---
title: What does PRA SS1/23 mean for AI agents in UK banks?
url: https://opencontroller.lyzr.com/blog/pra-ss1-23-model-risk-and-ai-agents
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What does PRA SS1/23 mean for AI agents in UK banks?

**PRA SS1/23 applies to all models used to inform business decisions, regardless of technology and including vendor models. An AI agent is built on a model and acts on its outputs, so its model belongs in the model inventory and the agent itself should be tiered, validated and monitored. The five principles give a ready structure for agent evidence: identification, governance, development and use, independent validation, and mitigants.**

This is general information, not regulatory advice.

## What is PRA SS1/23?

SS1/23, "Model risk management principles for banks", is a Prudential Regulation Authority supervisory statement. It was published on 17 May 2023 and took effect on 17 May 2024. The PRA published an updated version in April 2026, so check the current text on the Bank of England site.

Its expectations are aimed at UK-incorporated banks, building societies and PRA-designated investment firms with approval to use internal models for regulatory capital. Many other firms use it as their benchmark for model risk.

The scope is broad. The principles cover all types of models that inform business decisions, built in-house or bought from vendors, and models used for financial reporting. A sub-principle addresses risks from AI and machine learning techniques where they apply to models generally.

## Is an AI agent a "model" under SS1/23?

The large language model inside an agent is a model: a quantitative method that turns inputs into outputs. Usually it is a vendor model, which SS1/23 explicitly covers.

The agent adds more: prompts, tools, retrieval and the decision to act. Treat the agent as a model use, or a system of models, rather than arguing it is out of scope. Three practical reasons:

1. The agent's output informs or makes business decisions, which is the trigger SS1/23 uses.
2. Changing the prompt or tool list changes behaviour as much as changing model weights.
3. A supervisor will look at the outcome, not at where the model boundary sits.

## How do the five principles apply to AI agents?

| SS1/23 principle | What it expects | How it applies to agents | Evidence to keep |
|---|---|---|---|
| 1. Model identification and risk classification | A complete model inventory and a tiering approach | Register every agent and the models it uses, including agents found outside approved platforms | Agent registry with model, version, owner and tier |
| 2. Governance | Board oversight and a senior manager accountable for the framework | Agent risk sits under the same accountable executive and policies | Policy, committee minutes, owner per agent |
| 3. Model development, implementation and use | Documented design, data, testing and use limits | Documented purpose, tools, permissions and prompts, versioned | Agent card per version, test results |
| 4. Independent model validation | Validation independent of developers, scaled to tier | Independent evals and red-team testing before release and after material changes | Validation reports tied to version |
| 5. Model risk mitigants | Restrictions on use, adjustments and controls for weaknesses | Guardrails, tool allowlists, spend limits, quarantine and kill | Runtime control config and incident logs |

## What changes for validation when the model is a vendor LLM?

You cannot validate the vendor's training data or weights. SS1/23 still expects you to understand and test vendor models within your use. For agents, that means:

- **Test the use, not the model.** Validate the agent on your tasks, with your data and tools, across multi-turn conversations.
- **Treat model version changes as material.** A provider swapping a model alias is a change to your model. Pin versions where you can.
- **Monitor in production.** Validation at release is a snapshot. Score live traffic for groundedness and task adherence and alert on drift.
- **Keep a fallback.** If a vendor model is withdrawn or degrades, a tested alternative is a mitigant under Principle 5.

## How should AI agents be tiered?

SS1/23 tiers by materiality and complexity. For agents, add autonomy:

1. **Materiality:** what decisions or money the agent touches, and how many customers.
2. **Complexity:** number of models, tools and agent-to-agent hops.
3. **Autonomy:** whether a human reviews every action, samples them, or none.

A drafting assistant with human review may be low tier. An agent that moves payments or changes credit limits without review is high tier and needs the deepest validation. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents) and our broader [AI agent controls for bank model risk](/blog/ai-agent-controls-for-banks-model-risk). Canadian banks face a similar exercise under [OSFI E-23](/blog/osfi-e-23-model-risk-and-ai-agents).

To connect this to other frameworks, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps a governed registry of agents, versions, tools and MCP servers, which supports Principle 1 inventory. Multi-turn and red-team evals and LLM judges on production traffic support validation and monitoring. Rogue-agent handling, per-agent Block and Quarantine, and LLM Gateway fallback act as mitigants. Policy as code maps controls to NIST and ISO 42001. This is evidence for your framework, not a regulatory sign-off. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does SS1/23 cover generative AI?**
Yes. It covers all models that inform business decisions regardless of technology, and includes a sub-principle on AI and machine learning risks.

**Are vendor LLMs in scope of SS1/23?**
Yes. The principles cover externally developed models. You validate the model within your own use.

**When did SS1/23 take effect?**
It was published on 17 May 2023 and took effect on 17 May 2024. An updated version was published in April 2026.

## Key facts

PRA SS1/23 Model risk management principles for banks: published 17 May 2023, effective 17 May 2024; updated version published April 2026. Aimed at UK banks, building societies and PRA-designated investment firms with internal model approval. Covers all models that inform business decisions, including vendor models, regardless of technology; includes a sub-principle on AI/ML risks. Five principles: model identification and risk classification; governance; development, implementation and use; independent validation; model risk mitigants. AI agents and their LLMs should be inventoried, tiered (materiality, complexity, autonomy), validated and monitored. Opencontroller provides registry, evals, LLM judges, quarantine/kill and policy mappings as evidence. Pricing starts from $250K per deployment per year.

## Related

- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [What does OSFI Guideline E-23 mean for AI agents at Canadian banks?](https://opencontroller.lyzr.com/blog/osfi-e-23-model-risk-and-ai-agents.md)
- [What does the UK FCA expect from firms using AI agents?](https://opencontroller.lyzr.com/blog/uk-fca-expectations-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
