Reference · Model risk management

PRA SS1/23 Explained: Model Risk Management Principles for AI and LLMs

Model risk management principles for banks, current version effective 23 April 2026

Last reviewed October 15, 2026

At a glance

First published: 17 May 2023, with policy statement PS6/23
Issued by: Bank of England, Prudential Regulation Authority
In effect since: 17 May 2024
Current version: Dated 16 April 2026, published and effective 23 April 2026 (LIAF01/26)
Applies to: UK-incorporated banks, building societies and PRA-designated investment firms with internal model approval
Artificial intelligence: Not named; reached through a technology-neutral model definition
Official text: SS1/23, April 2026 version (PDF)

SS1/23 is the Prudential Regulation Authority's supervisory statement on model risk management for banks. The PRA published it on 17 May 2023 with policy statement PS6/23, and it took effect on 17 May 2024.PS6/23 The current version is dated 16 April 2026 and was published, and took effect, on 23 April 2026 under the PRA's low impact amendments round LIAF01/26.PRA SS1/23 page

The statement treats model risk as a risk in its own right and organizes its expectations under five principles that cover the whole model lifecycle.SS1/23 · 1.1, 1.4 Its text never uses the words artificial intelligence or machine learning. It reaches a large language model (LLM) through its model definition, which the PRA applies to every model used to inform business decisions, regardless of technology and including vendor models.PRA SS1/23 page

SS1/23 binds firms with internal model approval

The statement is relevant to UK-incorporated banks, building societies and PRA-designated investment firms with internal model approval to calculate regulatory capital: Internal Ratings Based approaches for credit risk, the Internal Model Approach for market risk, or the Internal Model Method for counterparty credit risk. Through the rest of the text, these are the "firms".SS1/23 · 1.2, 1.3, fn. 2

Its expectations do not apply to firms without permission to use internal models for regulatory capital, or to third-country firms operating in the UK through a branch. Credit unions, insurers and reinsurers are out of scope. The PRA adds that firms outside the scope may find the principles useful and are welcome to consider them.SS1/23 · 1.2

Firms in scope self-assess against the principles at least annually, share the findings and any remediation plans with the board, and provide them to the PRA on request. The senior manager (SMF) accountable for model risk management makes sure remediation plans exist and have owners.SS1/23 · 1.7, 1.8 The expectations scale with a firm's size, business and model use, and within a firm with each model's tier.SS1/23 · 3.3, 3.4

An LLM meets the model definition

Principle 1.1(a) asks firms to adopt the PRA's definition of a model: a quantitative method, system or approach that applies statistical, economic, financial or mathematical theories, techniques and assumptions to process input data into output. Inputs can be quantitative, qualitative or based on expert judgement, and output can be quantitative or qualitative.SS1/23 · Principle 1.1(a) An LLM applies statistical techniques to input text and produces text, and qualitative output is expressly included.

Use is what brings a model into the framework. The statement defines model use as relying on a model's output to inform business decisions, and reads business decisions broadly, from general business and operational banking activities to risk, capital and reporting.SS1/23 · 2.1, fn. 4 An LLM that summarizes a credit file for an underwriter, or sorts complaints for review, informs such a decision.

The statement covers all models, developed in-house or externally, including vendor models.SS1/23 · 1.3 Principle 2.6 keeps the board and senior management ultimately responsible for model risk under an outsourcing or third-party arrangement. It asks firms to satisfy themselves that a vendor model was validated to their own standards, to verify the relevance of vendor-supplied data and assumptions, and to validate their own use of the product with ongoing monitoring and outcomes analysis "using their own outcomes."SS1/23 · Principle 2.6 Documentation of a vendor model has to be detailed enough to validate the firm's use of it.SS1/23 · Principle 3.5(a)

The PRA reads the statement the same way. Slides from its October 2025 roundtables with chief risk officers describe SS1/23 as setting expectations for all model types, including models that use AI and machine learning.PRA roundtable slides · slide 3

Five principles, one inventory

Principle 1, model identification and risk classification, asks for a firm-wide inventory of every model in use, under development or decommissioned, recording each model's intended use against its actual use, its assumptions and limitations, and its validation findings.SS1/23 · Principle 1.2 Each model gets a materiality and complexity rating, and for newer techniques the complexity assessment may weigh unstructured data, interpretability, explainability, transparency and the potential for bias.SS1/23 · Principle 1.3 An LLM shared by many teams has one entry and many uses, so comparing intended and actual use depends on a record of who called it and for what.

Principle 2, governance, has the board set a model risk appetite with measures for limits on model use, exceptions and tolerance for errors, and makes a named senior manager accountable for the framework.SS1/23 · Principles 2.1(c), 2.2 Policies stay current with the new and advancing technologies the firm is exposed to.SS1/23 · Principle 2.3(a) For LLMs and agents, the appetite statement is where a firm records which uses it accepts before the first one goes live.

Principle 3, development, implementation and use, sets standards for design, data, testing and documentation. Material model changes, including changes that build up over time in models that adapt on their own, call for development testing, and models run in systems under rigorous quality and change control.SS1/23 · Principles 3.3(c), 3.6(a) For an LLM, those systems hold the prompts, tool permissions and settings that shape its behavior, and a provider's new model version is a change the firm's process has to classify.

Principle 4, independent model validation, asks for a validation function independent of development, with separate reporting lines at firms with internal model approval, and for performance monitoring at a frequency set by the model's tier.SS1/23 · Principles 4.1(d), 4.4(d) Monitoring includes checking whether a model has been used beyond its intended use and whether that use delivered acceptable results.SS1/23 · Principle 4.4(b)(iv) For a model the firm did not build, validation works from the firm's own use: the vendor's documentation, the inputs and outputs the firm observes, and its own outcomes.

Principle 5, model risk mitigants, covers post-model adjustments, restrictions on use up to prohibiting a model for specific purposes, and approved exception and escalation procedures for material models.SS1/23 · Principles 5.1 to 5.3 An exception includes using a model that is not approved or validated for that use, or using one outside its intended purpose, and escalation should make key stakeholders promptly aware of it.SS1/23 · Principle 5.3, fn. 23 An LLM open to many teams turns an unapproved use into a single API call, so the exception record depends on seeing each call.

The April 2026 update clarified paragraph 1.9

LIAF01/26 amended paragraph 1.9. The expectations in SS1/23 are not conditions for internal model approval, and a firm that first receives permission to use an internal model for regulatory capital has 12 months from that grant to meet them.LIAF01/26 SS1/23 · 1.9 The May 2023 text gave the 12 months only to firms first approved after the statement's publication and said nothing about approval conditions.SS1/23 (May 2023) · 1.9

The PRA describes the amendments as clarifying its existing policy.LIAF01/26 Set side by side, the May 2023 and April 2026 texts differ in paragraph 1.9 and in no principle or sub-principle.

Before and after

Paragraph 1.9 before and after LIAF01/26

May 2023 textApril 2026 text

Approval condition: Not addressed; →; Not a condition of internal model approval
12 months to comply: Firms first approved after May 2023; →; Any firm on first approval

From paragraph 1.9 of the May 2023 and April 2026 texts of SS1/23, as amended by LIAF01/26.

SS1/23 and SR 26-2 side by side

The US agencies issued SR 26-2, the Revised Guidance on Model Risk Management, on 17 April 2026, six days before the PRA's update. Both keep validation, ongoing monitoring and vendor models inside model risk management, and they differ in who they reach, in what counts as a model, and in generative AI.Revised Guidance SR 26-2 places generative and agentic AI models outside its scope by footnote.Revised Guidance · footnote 3 SS1/23 has no such exclusion, and its definition covers qualitative output.SS1/23 · Principle 1.1(a)

A UK subsidiary may rely on its parent group's validation of a group model only when it can show that the group's framework and standards meet SS1/23, that the data and assumptions fit its own use, and that the validation is rigorous enough for the model's tier at the subsidiary.SS1/23 · Principle 2.6(c) A US group whose UK bank holds internal model approval therefore works to both documents: SS1/23 at the UK bank, and SR 26-2 for the US banking organization.

Topic	PRA SS1/23	SR 26-2
Issued by	Prudential Regulation Authority	Federal Reserve, OCC and FDIC, jointly
Who it reaches	UK banks, building societies and PRA-designated investment firms with internal model approval	Most relevant to banking organizations with over $30 billion in total assets
Legal effect	PRA expectations, with an annual self-assessment reported to the board	No enforceable standards or prescriptive requirements
What counts as a model	Quantitative method applying statistical, economic, financial or mathematical techniques; qualitative output included	Complex quantitative method applying statistical, economic or financial theory to produce quantitative estimates
Generative and agentic AI	Not named and not excluded	Out of scope (footnote 3)
Vendor models	Validated to the firm's own standards; the firm validates its own use against its own outcomes (Principle 2.6)	Principles apply without the vendor's code, data or methodology; validate and monitor outcomes (Section VII)
Tiering	Materiality and complexity rating for every model (Principle 1.3)	Oversight set by model purpose and exposure (Section III)
Validation timing	Monitoring and revalidation frequency set by the model's tier (Principles 4.4, 4.5)	Varies with purpose, methodology, model change and data limits (Section V)
Accountability	A named senior manager and a board-set model risk appetite (Principle 2)	Defined roles, with internal audit evaluating the framework (Section VI)

Source: SS1/23 (April 2026 text) and the interagency Revised Guidance on Model Risk Management attached to SR 26-2, status as of 15 October 2026. SS1/23, April 2026 (PDF)

The PRA's 2025 roundtables applied the principles to AI and machine learning

On 20 and 22 October 2025 the PRA held two roundtables with chief risk officers from 21 regulated firms on adopting AI and machine learning under SS1/23, and it published the slides on 24 November 2025.PRA roundtable notice Each observation in them is framed against a principle already in the statement, such as Principle 2.1(c) on risk appetite and Principle 3.3(c) on model changes.PRA roundtable slides · slides 5, 11

The slides report tiering policies that set a minimum complexity for AI and ML models while the same firms' inventories rated those models low, and note that one technique deployed across many portfolios can carry higher aggregate materiality than each use suggests.PRA roundtable slides · slide 6 They suggest that six-month monitoring may be too infrequent for a dynamic AI or ML model, and that firms consider pre-approved fallback or challenger models and, in extremis, kill switches.PRA roundtable slides · slides 12, 13

Control mapping

What a reviewer expects to be able to see.

Obligation	What the system must do	Evidence a reviewer expects
Model definition	Adopt the PRA's definition and apply it to every model that informs business decisions, vendor models included	A model policy that places each LLM use in or out of scope, with the reason
Model inventory	Record purpose, intended against actual use, limitations, validation findings and owners for each model	Inventory entries for each LLM and provider, with the applications and teams that call it
Tiering	Rate each model's materiality and complexity, weighing interpretability and bias where relevant	Tier rationale per model, reassessed at validation and revalidation
Model risk appetite	Set board measures for limits on model use, exceptions and performance tolerance	The approved appetite statement and board reports against it
Vendor models	Validate the firm's own use and monitor vendor model performance against its own outcomes	Vendor documentation, call-level usage records and the outcomes analysis built on them
Change control	Run models in systems under quality and change control, and test material changes	Change records for model versions, prompts, tool permissions and settings
Restrictions on use	Limit or prohibit use for specific purposes when deficiencies are found	Restrictions in force, with when and why each was set
Exceptions and escalation	Define exceptions for material models and make stakeholders promptly aware of each one	An exception log with the escalation path and who was told
Self-assessment	Self-assess at least annually and share findings and remediation plans with the board	A dated self-assessment, remediation plan and board record

Key dates

Timeline, drawn to scale

SS1/23 from publication to the April 2026 update

Published or in effectCurrent version

Dates from the PRA's SS1/23 page and its roundtable notice. Status as of October 15, 2026.

17 May 2023The PRA publishes SS1/23 with policy statement PS6/23, its response to consultation paper CP6/22.
17 May 2024SS1/23 takes effect. Firms complete an initial self-assessment before this date and update it at least annually after it.
24 November 2025The PRA publishes the slides from its 20 and 22 October 2025 roundtables on AI and machine learning under SS1/23.
23 April 2026The revised SS1/23, dated 16 April 2026, is published and takes effect under LIAF01/26, clarifying paragraph 1.9.

Sources cited

Common gaps

Where a self-assessment against SS1/23 most often finds an expectation only partly met for LLMs.

LLMs missing from the inventory. Any use that informs a business decision is model use under the statement. An LLM reached through a product feature or a developer's API key still needs an entry, an owner and a tier.
Intended use with no record of actual use. Principle 1.2 asks for the comparison, and Principle 4.4 asks whether use beyond the intended purpose delivered acceptable results. Both depend on a record of actual use.
Vendor validation that ends with the vendor's documents. A provider's documentation informs the first two asks of Principle 2.6. The third, validating the firm's own use against its own outcomes, rests on the firm's own records.
A tiering policy and an inventory that disagree. The PRA's 2025 roundtable slides reported AI and ML models rated low complexity in inventories whose policies set a higher minimum. Reconciling the two belongs before the next self-assessment.
Provider version changes outside change control. A provider releases new model versions on its own schedule. The firm's change process decides which releases are material and records each decision.
Monitoring on a calendar set for traditional models. The roundtable slides suggest six-month intervals may be too infrequent for a dynamic AI or ML model. Monitoring frequency follows the tier, and the tier has to reflect the model.

In practice

How Meilynx maps to SS1/23 →

Last reviewed October 15, 2026. This reference summarises publicly available regulatory guidance and is provided for general information. It is not legal advice. Obligations depend on an institution's charter, registration status, size, and activities. Verify against the primary sources cited above and consult counsel before relying on any summary here.