Model risk tiering decides which models get enhanced scrutiny and which face fewer requirements. Each identified model is assigned a tier through a consistent, firm-wide tiering approach. That tier then drives the scope and frequency of validation, monitoring and revalidation, and the level of authority needed to approve the model. Model owners and users can challenge a tier and escalate, but in established practice the head of the model risk function has final sign-off on tier assignments.
Classification drives everything downstream. A product recommender and the credit model approving loans at the same institution carry different levels of risk, so they get different validation depth, gate severity, monitoring cadence and sign-off authority. The PRA’s supervisory statement SS1/23 expects the intensity, prioritisation and frequency of validation, independent review, performance monitoring and revalidation to match the assigned tier. OSFI’s Guideline E-23 has the inherent risk rating drive review scope, documentation requirements, approval authority, monitoring intensity and the interval at which the rating itself is re-assessed.
Identification decides what enters the tiering at all
Only things the institution has identified as models get tiered, so the definition and the inventory come before any tier. SR 26-2 defines a model as a complex quantitative method, system or approach that applies statistical, economic or financial theories to process input data into quantitative estimates. The same guidance excludes simple arithmetic calculations such as those found in spreadsheets, along with deterministic rule-based processes that have no statistical, economic or financial theory behind them.
OSFI’s E-23 asks institutions to periodically identify models used throughout the enterprise, including vendor and third-party models. Each one is triaged on whether it carries non-negligible inherent risk, and new models get a provisional rating. One practitioner guide gives a workable rule to write down: anything that produces a quantitative or classification output driving a business decision is a model. A dispute over a tier can turn out to be a dispute over whether the thing is a model at all, and that question needs settling before anyone argues about high versus medium.
Has SR 26-2 replaced SR 11-7, and does that change how models are tiered?
Yes. On 17 April 2026 the Federal Reserve, OCC and FDIC issued SR 26-2, revised guidance on model risk management. It supersedes and replaces SR 11-7 (April 2011) and SR 21-8, the 2021 statement on models used for BSA/AML compliance. SR 26-2 moves away from detailed supervisory expectations towards a principle-based, risk-tailored approach. It puts more responsibility on each organisation to decide how model risk management applies, based on model purpose, exposure, complexity and risk.
| Status | Superseded | Current; supersedes SR 11-7 and SR 21-8 |
| Who it addresses | Fed-supervised banking organisations, applied according to size, complexity and use of models | Most relevant above $30 billion in total assets; may also apply to smaller firms with significant model risk |
| Model definition | Baseline definition | Narrowed: excludes spreadsheet arithmetic and deterministic rule-based processes |
| What drives governance depth | Materiality; risk rises with complexity, input uncertainty, extent of use and potential impact | Materiality (purpose plus exposure), read in the context of inherent risk |
| Review cadence | Periodic review at least annually | Validation rigour and monitoring frequency set by materiality |
| Binding force | Supervisory guidance | States it sets no enforceable standards and that non-compliance alone will not bring supervisory criticism |
| Generative and agentic AI | Not addressed | Explicitly outside scope |
Under SR 26-2, model purpose together with model exposure determines materiality. Model exposure is how significant the model’s output is to business decisions, and it can be measured quantitatively, for example by portfolio size. Model purpose is a qualitative consideration about the nature and importance of the model. Models built to help meet regulatory requirements or to manage financial risk exposures are generally treated as higher risk. Inherent risk sits alongside this. It reflects the assumptions made in development, complexity, input quality and data constraints, and it rises with complexity and with how many or how critical the assumptions are. Higher-materiality models warrant more comprehensive and rigorous oversight. For models judged immaterial, appropriate management may come down to identifying them and monitoring for the conditions under which they could become material.
So the logic of tiering by materiality carries over. SR 11-7 already tied the depth of the framework to materiality and potential impact. What changes is who owns the calibration. With fewer prescriptions, the institution’s own criteria and rationale carry more of the weight, and a departure from the guidance holds up when a sound risk-based rationale supports it. That makes the documented tiering method more important, not less.
One boundary needs stating once. SR 26-2 applies to traditional statistical and quantitative models and to non-generative, non-agentic AI. Generative and agentic AI are left to the institution’s broader risk management and governance practices. Those systems belong in an AI governance change process, and a team should not claim SR 26-2 coverage for them.
Which published framework can we point to for our tiering decision?
A defensible tiering process adapts a recognised model risk framework rather than inventing one. That lifecycle runs from model definition, inventory and tiering through validation, monitoring and named accountability. The PRA’s SS1/23 is built on five high-level principles covering the whole model lifecycle. ASOP No. 56 covers actuarial modelling in any practice area, and the NIST AI RMF is voluntary, not tied to any sector and agnostic to use case.
TechnoLynx can act as an independent validator of churn and dynamic-pricing model methodology before production promotion. This sits within its AI Governance and Trust assurance work, and the AI Readiness Scorecard is the first step on that path.
What criteria decide whether a model is high, medium or low risk tier?
Most financial firms use explicit procedures and decision tools to assign models to a small number of tiers rather than ranking every model individually. Kiritz, Ravitz and Levonian’s 2019 survey of industry practice found that these tools are either scorecards or decision trees, with scorecards the more common choice. One approach treats model risk tiering like credit analysis. They score the probability of model failure and the impact of model failure as two separate dimensions.
To gauge impact, firms combine the value of the items modelled, the sensitivity of those values to model error, the number of customers exposed and how much the final result relies on the model. Firms also make categorical distinctions among model uses. General business uses such as direct mail targeting are usually seen as lowest impact, and financial accounting and regulatory reporting as highest. OSFI asks for clear, measurable criteria that mix quantitative factors (portfolio size and growth, operational or financial impact) with qualitative ones (business use, complexity or autonomy, reliability of data inputs, customer impact, regulatory risk). The FHFA’s 2013 advisory bulletin reduces this to three drivers: complexity, business impact and extent of use.
For a loan approval model, we read every impact factor as pointing the same way: a credit decision, many customers affected, regulatory and consumer consequences. It belongs in the top tier unless something unusual limits its reliance. On the choice of tool, a scorecard suits a portfolio where many factors trade off against each other. A decision tree, or a scorecard with explicit overrides, suits cases where one fact should settle the tier regardless of score, such as use in regulatory stress testing.
Tiering earns acceptance when it is transparent and consistent. The same survey’s principles warn against false precision: put the weight on the tier assignment rather than on intermediate scores that look like granular risk metrics, and accept that expert judgement plays a significant role. The tools systematise judgement rather than measure risk directly. The spread of models across tiers gives a quick calibration check. Kiritz, Ravitz and Levonian suggest that in a three- or four-tier system, a tier holding fewer than about 10% of inventoried models is probably too narrow, and a higher tier holding more than about 30-40% is probably too broad.
Tier on inherent risk, not residual risk. If governance factors feed the tier, the result is circular: low governance raises the tier, extra oversight lowers it at the next review, and the assignment never settles. Residual-risk tiering also makes a validated model impossible to compare with an unvalidated one. Trading algorithms are the known exception: the FMSB’s good-practice statement allows strong mitigating controls to place a high-inherent-risk algorithm in a low tier, with validation then focused on whether those controls work.
Tiering a longevity model before it touches statutory reserves
A longevity model that feeds statutory reserves falls in the use category firms most often treat as highest risk: financial accounting and regulatory reporting. For insurers, ASOP No. 56 makes the appropriate degree of model risk mitigation depend on the model’s purpose, its complexity, its governance and controls, any changes to the model or its environment, and the cost of mitigation set against the risk it removes. Delayed feedback complicates validation of these models. Globebyte’s guide to applying the NIST AI RMF in actuarial practice notes that ground truth for long-duration predictions only emerges over years, so the back-testing method has to address delayed-feedback bias explicitly. The same guide expects models feeding financial reporting to be monitored against actual experience on a quarterly or monthly basis, with a documented process for refreshing the model when experience diverges.
The monitoring commitments belong in the approval record before go-live. Under OSFI’s E-23, an approval decision affirms the assigned risk rating and the residual risk assessment, not just whether the model is suitable. The model risk policy defines the requirements and frequency of performance monitoring and who is responsible for it. Response thresholds should be set and documented before deployment, not improvised when something goes wrong. How the approval record, evidence and escalation path fit together at release is covered in our piece on building defensible governance around model releases.
The model owner informs the model risk function of changes and engages validation to re-evaluate the model’s risk regularly. The validation team evaluates whether the change triggers a validation event. A revised mortality table is a change to a key input, so the approval should name who raises it and who decides whether it needs full or targeted revalidation.
A standing monitoring cadence for the recommender fleet
Monitoring frequency for a fleet of recommendation models should be set by tier and written into policy. One AI governance playbook suggests weekly or even daily monitoring for high-risk systems in fast-changing environments, and monthly or quarterly reviews for lower-risk systems in stable ones. For comparison, the FHFA’s 2013 bulletin expected mission-critical or high-risk models to be monitored at least quarterly against thresholds the model risk group approves annually. Recommenders resemble general business uses, which most firms see as lowest impact, so daily monitoring is justified where a region’s data shifts quickly, not by default.
Escalation needs a named owner and thresholds set in advance. Monitoring needs an explicit owner role, because without one it loses out to new deployments. One playbook sets graded thresholds, with yellow flagging a model for review and red pausing it for investigation. The PRA expects escalation processes that promptly alert model owners, users, validators and senior management to an exception. Persistent breach of performance metrics is one of its defined exceptions. The mechanics of drift detection are covered in our article on validation, drift detection and continuous monitoring in the ML lifecycle.
A quarterly one-page summary of drift status by region fits what tiers are for: Kiritz and colleagues note that senior management relies on tiers for clear, concise model risk reporting, and the FHFA expects the CRO to report the aggregate level and direction of model risk to the board regularly. That summary is only as credible as the team’s ability to run the cadence the same way in every region.
Which parts of SR 26-2 apply to a churn or pricing model, or outside a bank?
SR 26-2 is supervisory guidance for banking organisations, most relevant to those above $30 billion in total assets. A media company’s pricing model sits outside it entirely. Regulation often does not reach a given model, so the first check is whether a given model’s use actually falls inside a regulated scope.
Legal scope and governance scope are separate questions. Globebyte gives the example of an insurer’s lapse model used only for hedging and reserving. It is less likely to be the direct legal focus of consumer-facing AI regimes than a system driving consumer pricing, but it still falls under ASOP No. 56, the firm’s own model risk policy and, where relevant, Solvency II validation expectations.
For a non-bank churn or pricing model, the core of the lifecycle carries over: an inventory, materiality read as purpose plus exposure (revenue and customers affected instead of portfolio size), independent challenge, and tiered monitoring. The NIST AI RMF offers a non-sector-specific basis for this. It raises initial priority where outputs affect people directly or indirectly, and it leaves risk tolerance for each organisation to set. An interview study of financial organisations found existing model risk frameworks broadly adequate when AI is built into existing models, with targeted adaptation needed when AI drives novel applications. Adapting a bank framework to a non-credit model can under-apply or over-apply its provisions, which is why the mapping itself deserves an independent read.
If a model’s tier is wrong or its inputs change, who re-tiers it?
The model risk function re-tiers it, with individual tier assignments independently reassessed. The established pattern puts ownership of the tiering tool with that function, gives its head final sign-off, and lets model owners and users challenge and escalate. For a vendor credit-scoring model heading to committee, the first step is factual: check the inventory for an existing tier. OSFI expects vendor and third-party models to be identified, triaged and given a provisional rating like any other.
Buying a model does not hand off the obligation. Under SR 26-2, sound practice for vendor models means understanding the model’s conceptual soundness, design, development data and performance, then running ongoing monitoring and outcome analysis. The PRA accepts that vendor documentation will be thinner than for internal models, but still expects enough detail to validate the firm’s use of the model. One practitioner guide lists opacity among the factors for rating inherent risk and materiality. More generally, the PRA wants the extent of challenger or benchmark comparison considered in tier classification, reflecting higher model uncertainty. Under OSFI’s E-23, the rating then sets the frequency, intensity and scope of model review.
Re-tiering is triggered by events and by schedule. OSFI expects a rating to be updated when a model has materially changed in substance or usage. Strong triggers include a new decision use, expansion to a materially different population, changed output authority and a new material limitation, and policy should map each trigger to an approval authority. New information about model risk is one reason to change a tier. Frequent post-implementation revisions undercut acceptance of the whole framework, though, so stability is a goal and not an absolute.
On schedule, most firms re-evaluate tier assignments at least annually, including the profile elements used to set them. The FHFA expected its model risk group to review all risk classifications annually. The PRA wants individual tier assignments independently reassessed during validation and revalidation. One practitioner puts the stakes plainly: a classification with no review date is a guess with a number attached.
Disputes need a defined path. The FHFA asked for an escalation policy and dispute resolution process with time limits, with the management-level model risk committee as the last stop before the board. Whoever approves the model must make sure the validation’s remediation recommendations are actioned. One caution for anyone benchmarking against peers: a similar model can present different risks at two institutions, so peer practices do not transfer automatically.
Does a logistic-regression scorecard need the same tier as a neural network or a vendor gradient-boosted model?
If both tools serve the same decision and an error has the same consequence, they start in the same tier. Tiering follows use and impact, not algorithm family. Kiritz and colleagues make the point directly: inherent model risk refers to the risk of a model and its uses, not the intrinsic riskiness of its design.
The IMDRF’s categorisation of software as a medical device offers a parallel from healthcare. It categorises by two factors: how significant the information is to the healthcare decision, and the state of the healthcare situation or condition. That gives four categories, with Category IV the highest impact and Category I the lowest. Image analysis for treatment decisions in acute stroke is a Category IV example. Heart-rate analysis that helps a clinician diagnose arrhythmia is Category II. Algorithm type is not one of the two factors, so a regression score and a neural network with the same intended use land in the same category. The IMDRF framework is not a regulatory classification and does not prescribe controls by itself. The category should be re-evaluated whenever the tool’s definition statement changes. A list-based legal tier and a contextual measured risk can disagree, and a sound taxonomy records both rather than averaging them.
SR 26-2 says inherent risk increases with complexity. The PRA expects tiering to consider at least one metric for complexity. OSFI allows extra requirements by model type on top of the primary risk-based ones and expects heavy AI/ML use to come with matching governance maturity. So use and impact drive the tier, and complexity, opacity and data uncertainty can raise it or add requirements. Simplicity alone never justifies dropping a high-impact model a tier.
For an examiner, run both models through the same tool with the same factors and record each factor and the rationale. Show where the gradient-boosted model’s opacity, vendor dependence or data complexity moved it, rather than arguing that one is “a scorecard”. One nuance from SR 26-2’s narrower definition: the spreadsheet exclusion covers simple arithmetic, not the medium. A logistic regression fitted to data applies statistical theory and stays a model wherever it is computed. A hand-set points table with no statistical basis may genuinely fall outside the definition, and that is a scope decision to document, not a tiering one.
What changes if a model is re-tiered one level up or down?
Moving a model one tier changes the scope and rigour of validation, the depth of independent review, monitoring and revalidation frequency, documentation, and the level of authority needed to approve it. TechnoLynx, through AI Governance and Trust, helps where the gap lies in the technical evidence a tier calls for, not in the tier decision itself.
| What moves | One tier up | One tier down |
|---|---|---|
| Validation scope and rigour | Larger scope, more and more severe tests across all areas of the model | Narrower, proportionate review |
| Independent review | Deeper review, set by the tier | Lighter review, still independent |
| Monitoring frequency | More frequent, as set by the tier | Less frequent |
| Revalidation | More frequent periodic revalidation | Longer revalidation interval |
| Documentation | More extensive, in line with the model’s importance | Less extensive |
| Approval authority | Higher level of authority to approve | Lower approval level |
| Usage limits and safeguards | More likely constraints, extra safeguards, closer monitoring | Fewer constraints |
Before finalising documentation depth, confirm the tier and its provenance. The PRA expects the documented tiering approach to cover its data sources, how tiers set the scope, intensity and frequency of validation, who assesses materiality and complexity, and how often tiers are re-assessed. It also expects confirmation that the model is assigned to the correct tier and has been validated in line with it. If the documentation meets or exceeds what policy requires for the tier, approval can proceed. If it falls short, either remediate the documentation or reopen the tier on its complexity and impact. The second route has a risk: tiering a borderline model down to shrink the documentation burden understates its risk, so the tier record should show why the move was made. How a confirmed tier becomes a specific evidence pack is covered in our piece on scoping validation depth from risk tiering to evidence packs.
Internal tier tables also need to track regulatory dates. OSFI’s revised E-23, for example, only takes effect on 1 May 2027, so an internal tier-to-documentation table written against older guidance may need revisiting.
Scaling tier and evidence depth across many sites without duplicating the pack
None of the sources reviewed here addresses multi-site deployment directly, so what follows is our judgement built from general tiering principles. Tier the model and its use once, keep one core evidence base for the model, and add a site-level delta only where something material differs at that site.
Kiritz and colleagues describe firms that tier each model-and-use pair and base requirements on the highest tier, as well as the more common approach of one tier per model covering all its uses. The Society of Actuaries’ newsletter argues that SR 26-2 makes room to vary validation depth and frequency by use case instead of revalidating the whole model at once. In a cross-framework setting, one study describes a single control register, with one control library and one evidence base, that cuts the cost of running parallel programmes. The 2021 BSA/AML statement, now superseded by SR 26-2, said there was no requirement to duplicate independent testing.
A site warrants its own record when it has a distinct purpose, owner, population, methodology or monitoring plan. Otherwise the same inventory record continues with version history. Expansion to a materially different population is itself a re-tiering trigger. In practice, the core pack covers conceptual soundness and development data once, and each site adds evidence that the model still holds on its own inputs, plus monitoring broken out by site so a local failure does not hide inside an aggregate.
Who can help us put defensible tiering and validation around our AI models?
TechnoLynx provides the independent technical challenge a tiering decision relies on, such as validating churn and dynamic-pricing model methodology before production promotion, while your model risk function keeps the tier and the sign-off. On building in-house versus bringing in a specialist: the FHFA notes that validators may be internal staff, external consultants or both, and that more complex models may warrant outside experts to supplement internal validation. When choosing a partner, check independence. Validators should have no responsibility for developing or using the model and no stake in the outcome. They also need enough familiarity with the business use to give an objective opinion. The engagement runs through AI Governance and Trust, starting with the AI Readiness Scorecard.
This does not suit a team that wants an outside party to set its tiers, give a regulatory opinion or provide actuarial sign-off. The tier stays with the institution, where established practice gives the model risk function’s head final sign-off. For actuarial models, ASOP No. 56 points to review by a second qualified professional, with the depth of review increasing with model complexity and materiality. A firm with an established, independent validation team and the ML skills in-house may need no outside help at all.
The tiering method needs reviewing too
The PRA expects the firm-wide tiering approach to be periodically validated or critically reviewed by an informed party. That review covers whether the information used to set tiers is still relevant and available, and whether tiers are recorded accurately in the inventory. Kiritz and colleagues describe an emerging practice in some firms of subjecting tiering tools to the same documentation, validation, periodic review and tiering as other models.
The SOA newsletter offers a corrective here: inventories and tiering are supporting tools, not the end point. What a committee needs to know is whether model-informed decisions are producing the intended outcomes.
FAQ
What is model risk?
The PRA defines model risk as the potential for adverse consequences from model errors or from inappropriate use of model outputs to inform business decisions. SR 11-7 noted that it increases with complexity, uncertainty about inputs and assumptions, breadth of use and potential impact.
What are the key components of a model risk management framework?
Model risk management policies typically cover model and model risk definitions, stakeholder roles, risk assessment, development and use practices, validation, and governance and controls over the process. One current guide lists them as identification and classification, governance, development, independent validation and mitigants.
Who is responsible for model risk management in a bank?
The board of directors and senior management are ultimately responsible for a sound framework that stays within the board’s model risk appetite. The work can follow three lines of defence: model owners, model risk management and internal audit.
