PRA SS1/23 AI Governance: UK Model Risk Management Supervisory Statement Applied to AI and Machine Learning

PRA Supervisory Statement SS1/23, published May 17, 2023 and effective from May 2024, is the toughest model risk management standard currently in force in any major banking jurisdiction. It was written with AI and ML explicitly in scope from day one — unlike the US Federal Reserve's SR 11-7 (2011), which predates the machine learning era and had AI governance bolted on informally by examiners over the following decade.

SS1/23 applies to all PRA-regulated firms: UK banks, building societies, PRA-designated investment firms, and — the part banks with Indian parents sometimes underweight — every UK subsidiary and branch of an international bank. HDFC Bank UK, ICICI Bank UK, SBI UK, Bank of Baroda UK, Axis Bank UK, and Yes Bank UK all answer to SS1/23 for every model running in their UK books, regardless of who built the model or where. A model built and validated in Mumbai does not arrive in London pre-cleared.

The PRA's 2023 Dear CEO letter on AI reiterated that existing requirements, SS1/23 included, already apply to AI systems — there is no AI carve-out waiting in the wings the way there was with the Fed's eventual SR 26-2. The Bank of England and FCA's joint Discussion Paper DP4/22 (October 2022) flagged where AI creates genuinely novel governance challenges, and signaled more guidance was coming. Until it arrives, SS1/23 is the operative standard, and PRA examiners are already using it to test AI governance in practice.

The Five SS1/23 Principles, Applied to AI

Model Identification and Risk Classification

Every model needs registration and a risk tier, and for AI the classification logic has to account for role in consequential decisions, opacity, susceptibility to distribution shift, and degree of autonomy. An AI system used in customer credit decisions that gets waved through as low-risk with no documented justification is exactly the kind of thing a PRA examiner flags on sight. Generative AI makes this harder, not easier: a model drafting regulatory submissions is a different risk animal from one summarizing documents, and firms need classification logic that tells the two apart rather than a single blanket "GenAI" tier.

Governance and Ownership

Every model needs a named owner accountable for keeping it inside its design scope — and for AI used in consequential decisions, that owner needs real seniority, typically director level or above. This is where traditional model risk management and AI governance collide: the credit risk team obviously owns the credit scoring model, but who owns a model a data science team built and a different business line now relies on? SS1/23 forces that ambiguity to get resolved, and under SMCR, material AI model risk has to trace to a named Senior Manager whose Statement of Responsibilities actually says so — not implies it.

Model Development, Implementation, and Use

Documentation has to cover data sourcing, feature engineering, training methodology, and hyperparameter choices to a standard an independent reviewer can actually work from. The architecture choice — gradient boosting versus neural network versus transformer — needs a justification, not just a record that a choice was made. "We used this foundation model because it performed best on our evaluation set" does not clear the bar: the PRA expects the evaluation methodology, the dataset's representativeness, and the alternatives considered to all be written down, which is a materially higher bar than most banks currently clear for vendor GenAI deployments.

SS1/23's use principle also covers the practical reality of model drift: a model validated on 2024 data can behave differently in 2026 production if the data environment moved underneath it. Implementation controls need to catch AI being used outside its validated scope before an examiner does — a retail credit model should not quietly start making commercial lending calls without revalidation. For autonomous agents, this principle requires the boundary between autonomous action and human referral to be defined and enforced in the agent's actual design, not just described in a policy document nobody reads.

Independent Model Validation

Validators need genuine independence — no reporting line to model owners, full access to documentation and data, and real authority to challenge development decisions. For AI, validation scope has to cover conceptual soundness, data quality, out-of-sample performance, stability across subpopulations and over time, and bias/fairness testing for models feeding consequential decisions. For generative AI specifically, validation must address hallucination rate, factuality, and output-constraint compliance — and the PRA explicitly expects firms to build AI-specific validation methods rather than bolt traditional statistical validation onto systems it was never designed to evaluate.

Model Risk Mitigants: Monitoring, Limits, and Reporting

The fifth principle covers the ongoing controls that keep a validated model safe in production — performance monitoring, documented limits of use, and risk reporting to senior management and the board. For AI, monitoring needs to track output quality, behavioral consistency against validated behavior, and adversarial resilience; for generative AI specifically, hallucination rate and output-constraint compliance belong in that monitoring. Board-level model risk reporting has to include AI incidents and emerging risks alongside quantitative validation status — a report that only covers traditional models is an incomplete report once the bank has deployed any AI at scale.

SS1/23 is a supervisory statement, which in PRA language means compliance is expected, not aspirational. Banks running AI without SS1/23-aligned governance aren't lagging best practice — they are carrying an open supervisory gap the PRA can and does cite in its annual governance assessment. We've watched model risk assessments shift from "do you have a validation function" five years ago to "show me the AI-specific validation findings" today, and that shift isn't slowing down.

SS1/23 vs. SR 11-7/SR 26-2: What International Banks Need to Know

SR 26-2 replaced SR 11-7 outright on April 17, 2026 — the first rewrite of US model risk guidance in fifteen years — and its headline decision was to exclude generative and agentic AI from formal scope on the grounds that the technology is moving too fast for a settled rulebook. That carve-out does not mean US examiners have stopped asking about GenAI governance; if anything, the absence of a prescriptive rule has made examiners lean harder on a bank's own risk framework to see what it covers voluntarily. A UK subsidiary of a US bank inherits the stricter of the two regimes by default, since SS1/23 draws no such exclusion.

SMCR Alert: if your CRO's (or equivalent Senior Manager's) Statement of Responsibilities doesn't explicitly reference AI model governance, that is a gap the PRA's model risk assessment will surface — not a drafting nicety you can clean up later.

India Banks With UK Operations: Dual-Jurisdiction Compliance

Every UK subsidiary of an Indian bank is independently PRA-regulated and independently on the hook for SS1/23 — the parent's RBI-compliant governance back home does not transfer. The UK entity needs its own inventory, its own tiering, its own validation function, and its own SMCR-compliant accountability chain. For models shared between an India parent and its UK subsidiary — a fraud model trained partly on UK data, a group credit scoring engine — the governance structure has to name who owns the UK use, who validates for the UK's regulatory purposes, and how the UK entity maintains SS1/23-grade documentation for a model the parent primarily built.

Structurally, RBI's direction (inventory, tiering, validation, monitoring) and SS1/23 point the same way, but SS1/23 is the more detailed framework. For India banks with UK operations, implementing SS1/23 first and checking RBI alignment against that baseline is more efficient than building two separate governance programs from scratch.

SS1/23's development principle requires training data to be documented for quality and representativeness — and that documentation obligation collides directly with cross-border data rules. Models developed in India and deployed in the UK have to satisfy both UK GDPR (post-Brexit) and India's DPDP Act 2023, whose Rules were notified on November 14, 2025 with a hard compliance deadline of May 13, 2027. Personal data used to train AI models cannot move from UK operations back to India for retraining without a proper transfer mechanism in place — a requirement that pushes India-UK banking groups toward data governance for AI training pipelines that most haven't formalized yet.

What to Do About It

• Rewrite your model risk tiering criteria to explicitly name opacity, distribution shift, and degree of autonomy as factors — "we classify by materiality" is no longer a defensible answer on its own.
• Confirm your CRO's (or relevant SMF holder's) Statement of Responsibilities names AI model governance explicitly; if it doesn't, fix it before your next PRA engagement, not after.
• Build a documented AI-specific validation methodology separate from your statistical model validation playbook — hallucination rate and factuality testing need their own standard.
• For any UK subsidiary running a group-developed model, confirm it has its own named owner, its own UK-context validation, and its own SS1/23 documentation trail.
• Map your India-UK AI training data flows against both UK GDPR and DPDP Act 2023 transfer rules before the May 2027 DPDP deadline, not in the quarter before it.

Zytra's governance platform produces the audit trail SS1/23 requires: model inventory with risk tiering documentation, validation scope records, performance monitoring evidence, and board-ready model risk reports. Used by UK-regulated entities and international banks managing dual-jurisdiction model governance obligations.

Talk to sales