocols.

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

Short answer

Senior compliance leaders in wealth management must navigate the SR 26-2 transition by distinguishing between traditional models and generative AI. Because SR 26-2 excludes generative and agentic AI from its model definition, firms can optimize their AI approval gate by applying specialized governance to these tools rather than treating them under legacy model risk management protocols.

How it works

The implementation of SR 26-2 represents a critical inflection point for wealth management firms managing complex technology stacks. As these organizations integrate artificial intelligence into investment research, client service, and risk analysis, the ability to distinguish between traditional models and new AI technologies determines the efficiency of the AI approval gate. Failure to correctly categorize these tools can lead to unnecessary administrative burdens on the second line of defense or, conversely, insufficient oversight of high-consequence generative AI applications.

At the wealth management AI approval gate, the control mechanism functions through a refined classification process. Under SR 26-2, the definition of a model has been narrowed to exclude generative and agentic artificial intelligence, as well as simple arithmetic or deterministic rules-based calculations [3]. This allows wealth management compliance teams to perform aggressive inventory de-scoping, separating legacy models from advanced AI. This enables a risk-based approach where materiality is determined by both Model Exposure, which is the quantitative impact, and Model Purpose, which is the qualitative significance [5].

Regulatory guidance and research indicate a significant shift in the governance landscape. SR 26-2, issued on April 17, 2026, supersedes SR 11-7 and emphasizes a tailored approach that recognizes model risk management practices vary based on specific risk profiles [1]. While SR 26-2 excludes generative AI from the formal model definition, the OCC has noted that generative and agentic AI remain outside that specific scope [6]. Furthermore, the NIST AI Risk Management Framework provides a voluntary baseline to improve trustworthiness in AI products through functions like govern, map, measure, and manage [2, 8].

To maintain a compliant AI approval gate, wealth management practitioners should follow this checklist: first, conduct a full inventory to identify all deterministic tools, traditional models, and generative AI systems; second, apply the SR 26-2 tiering to designate immaterial models for expedited oversight; third, evaluate generative AI for output fidelity, ensuring results remain faithful to approved sources and business rules; and fourth, establish accountable decision rights and independent challenge mechanisms for high-consequence AI use cases [3, 5, 6, 7].

Failure modes at the AI approval gate often stem from inaccurate classification or data issues. Wealth management firms may encounter hallucinations, where generative AI produces inaccurate or misleading information presented as fact [4]. Other risks include bias, where model outputs are skewed due to limited or outdated training data, and concept drift [4]. Additionally, firms risk losing adverse-action fidelity if generative AI transforms technical evidence into customer-facing language in a way that is not traceable to the principal driving factors of a decision [7].

Teams should begin by re-evaluating their existing model inventory to align with the new SR 26-2 definitions [5]. This involves identifying which deterministic tools can be de-scoped from the legacy SR 11-7 framework to reduce the administrative load on compliance officers. Simultaneously, leadership must develop enterprise-level supervisory processes specifically for generative AI to address unique risks like summarization and information extraction errors [4].

Frequently asked questions

How does SR 26-2 change the way wealth management firms handle generative AI?

Under SR 26-2, generative and agentic AI are explicitly excluded from the formal model definition that was previously governed under SR 11-7. This means the AI approval gate for wealth management must transition from traditional model risk management to a specialized governance structure that addresses unique AI risks like hallucinations and bias.

What is the significance of the 'immaterial' designation in the new guidance?

The designation allows wealth management firms to use a risk-based tiering system to prioritize resources. Immaterial models only require identification and performance monitoring, which streamlines the AI approval gate by allowing compliance teams to focus rigorous oversight on complex, high-exposure models that impact client outcomes.

What specific risks must be addressed during the wealth management AI approval gate?

Wealth management leaders must mitigate risks such as hallucinations, where generative AI provides misleading information, and bias resulting from skewed training data. Additionally, ensuring adverse-action fidelity is critical when AI transforms technical data into customer-facing language to ensure decisions remain traceable and accurate.

How should a wealth management firm define model materiality under SR 26-2?

Materiality is no longer implied but is explicitly determined through a structure of Inherent Risk, Exposure, Purpose, and Materiality. For the AI approval gate, this means evaluating both the quantitative impact of the tool and its qualitative significance to the wealth management business functions.

Sources

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