The Impact of GenAI on Model Risk Management

Generative AI does not need a separate governance regime. It needs the existing model risk discipline, extended to cover how these systems actually behave.

GenAI and the Model Risk Perimeter

Which parts of the MRM framework still hold for generative AI, which parts need to change, and how regulators are already drawing the new boundary.

Where the Traditional Questions Run Out

Generative AI (GenAI) now sits inside customer service, document intelligence, fraud operations, credit analysis and decision support at many financial institutions. The technology inside the model has changed, and so has the nature of the model risk it creates.

Traditional model risk management (MRM) asks whether a model is conceptually sound, whether its data is appropriate, whether performance holds up and whether it has been independently validated. Those questions still apply to GenAI. They do not reach hallucination, prompt sensitivity, output variability, autonomy or dependence on a third-party foundation model, none of which has a natural place in a framework built for statistical models with stable inputs and outputs.

The practical answer is to govern GenAI through the same MRM disciplines of inventory, risk tiering, validation, monitoring and control, extended to cover how these systems behave. Institutions are already moving: in a September 2025 Solytics Partners webinar poll, 80.3% of respondents said they were building frameworks specifically for generative models, and 74.7% already validated or oversaw AI/ML models within their MRM programmes.

Why GenAI Doesn't Fit the Old Questions

A conventional model produces a score from relatively stable inputs. A GenAI system can return a different answer to a nearly identical prompt, produce content that reads as credible but is wrong, and depend on a foundation model that changes on the vendor's release schedule rather than the institution's. Five risk areas follow from that:

Hallucination: plausible but incorrect output, such as a wrong summary, a fabricated figure or a misstated policy. It becomes conduct and operational risk the moment it reaches a customer or a decision. NIST's Generative AI Profile lists it, under the name confabulation, among the 12 risks it identifies as unique to or exacerbated by generative AI.

Prompt sensitivity and output variability: small changes in wording can shift outputs materially, and the same prompt can return different results across runs. Regression testing designed for deterministic models becomes harder to rely on, and the object under governance extends to the prompt, the model version, the retrieval layer and the guardrails around them.

Explainability and bias: foundation models are harder to interpret than the regressions and scorecards MRM teams usually review, and they can reproduce bias present in training data or retrieval sources. Fairness and explainability therefore belong in core validation.

Provider dependency: when the foundation model belongs to a vendor, an update can change behaviour with no change on the institution's side, and the documentation available for independent review is often limited.

Security and data leakage: customer-facing GenAI adds exposure to prompt injection and adversarial inputs, and to sensitive data surfacing in outputs.

Why It Matters: Regulators Are Extending MRM to Cover GenAI

The Reserve Bank of India's draft Guidance on Regulatory Principles for Model Risk Management, released for consultation on 24 June 2026, is the most detailed recent example. It defines a model broadly enough to cover AI/ML systems, scoring algorithms, decision rules and material spreadsheets, whether or not the institution recognises them as models. It would require independent validation of third-party models regardless of any validation, certification or assurance the vendor supplies. For AI models, tiering would additionally weigh how far decisions rely on model outputs and how much autonomy the model is given. The draft also calls for controls against hallucination, red-teaming of customer-facing and generative models, and human-in-the-loop and kill-switch arrangements.

Other supervisors are moving the same way, with different scopes and timetables:

United Kingdom: the PRA's SS1/23, in force since 17 May 2024 for banks with internal model approval for regulatory capital, applies its principles to all model types, including models built with AI and machine learning.

Canada: OSFI's Guideline E-23, effective 1 May 2027, applies enterprise-wide to AI/ML and third-party models, and names a model's level of autonomy among the qualitative factors in its risk rating.

Singapore: MAS consulted from November 2025 to January 2026 on Guidelines on AI Risk Management covering generative AI and AI agents, with risk materiality assessed on impact, complexity and reliance. Final guidelines had not been issued at the time of writing.

European Union: under the AI Act, Article 9 requires providers of high-risk AI systems to run risk management as a continuous, iterative process across the system's lifecycle, and Article 14 sets a separate human-oversight requirement. High-risk uses include life and health insurance pricing and creditworthiness assessment of individuals, though AI used to detect financial fraud is excluded. Following the Digital Omnibus, which entered into force on 27 July 2026, these obligations apply to such systems from 2 December 2027. Article 50 transparency duties for chatbots and AI-generated content began applying on 2 August 2026.

United States: the exception. SR 26-2, issued on 17 April 2026, replaced SR 11-7 but places generative and agentic AI outside its scope, with the agencies indicating that separate guidance will follow. Traditional machine learning models remain in scope, and US banks govern GenAI through their existing risk management practices in the meantime.

What Changes in Practice

The core MRM principles stay the same. What changes is how each stage of the lifecycle is applied:

Inventory that finds AI wherever it sits: standalone models, GenAI applications, third-party foundation models, and AI capabilities embedded in business processes or vendor software that were never formally registered. The RBI draft would bar the use of any model not recorded in the inventory.

Tiering with a reliance and autonomy lens: a GenAI tool summarising internal documents does not carry the risk of one whose output drives a customer-facing credit decision without human review. RBI, OSFI and MAS each name reliance or autonomy as a tiering dimension, and the RBI draft adds that one factor should not dilute another, so low technical complexity should not pull down the tier of a highly material use case.

Lifecycle governance in place of one-time approval: registration, validation, deployment, monitoring, change management and revalidation run as connected stages, because prompts, retrieval sources and underlying models keep changing after go-live. A material change to any of them should trigger revalidation.

Validation that tests behaviour: alongside performance, testing covers atypical and adversarial inputs, output variability across repeated runs, grounding against source documents, explainability and the adequacy of human oversight.

Monitoring closer to continuous: GenAI monitoring tracks accuracy and drift alongside output reliability, content safety, data leakage and behavioural shifts after a provider update, with thresholds that trigger review.

Third-party accountability that stays with the institution: a vendor's certification does not discharge the obligation to validate independently, document limitations and maintain oversight. The RBI draft states this explicitly, and E-23 expects externally developed models to be risk-rated on a standalone basis.

A Wider Definition of the Model

GenAI widens what “the model” means. For governance purposes, the record for a GenAI system has to capture the foundation model and its version, the prompts and system instructions, the retrieval corpus, the guardrail configuration and the third-party dependencies, because a change to any one of them can change behaviour. An inventory entry that records only the application name and the vendor cannot support validation, change control or a supervisor's questions.

How Solytics Partners Can Help

Solytics Partners' AI Governance ecosystem brings GenAI and agentic systems under the same governance, validation and monitoring discipline as the rest of the model book, and integrates with the testing tools and pipelines institutions already run.

One inventory for every model type: MRM Vault holds traditional, ML, GenAI and agentic models, including third-party and embedded AI use cases, with risk tiering, ownership, lineage and approval trails.

Configurable risk tiering: tiering follows the institution's own framework, so reliance and autonomy can be scored for AI use cases while traditional models keep the standard methodology.

Lifecycle workflows and audit trail: a configurable rule engine routes approvals, actions and revalidation by team and risk tier, and an audit trail records major and minor changes across the inventory.

Validation and monitoring in one record: NIMBUS Uno supports model validation and performance monitoring, including GenAI and RAG evaluation. Monitoring results flow into MRM Vault, where a threshold breach can trigger review and the findings are recorded against the model.

Automated documentation: VaultBot and Nimbus AI Mate capture metadata and generate documentation, reducing the manual effort of keeping GenAI records current.

Use our ecosystem to bring GenAI to establish a centralized inventory, apply AI-specific risk tiering and governance workflows, connect validation and continuous monitoring, and maintain a complete, auditable record of risks, controls, changes, evidence, and approvals throughout the AI lifecycle.