What does PRA SS1/23 say about AI in banking?

What are the key points of PRA SS1/23?

• Five principles: (1) model identification and model risk classification, (2) governance, (3) model development, implementation and use, (4) independent model validation, (5) model risk mitigants.
• Scope is all models informing business decisions, regardless of technology, whether in-house or vendor-supplied, including models used for financial reporting.
• AI and ML are addressed through general principles, not by name: Principle 1.3(c) complexity factors for 'newly advanced approaches or technologies' (alternative and unstructured data; interpretability, explainability, transparency and potential for designer or data bias) and Principle 3.3(c) testing of material changes in dynamic models that adapt or recalibrate autonomously. The PRA's publication page summarises this as a sub-principle on AI in modelling techniques such as machine learning.
• Accountability for the overall MRM framework must be allocated to the most appropriate SMF holder under the Senior Managers regime.
• Effectiveness of MRM for financial reporting must be reported to the audit committee.
• Proportionate implementation across model tiers and across firms; the expectations apply only to internal-model banks, and the PRA says other firms may find the principles useful.
• In October 2025 the PRA held CRO roundtables with 21 firms on applying SS1/23 to AI/ML, covering risk appetite, model tiering, explainability, overfitting, validation and monitoring.
• Amended 23 April 2026 (LIAF01/26): the expectations are not conditions for internal model approval, newly permitted internal-model firms have 12 months to comply, and internal-model firms assess stress-test models against SS1/23 rather than SS3/18 alone.

What did PRA SS1/23 change for banks?

Before SS1/23 the PRA had no consolidated model-risk standard; UK banks often borrowed SR 11-7. SS1/23 made model risk a standalone discipline with named senior-manager accountability, and by covering all models regardless of technology, including vendor models, it became the main channel through which the PRA supervises AI in banks. The BoE/PRA told government in April 2026 they intend to build on it further in 2026 using supervisory insights on good practice.

What does PRA SS1/23 require, and how do its model risk management principles apply to AI and machine-learning models?

Supervisory Statement 1/23, 'Model risk management principles for banks', is the Prudential Regulation Authority's model risk management standard. Published with Policy Statement 6/23 on 17 May 2023, in force since 17 May 2024 and amended on 23 April 2026, it sets five principles: model identification and model risk classification; governance; model development, implementation and use; independent model validation; and model risk mitigants. Its formal scope is UK-incorporated banks, building societies and PRA-designated investment firms with internal-model permissions for credit, market or counterparty credit risk, and the PRA says other firms may find the principles useful. The model definition covers all models that inform business decisions regardless of technology, including vendor models and models used for financial reporting, and, without naming artificial intelligence or machine learning, its tiering principle (1.3(c)) weighs unstructured data, interpretability, explainability, transparency and bias for newly advanced approaches, while Principle 3.3(c) requires testing of material changes in dynamic models that adapt or recalibrate autonomously. Accountability for the whole MRM framework must be allocated to a named Senior Management Function holder.

SS1/23 is the UK's closest counterpart to the US interagency model risk guidance, but it differs in three ways that matter for AI. First, accountability is personal: the framework must be owned by a named SMF holder under the Senior Managers regime. Second, scope is deliberately technology-neutral — the definition captures any quantitative method that applies statistical, economic, financial or mathematical theories and techniques to process input data into an output used for business decisions, which brings machine-learning models, vendor-supplied scoring engines and generative tools that influence decisions inside the inventory. Third, AI sits inside the framework rather than beside it: the text never names AI or machine learning, but its tiering factors for newly advanced approaches (unstructured data, interpretability, explainability, bias) and its testing of dynamic models that change autonomously make these ordinary model risk questions, not a separate programme.

Compliance in practice runs on the model lifecycle. Identification and tiering (Principle 1) decide how much development documentation, validation depth and monitoring frequency each model gets; the PRA expects the tiering to consider materiality, complexity and, for AI, opacity. Independent validation (Principle 4) is where the October 2025 roundtable found the most inconsistency: several firms were applying core SS1/23 expectations unevenly where AI models introduce opacity, uncertainty and faster rates of change, and the PRA's slides set out its thinking on explainability, overfitting and ongoing monitoring as validation requirements. Model risk mitigants (Principle 5) — overlays, use restrictions, escalation — are what let a firm keep using a model whose limitations are known while remediation proceeds.

Two 2026 developments change the picture. The 23 April 2026 low-impact amendments (LIAF01/26) make explicit that SS1/23 is not a condition of internal-model approval and give newly permitted firms 12 months to comply, while folding stress-test model risk under SS1/23 for internal-model firms. And the Bank's February 2026 summary of its AI roundtables records that firms want SS1/23 applied, not replaced: the challenge they raised is that the traditional 'inputs to outputs' understanding at the heart of validation does not scale to generative AI and agentic systems, so risk management must move toward testing, monitoring and guardrails around outcomes. The PRA and Bank told Parliament in April 2026 that they intend to build on SS1/23 during 2026 using supervisory insights on good practice.

WHAT THIS MEANS IN PRACTICE

• SS1/23 compliance requirements start with the inventory: if a generative AI tool, a vendor fraud score or a spreadsheet influences a business decision, it is a model and needs a tier, an owner and a validation plan.
• SS1/23 model validation for AI/ML models must cover data provenance, explainability, overfitting and drift monitoring — the PRA's October 2025 slides treat these as ordinary validation questions at the right tier.
• Genai and agentic systems are in scope; the PRA's own summary says traditional validation will not be sustainable for them, so build outcome testing, monitoring and guardrails (Principle 5 mitigants) into the design rather than waiting for a validation exception.
• Name the SMF holder for model risk and give the audit committee the required report on MRM effectiveness for financial reporting — both are examinable evidence under Principle 2.
• Firms without internal-model permission are outside the formal scope; the PRA says they may find the principles useful and are welcome to consider them.
• New internal-model firms have 12 months from permission to comply (LIAF01/26); everyone else has been expected to comply since 17 May 2024, with self-assessment and remediation plans available to supervisors.

How does PRA SS1/23 compare?

• PRA SS1/23 vs SR 26-2 — UK and US Model Risk Compared

Which banks' AI programmes does PRA SS1/23 reach?

3 of the 120 largest US banks profiled on this site cite PRA SS1/23 among the documents their AI work answers to.