SS1/23 and SR 26-2: model risk management beyond the model
Both regimes centre on the model. A large share of model risk sits in the process around it.
SS1/23 is the Prudential Regulation Authority's supervisory statement setting out five principles for model risk management in UK banks. It took effect on 17 May 2024 for banks, building societies and PRA-designated investment firms with approval to use internal models for regulatory capital, and the PRA amended it on 23 April 2026 to clarify that its expectations are not conditions for internal model approval. It covers every model used to inform business decisions, including vendor models, financial reporting models and AI. In the United States, SR 26-2 replaced SR 11-7 in April 2026 as the banking agencies' model risk guidance.
The five SS1/23 principles
SS1/23 treats model risk as a risk discipline in its own right, with responsibility for the framework allocated to a Senior Management Function holder.
- Model identification and model risk classification. A clear definition of what counts as a model, a complete inventory, and a tiering of each model by its risk.
- Governance. Board-level oversight, a model risk management policy, and a Senior Management Function holder responsible for the framework.
- Model development, implementation and use. Robust development standards, including the data a model is built on, testing before use, and use within the model's intended purpose.
- Independent model validation. Validation independent of development, scaled to the model's tier, with ongoing performance monitoring.
- Model risk mitigants. Controls for when a model is uncertain or limited, including the governance of post-model adjustments, and restrictions on use where needed.
The PRA expects firms to manage the risks of AI and machine learning techniques within this framework, as far as they apply to models generally, rather than in a separate one.
SR 26-2: what replaced SR 11-7
On 17 April 2026 the Federal Reserve, the OCC and the FDIC issued SR 26-2, Revised Guidance on Model Risk Management, superseding SR 11-7 (2011) and the 2021 interagency statement on models supporting anti-money-laundering compliance. The agencies expect it to be most relevant to banking organisations with more than $30 billion in total assets. It is guidance, not a rule: it sets no enforceable standards.
The main shifts are towards proportionality. The approach is risk-based and tailored to each organisation's model risk profile, and a model's materiality is set by its purpose and exposure. A model is defined as a complex quantitative method that applies statistical, economic or financial theories to turn input data into quantitative estimates. Simple arithmetic calculations, such as those found within spreadsheets, fall outside that definition, and so do deterministic rule-based processes with no such theory behind them. Validation, effective challenge by people with sufficient independence, and ongoing monitoring all remain, at a rigour matched to the risk. The principles still apply to vendor models, and each organisation remains responsible for its own model risk management.
SR 26-2 also draws a line around AI. A footnote states that generative AI and agentic AI models are novel and rapidly evolving and are not within the scope of the guidance, which applies to traditional statistical and quantitative models and to non-generative, non-agentic AI models. Governance of generative and agentic AI is left to each banking organisation to determine.
Where model risk actually sits
Both regimes centre on the model: its inventory, development, validation and limits. In practice a large share of model risk sits in the process around the model.
Upstream, in the inputs. A model is validated on the data it was built with. In production it receives data that has passed through extraction jobs, mappings, transformations in application code and, often, a spreadsheet. If a mapping truncates a field, or a rule upstream changes, the model's inputs change with no change to the model. Validation does not see it, because nothing in the model changed.
Downstream, in the use. The model's output becomes a capital figure, a provision, a price or a credit decision only after further steps: overlays, post-model adjustments, aggregation, manual review. SS1/23 asks for those adjustments to be governed. Many live in spreadsheets and email approvals.
At the boundary, in what counts as a model. SR 26-2 excludes simple spreadsheet arithmetic and deterministic rule-based processes. That is sensible for proportionality, but it moves the risk rather than removing it. A spreadsheet or a rule that is not a model can still change the outcome a model produced.
So the question a validator or a supervisor eventually asks, "how was this outcome produced?", is not answered by the model documentation alone. It needs the path from the source data, through the model, to the outcome that was reported or the decision that was made.
Model risk management as process context
xflow builds process context: one executable view of the process a model sits in, from the objective the outcome serves, through the steps, code, data, spreadsheets and manual adjustments, to the output. For model risk that gives three things.
The inputs are traceable to source at the grain the model uses, including the transformations held in code, so a change upstream is visible before it reaches the model. The adjustments and overlays after the model are part of the same view, with their owners and approvals, rather than a separate log. And because real data can be run through the represented process, a figure can be tested against the objective it serves, which is a stronger check than agreement with its sources.
That view matters more as models become agents, and SR 26-2 leaves agentic AI outside its scope, so the governance is for each institution to design. An agent applying a rule in a regulatory process needs the approved rule, not its own copy in a prompt. When the rule is held once as a governed definition and the agent points at it, the rule applied at runtime is the one that was designed, validated and evidenced.
At a global tier one bank, lineage for regulatory reporting stopped where the application code began. xflow connected to the code repositories and produced the first analysis in hours. Transformation gaps were closed and written back to the bank's data governance layer, giving end-to-end lineage from policy to process to code to data.
Questions about SS1/23 and SR 26-2
Who does SS1/23 apply to?
UK-incorporated banks, building societies and PRA-designated investment firms with approval to use internal models for regulatory capital (credit, market or counterparty credit risk). Other firms continue to apply the existing supervisory expectations relevant to them.
When did SS1/23 take effect?
17 May 2024. It was published on 17 May 2023 and amended on 23 April 2026 to clarify that its expectations are not conditions for internal model approval.
Does SR 26-2 cover AI agents?
No. A footnote places generative AI and agentic AI models outside its scope, as novel and rapidly evolving. It applies to traditional statistical and quantitative models and to non-generative, non-agentic AI models, and leaves governance of the rest to each banking organisation.
Does SS1/23 cover AI models?
Yes. It applies to all models used to inform business decisions, whatever the technique, and expects AI risks to be managed within the model risk framework.
What is SR 26-2?
The US banking agencies' revised model risk management guidance, issued on 17 April 2026, which superseded SR 11-7.
What changed from SR 11-7 to SR 26-2?
A more explicitly risk-based approach, a definition of a model that excludes simple spreadsheet arithmetic and deterministic rule-based processes, validation and monitoring at a rigour matched to risk, and continued application of the principles to vendor models. Generative and agentic AI are outside its scope.
Is a spreadsheet a model?
Under SR 26-2, simple arithmetic in a spreadsheet is not. A spreadsheet that applies statistical, economic or financial theory to produce estimates could still meet the definition. Under SS1/23, the test is whether it is a quantitative method used to inform business decisions. Either way, a spreadsheet in the path of a reported outcome is part of that outcome's risk.