PRA SS1/23

• The PRA (part of the Bank of England) published SS1/23 on 17 May 2023; it took effect on 17 May 2024.
• It treats model risk as a risk in its own right, like credit or market risk, and asks for a firm-wide framework with board-set model risk appetite and one named Senior Manager accountable.
• It is built on 5 principles split into 23 numbered sub-principles: (1) model identification and risk classification, (2) governance, (3) development, implementation and use, (4) independent validation, (5) model risk mitigants.
• It covers all models, whatever the technology, built in-house or bought from vendors, including models used for financial reporting.
• Status (as of 2026-10-19): in effect. A light-touch update in April 2026 (LIAF01/26) only clarified paragraph 1.9. No AI-specific supplement has been issued; the PRA instead said in its October 2025 roundtables that SS1/23 is the yardstick for AI and ML.

• UK-incorporated banks, building societies and PRA-designated investment firms that hold internal model approval for regulatory capital: credit risk (IRB), market risk (Internal Model Approach) or counterparty credit risk (Internal Model Method). Para 1.2 and footnote 2.
• Not in scope: firms without internal model permission, UK branches of third-country (non-UK) banks, credit unions, insurers and reinsurers. The PRA says out-of-scope firms may still find the principles useful.
• A firm that newly receives internal model permission has 12 months from that grant to comply (para 1.9).
• Vendors and consultants are not addressed directly, but footnote 3 notes they may find the SS useful because it sets the minimum bar their bank clients must meet.
• Not extraterritorial. A non-UK company is reached only as a vendor, through the bank's own duties (Principle 2.6).

• The SS states no penalties. It sets out what the PRA expects; it is not itself a rule with fines attached.
• In practice the PRA tests it through supervision: firms must keep a self-assessment and remediation plan and give them to the PRA on request (para 1.8).
• Accountability is personal: the chosen SMF must have MRM written into their Statement of Responsibilities (Principle 2.2), which ties model failures to the UK Senior Managers regime.
• April 2026 clarification: meeting SS1/23 is not a condition for getting internal model approval (para 1.9), but the expectations still apply once approval is held.

• Indian banks with UK operations: check how you operate. A UK branch of an Indian bank is out of scope (third-country branch). A UK-incorporated subsidiary is in scope only if it holds internal model approval for capital. If neither applies, SS1/23 is optional good practice, though UK supervisors may still use it as a benchmark.
• IT and analytics vendors (model builders, scorecard vendors, AI platform providers) selling to UK banks: your client must validate your model to the same standard as its own (Principle 2.6), verify your data and assumptions, and monitor your model on its own outcomes. Expect requests for development documentation detailed enough for an independent expert to replicate results (Principle 3.5), data lineage, bias testing evidence and monitoring packs.
• Group model hubs in India (for example a captive centre building models for a UK entity): a UK subsidiary may rely on the parent's validation only if the group's standards match SS1/23 and the subsidiary checks relevance and tier on a standalone basis (Principle 2.6(c)).
• Comparison with RBI: RBI's draft Guidance on Regulatory Principles for Model Risk Management (24 June 2026, still draft as of 2026-10-19) covers AI and ML models and picks up themes from RBI FREE-AI such as board-approved frameworks and model inventories. SS1/23 is a useful, more detailed template for what "good" looks like: model definition, tiering by materiality and complexity, SMF-style accountability, PMA governance.
• One SS1/23-aligned MRM framework, extended with AI security and fairness controls, can serve UK, RBI and OSFI E-23 expectations together.

Principle and sub-principle numbers are from section 4 of the April 2026 SS text; paragraph numbers are from sections 1 and 3.

• The latest annual self-assessment against SS1/23, remediation plan and board minutes showing they were presented.
• The SMF Statement of Responsibilities naming MRM.
• Board-approved model risk appetite statement with its measures and the latest report against it.
• MRM policy and procedures, with version history and date of last review.
• Model inventory extract (including decommissioned and in-development models) with tier, owner, validation dates and open findings.
• Tiering methodology, plus evidence that AI and ML models were rated using explainability, bias and data complexity factors.
• Development documentation and validation reports for a sample of high-tier models, including vendor models.
• Monitoring packs, threshold breaches and actions taken.
• PMA register with approvals, independent review and results shown with and without PMAs.
• Exceptions log and escalation records.
• Internal audit reports on MRM; the annual audit committee report on financial reporting models.

• Assuming SS1/23 covers only capital models. Once a firm is in scope, it covers all models used for business decisions, including AI and ML scoring models, marketing models and financial reporting models.
• Writing a tiering policy that puts AI and ML at medium complexity or higher, then logging them as low in the inventory. The PRA called out this exact gap in its 2025 roundtable.
• Monitoring dynamic AI models on a six-monthly cycle; the PRA said that may be too slow and suggested fallback models and, in extremis, kill switches.
• Treating SHAP or LIME outputs as proof of explainability. The PRA's slides warned these tools can mislead when input features are correlated.
• Relying on a vendor's own validation without checking it against internal standards or monitoring on the bank's own outcomes.
• Letting overlays (PMAs) pile up year after year instead of fixing the model.
• Assuming a UK branch of an Indian bank must comply; branches of third-country banks are out of scope.

• PRA: Prudential Regulation Authority, the part of the Bank of England that supervises banks, building societies and insurers for safety and soundness.
• Supervisory Statement (SS): a PRA document setting out what it expects firms to do; it guides supervision rather than creating fines by itself.
• Model risk: the chance of loss or harm from a model being wrong or being used for the wrong job.
• IRB (Internal Ratings Based): an approach where a bank uses its own approved models to calculate capital for credit risk.
• SMF (Senior Management Function): a named senior role under the UK Senior Managers regime, with personal accountability.
• Model tiering: ranking models by how much damage they could do (materiality) and how hard they are to get right (complexity).
• Validation: an independent check that a model is sound and fit for its intended use.
• Post-model adjustment (PMA): an overlay, override or manual tweak applied to a model's output because the model misses some risk.
• Operating boundaries: the range of data a model was built on; using it outside that range adds risk.
• Dynamic model: a model that recalibrates or changes its parameters by itself as new data arrives, which includes many machine learning models.
• Challenger model: an alternative model used to compare against and test the main one.

• SS1/23 is the UK's model risk rulebook for banks with internal capital models; in effect since 17 May 2024 and still current after a minor April 2026 clarification.
• It is technology neutral, so predictive AI and ML models are covered as ordinary models, with AI traits (opacity, bias, unstructured data, self-updating) pushing them into higher tiers.
• 30 obligations across 5 principles and 23 sub-principles, plus self-assessment and audit committee reporting duties.
• The PRA's 2025 AI roundtable gave practical signals: align tiering with inventory, monitor dynamic models more often, be sceptical of explainability tools, plan fallbacks and kill switches.
• Indian banks' UK branches are out of scope; Indian vendors are reached through their UK bank clients' vendor validation duties.

• SS1/23 Model risk management principles for banks, PRA publication page, accessed 2026-10-19
• SS1/23 current text, April 2026 (LIAF01/26 appendix), accessed 2026-10-19
• SS1/23 original text, May 2023, accessed 2026-10-19
• LIAF01/26 Low Impact Amendments Finalisation, April 2026, accessed 2026-10-19
• PRA model risk management roundtable on AI and ML (November 2025), accessed 2026-10-19
• Roundtable slides, October 2025, accessed 2026-10-19