The PRA judges AI agents by whether they threaten a bank's safety and soundness. It has no AI rulebook; it applies existing expectations. For agents, four matter most: model risk management under SS1/23, operational resilience of the services agents support, management of the third parties behind them, and a named senior manager answerable for each.
What does the PRA require, in plain words?
The Prudential Regulation Authority, part of the Bank of England, supervises banks, building societies, credit unions, insurers and major investment firms for safety and soundness. The UK takes a principles-based, regulator-led approach to AI rather than passing a single AI law, so the PRA applies the expectations it already has:
- Model risk management. SS1/23 sets principles for banks: identify and classify models, govern them, develop and use them soundly, validate them independently, and apply mitigants where model risk is high.
- Operational resilience. Firms must identify their important business services, set tolerances for how much disruption each can bear, map what they depend on and test that they can stay within tolerance in severe but plausible scenarios.
- Outsourcing and third parties. Firms remain responsible for services they rely on others to provide, and must manage the risks through the life of the arrangement.
- Accountability. Under the Senior Managers and Certification Regime, named individuals are accountable for areas of the business, including model risk and operational resilience.
Who does it apply to?
PRA-authorised firms. SS1/23 applies primarily to banks and building societies with permission to use internal models for regulatory capital, though other firms may find its principles useful. Resilience, outsourcing and accountability expectations apply more widely. The FCA supervises the same firms for conduct, and the ICO for data protection; our piece on ICO expectations for agents processing personal data covers the latter.
What does the prudential lens mean for agents that act?
The PRA's concern is whether a firm can absorb errors without failing or harming the system. Agents raise four prudential questions.
Is the agent a model? Usually yes, in the sense that matters: it turns inputs into outputs used in decisions. The question is how to classify and tier it. Our short answer on whether SS1/23 covers AI agents works through it.
Does an important business service depend on it? If an agent handles payment exceptions or customer authentication, its failure could take a service outside tolerance. Agents must appear in resilience mapping, and scenario tests should include the agent and its model provider failing or behaving wrongly.
Is the bank dependent on a few providers? Many agents may run on one model provider. That is a concentration the PRA will want to see understood, with exit and substitution plans.
Who answers for it? The PRA expects a senior manager to own model risk management. For agents, that individual needs to know which agents act without a person and on what evidence. A responsibilities map that names an owner for model risk but cannot list the agents that act on their own is a gap a supervisor will find quickly.
Take an illustrative case. A UK building society uses agents to process mortgage payment holiday requests and to triage arrears contact. Its resilience mapping shows arrears handling supports an important business service. The society records both agents in its model inventory, validates them on its own cases, sets a scenario test where the model provider is unavailable for a day, and confirms that staff can take over within tolerance. The senior manager responsible for model risk signs off each agent's authority.
How do you meet PRA expectations for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Model identification and tiering | Agents in the model inventory, tiered by impact and what they can do | Inventory entries with tier |
| Model governance | Senior manager accountable; agent authority approved at the right level | Responsibilities map; approvals |
| Sound development and use | Brief, sources, tools and limits documented; limits enforced before action | Design record; verdict log |
| Independent validation | Validation on the firm's own cases by a separate team | Validation reports |
| Model risk mitigants | Lower authority, held actions or extra review for high-tier agents | Mitigant register |
| Operational resilience mapping | Agents and model providers in maps of important business services | Resilience maps |
| Scenario testing | Tests of agent failure and wrong behaviour within impact tolerance | Scenario test results |
| Third-party risk | Agent platform and model provider managed through the relationship; exit plans | Third-party file; exit plan |
What will a PRA supervisor ask?
- Which agents support important business services?
- How are agents tiered in your model inventory, and who validated them?
- What happens to those services if your model provider fails or changes its model?
- Which senior manager is accountable, and what reports do they receive?
- How concentrated are your agents on a single provider?
What does the PRA not cover?
It does not supervise conduct towards customers; the FCA does, including under the Consumer Duty. It does not enforce data protection. And its approach differs from that of other supervisors; a US national bank would face the OCC's lens, described in OCC expectations for AI agents in national banks. For banks operating in both the UK and EU, see how AI rules differ between the UK and EU. For the full map, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does the PRA require banks to notify it before deploying AI agents?
There is no general notification requirement for using agents. Firms should consider whether a deployment is material enough to raise with their supervisor, for example where it affects an important business service or involves a significant outsourcing.
Do insurers face the same PRA expectations as banks?
They face the PRA's resilience, outsourcing and accountability expectations. SS1/23 is aimed at banks, but insurers using agents in pricing or reserving will meet model risk scrutiny through other routes.
Is a model provider a critical third party?
The UK has a regime for third parties critical to the financial sector, with designation by the government. Whether or not a provider is designated, firms remain responsible for managing their own reliance on it.