Finantrix
Subscribe
ArticleRisk & Compliance

How to Tier Models and AI by Risk to Scale Validation

US, UK and Canadian model risk guidance ties validation effort to risk but leaves tiering to the firm. Here are seven steps to build a defensible scheme.

Finantrix Editorial Team 9 min readOctober 29, 2026

Key Takeaways

  • The US, UK and Canadian model risk texts reviewed all tie validation depth and frequency to model risk, but none prescribes tier counts or scoring formulas.
  • Rate materiality (exposure and purpose) and inherent risk (complexity, data, assumptions) separately so the reason for each tier is visible.
  • Add AI factors such as explainability, autonomy, retraining and human review to inherent risk; SR 26-2 excludes generative and agentic AI.
  • Map each tier to validation scope, monitoring frequency, revalidation cycle and approval authority, and record why each interval was chosen.
  • Re-tier on trigger events, reassess the tier at every validation, and have the tiering scheme itself reviewed.

Current as of October 29, 2026. This guide draws on US, UK and Canadian model risk guidance, US state insurance rules and guidance, the EU AI Act and NIST, as published on that date. Which of them apply depends on the firm; it is not legal advice.

No model risk function can review every model in equal depth. The three model risk texts reviewed for this guide give the same answer: depth and frequency should follow the model's risk. The US interagency guidance issued by the Federal Reserve, the OCC and the FDIC in April 2026 says "The nature and rigor of validation generally align with the model's approach, use, and materiality." None of the three tells a firm how many tiers to have, how to score a model or how often to revalidate.

This guide shows how to build a tiering scheme that can be followed from criteria to validation plan. Unless a step quotes a rule, it is a recommendation, labeled (interpretation); the sample scheme is illustrative and no regulator reviewed here requires it.

What is settled, and what is not
  • Settled: within their scope, the three model risk texts reviewed tie validation effort to model risk. The UK PRA expects the rigor, intensity and frequency of validation, monitoring and revalidation to be "commensurate with the associated model tier assigned to a model". Canada's OSFI, in a guideline that takes effect May 1, 2027, says "The extent and frequency of model reviews should be commensurate with the model risk rating."
  • Settled: none of the three model risk texts prescribes a number of tiers, a scoring formula or fixed revalidation intervals. The PRA builds its expectations around a "model tier"; OSFI uses a "model risk rating", which it defines as "a categorical model risk tier"; SR 26-2 does not use the word tier and speaks of materiality.
  • Not settled: how to rate generative and agentic AI at US banks. The 2026 US guidance says these models "are not within the scope of this guidance" and leaves them to the bank's own risk management and governance practices. OSFI's guideline, which does cover AI and machine learning methods, does not take effect until May 1, 2027.

On scope: the US guidance, SR 26-2 (April 17, 2026), supersedes SR 11-7 and SR 21-8 and is non-binding: it states that non-compliance "will not result in supervisory criticism", while its footnote 1 adds that "supervisory action may result for any violations of law or unsafe or unsound practices stemming from insufficient management of model risk." It is expected to be most relevant to larger banking organizations, above an asset threshold set in the guidance, though it says it may also be relevant to smaller banks with significant exposure to model risk. The PRA's SS1/23 (April 2026 version) applies to UK-incorporated banks, building societies and PRA-designated investment firms with internal model approval for regulatory capital, not to insurers. OSFI's Guideline E-23, final since September 2025, will apply from May 1, 2027 to all federally regulated financial institutions in Canada, banks and insurers included, including foreign bank and insurance branches.

Step 1: Decide what enters the tiering, and give low-risk models a light path

Tiering starts with the inventory. SR 26-2's model definition leaves out simple spreadsheet arithmetic and deterministic rule-based processes with no statistical, economic or financial theory behind them, and for models in scope that matter little it allows a lighter path: "Banking organizations may deem certain models immaterial based on model exposure and purpose. In those cases, model risk management may consist of identifying those models and monitoring model performance and conditions under which the use of those models may become material to the banking organization in the future."

OSFI adds a control: "Institutions may include a risk rating category that implies a negligible level of inherent model risk and therefore exempts such models from the full model lifecycle governance requirements. There should be a robust process to approve and track such exemptions." Record the reason for each exemption and the condition that would end it (interpretation).

SR 26-2 says its principles "apply to traditional statistical and quantitative models and non-generative, non-agentic AI models", so a bank's machine learning credit or fraud models belong in the same tiering as its statistical models. Generative and agentic AI can sit on the same inventory, flagged as governed under the bank's own framework (interpretation).

Step 2: Rate materiality and inherent risk as separate dimensions

The PRA asks firms to assign each model a materiality and a complexity rating; SR 26-2 describes model risk as a model's inherent risk in the context of its materiality; and OSFI's rating reflects "model vulnerabilities and materiality of model impacts". SR 26-2 says "Model purpose, together with model exposure, determines model materiality", and describes inherent risk as reflecting factors "such as the assumptions made in developing the model, the model's complexity, the quality of inputs for the model, and data constraints."

The PRA's Principle 1.3 on model tiering says: "Firms should implement a consistent, firm-wide model tiering approach that assigns a risk-based materiality and complexity rating to each of their models." Complexity is assessed through the factors that drive inherent risk; materiality through size-based measures, such as exposure or the number of customers a model applies to, and the model's purpose. OSFI asks for "clear, measurable criteria for each risk dimension", with qualitative factors that include "business use or purpose, model complexity or level of autonomy, reliability of data inputs, customer impacts, or regulatory risk."

Keeping the two ratings separate shows why a model sits in its tier: a simple model with a large exposure and a complex model with a small one need different validation plans (interpretation).

Step 3: Add AI-specific factors to the inherent risk rating

The PRA says that, where necessary, the complexity assessment may consider the use of alternative and unstructured data, and measures of a model's interpretability, explainability, "transparency, and the potential for designer or data bias" to be present.

NIST's AI Risk Management Framework, which is voluntary, points the same way for AI systems generally: "MANAGE 1.2: Treatment of documented AI risks is prioritized based on impact, likelihood, and available resources or methods." AI factors worth adding to inherent risk (interpretation): whether outcomes can be back-tested, whether the model changes after deployment, how far its outputs can be explained, whether it uses alternative data, and how much a person reviews its output before it is acted on.

Insurers face their own versions. Colorado's regulation for insurers offering individual life, private passenger auto and health benefit plans that use external consumer data, and algorithms and models that use it, requires a governance framework aimed at detecting and remediating unfair discrimination, including "Documented policies, procedures, and processes for assessing and prioritizing risks associated with the deployment of ECDIS, as well as algorithms and predictive models that use ECDIS". New York's Circular Letter No. 7, addressed to insurers authorized in the state and covering AI and external data in underwriting and pricing, expects an approach "reasonable and appropriate to each insurer's business model and the overall complexity and materiality of the risks inherent in using ECDIS and AIS."

Legal classification is a separate input. Under the EU AI Act, AI used to evaluate individuals' creditworthiness or set their credit score (other than for detecting financial fraud), or for risk assessment and pricing of individuals in life and health insurance, is high-risk, with the Act's high-risk requirements applying from December 2, 2027 under Regulation (EU) 2026/1744. Treat that status as raising the tier, not replacing the rating (interpretation).

Step 4: Write tiering criteria, not a formula, into the policy

The PRA expects firms' policies to cover the model tiering approach, including "how model tiers are used to determine the scope, intensity and frequency of model validation" and how often tiering is re-performed. For firms with fewer or less complex models, "the criteria for classifying models into tiers can be materially simpler than for firms with a wider range of models or more complex models."

An illustrative scheme, required by no regulator (interpretation):

  • Highest tier: regulatory capital, financial reporting or decisions on individual customers' credit or insurance; large exposure; hard-to-explain methods, alternative data or little human review; or high-risk status under law.
  • Middle tier: informs material decisions with a person judging the output; moderate exposure; methods that can be benchmarked or back-tested.
  • Lowest tier: limited exposure and purpose, simple methods, easily observed outcomes.

Three tiers or five is the firm's call; two validators applying the criteria should land in the same place (interpretation).

Step 5: Tie each tier to validation scope, intensity, frequency and approval

OSFI spells out what the rating should drive. Under its Principle 2.3, "The scope, scale, and intensity of MRM should be commensurate with the risk introduced by the model", and the inherent model risk rating should drive the "frequency, intensity, and scope of model review", documentation requirements, the level of authority required to approve the model, the "frequency, intensity, and scope of model monitoring" and the interval at which the rating is reassessed.

The PRA ties monitoring and revalidation to the tier: monitoring "should be conducted on an ongoing basis with a frequency determined by the model tier", and periodic revalidation "should be carried out with a frequency that is consistent with the model tier." SR 26-2 makes the same link for testing: "The rigor of effective testing is commensurate with model complexity and materiality."

Put this in the policy as one row per tier, from full independent validation before first use, frequent monitoring and senior approval at the top to a targeted review and lighter monitoring at the bottom (interpretation). Record why each interval was chosen; the three model risk texts do not set them, though New York's circular letter describes testing AI system output at least annually for insurers' underwriting and pricing, and Colorado requires annual reviews of the governance framework.

Step 6: Keep a floor of challenge in the low tiers and watch aggregate risk

A low tier lowers effort; it should not remove scrutiny. SR 26-2 warns that even a sound, accurate model can carry high model risk "if it is misapplied or misused", and says "Sound practice involves assessing model risk both individually and in aggregate." Review dependencies across low-tier models that share data, assumptions or vendors (interpretation).

OSFI says "Externally developed models should be assessed for model risk ratings on a standalone basis".

Step 7: Re-tier on triggers, and validate the tiering itself

OSFI says "Model risk ratings should be regularly reviewed and updated as appropriate, including when a trigger event occurs (for example, a decrease in performance or a material change in the model's use, data, or infrastructure)."

The PRA wants the firm-wide tiering approach periodically validated or reviewed by an informed party, and individual tier assignments "should be independently reassessed as part of the model validation and revalidation process".

Checklist

Items restate the recommendations above (interpretation).

  • Model records show which models are in scope, which are immaterial or negligible-risk, and why
  • Exemptions from full lifecycle governance are approved, tracked and have an end condition
  • Generative and agentic AI are flagged where the US guidance does not reach them
  • Materiality and inherent risk are rated separately, with AI factors feeding inherent risk
  • Legal classifications such as EU AI Act high-risk status are recorded and raise the tier
  • Tiering criteria are written in the policy and give consistent results across validators
  • Each tier maps to validation scope, monitoring frequency, revalidation cycle, documentation and approval authority
  • Aggregate risk across shared data, assumptions and vendors is reviewed
  • Tiers are reassessed on trigger events and at each validation, and the scheme itself is reviewed

How we sourced this guide: every factual statement in it is drawn from published rules, regulatory guidance and other public documents, and we keep a full set of sources and citations for it. If you would like them, or want to question a point, get in touch through our contact page and we will share them.

Frequently Asked Questions

Do regulators require a set number of model risk tiers?

Not in the texts reviewed for this guide. The US interagency guidance, the UK PRA's supervisory statement and Canada's OSFI guideline all expect validation effort to follow model risk, but none prescribes how many tiers to use, a scoring formula or fixed revalidation intervals. The PRA allows smaller or simpler firms to use materially simpler tiering criteria.

Does SR 26-2 require model tiering?

SR 26-2 does not use the word tier. It asks for a risk-based approach in which materiality, meaning model exposure and purpose, together with inherent risk shapes the rigor of oversight, and it says the nature and rigor of validation generally align with a model's approach, use and materiality. The guidance is non-binding, though supervisory action may still follow from unsafe or unsound practices.

How should generative AI be tiered at a US bank?

SR 26-2 places generative and agentic AI outside its scope and leaves them to the bank's own risk management and governance practices. Applying the same materiality and inherent risk criteria, with added factors such as autonomy and human review, is one reasonable approach (interpretation). Non-generative AI models, such as machine learning credit models, are within the guidance and belong in the ordinary tiering.

Can a low-risk model skip validation?

The texts reviewed allow a lighter path, but not an unrecorded one. SR 26-2 says oversight of immaterial models may consist of identifying them and monitoring for conditions that would make them material; it does not mention validating them. OSFI, from May 1, 2027, allows a negligible-risk category exempt from full lifecycle governance, but expects a robust process to approve and track those exemptions.

Share: