AI
Generative AI in banking: use cases for risk teams
Generative AI is crossing from proof of concept into production across UK banking. For risk teams, the question is no longer whether to engage with it but which use cases will hold up under governance scrutiny.
A model validator approaching a quarterly documentation review has come to expect a particular kind of friction. The pack runs to several hundred pages; the regulatory text has been updated twice since the last cycle; the stress scenarios reference macro assumptions reconciled only days earlier. None of this is poor process. It is the cumulative weight of documented, auditable risk management. The result is that experienced professionals spend a disproportionate share of their time on information assembly rather than on the judgement those documents are meant to contain.
Generative AI is being positioned as a solution to precisely this friction. In practice, the use cases that produce genuine value in a risk context are narrower than the capability claims suggest, and even those that hold up carry governance obligations the supervisory frameworks only partially address.
How far has AI adoption outrun understanding?
The scale of adoption is no longer in question. The Bank of England and Financial Conduct Authority’s 2024 survey of 118 UK-regulated firms found 75% already using AI, with a further 10% planning to, and foundation models already account for 17% of use cases1. Governance scaffolding has followed: 84% of firms have a designated accountable person for AI. We unpack this survey, and what it means for a risk function’s AI maturity, in our companion piece on AI maturity assessment2.
The more revealing figure in the same survey is 34%: the proportion of respondents reporting a complete understanding of the AI they deploy.1 For the remaining two-thirds of the sector, AI applications run in production while the firms themselves can’t fully characterise the systems’ behaviour. The understanding gap is concentrated in third-party models, where training data and methodology aren’t open to independent inspection.
With the frontier models from OpenAI, Anthropic, and their peers, the gap runs deeper than access: even with the training data in hand, systems of that scale can’t be comprehensively understood. That disconnect will hold for the foreseeable future, and regulation will need to recognise it and work with both sides of it rather than assume it away. Broad adoption, incomplete understanding, and no quick fix: this is the context in which risk teams are now making decisions about generative AI.
The problem isn’t AI adoption. It’s the distance between what firms have deployed and what they can account for.
Three use cases hold up: model development documentation, stress-testing intelligence, and regulatory search. Each solves a different bottleneck, and each carries its own validation demand.
What can generative AI do in an IRB model change?
Start with where the hours actually go. An IRB (internal ratings-based) model change generates hundreds of pages: data quality reports, methodology write-ups, parameter narratives, submission packs. Drafting them consumes qualified modeller time without requiring the analytical judgement that justifies the cost. That is the bottleneck generative AI can credibly relieve, and it is assembly work, not judgement work.
Agentic AI systems chain multiple large language model calls with tool use, memory, and sequential logic. A system that can draft documentation, check it against a template, and route it for human review handles the procedural load without touching the analytical substance. The key distinction: these systems surface work for human decision, they don’t complete it autonomously.
How can generative AI keep stress-test scenarios current?
Management stress testing is the second domain where generative AI has proved its value, and the bottleneck is different. Here the problem isn’t drafting volume but staying current. Designing a scenario means turning a picture of the world that changes weekly, macro data, geopolitical events, policy announcements, into structured quantitative inputs, on a fixed cycle and with modelling teams waiting downstream.
Retrieval-augmented generation (RAG) is built to keep those inputs current as the picture changes. Instead of answering from its training data, which may be months old, a RAG system looks up the current version of the relevant document at the moment a question is asked, and answers from that. When a new policy statement lands, the knowledge base absorbs it the same day. Nothing needs retraining.
The validation obligation for a RAG system shifts accordingly. What needs verification is not whether the generated text is coherent but whether the retrieved document is the right one and the knowledge base is current. Retrieval precision and document currency are the properties to specify before deployment.
How can risk teams search regulatory text faster?
A third use case, less visible in vendor literature but practically significant, is regulatory intelligence: querying a corpus of supervisory statements, consultation papers, and internal methodology documentation in natural language and receiving a source-cited answer.
Risk teams maintain large and frequently updated libraries of regulatory guidance. The operational problem is not access but search: locating a specific paragraph across successive consultation papers, tracing how a rule has evolved, or identifying which internal models are affected by new guidance. These tasks currently rely on systematic manual search, or on the institutional memory of whoever holds it: knowledge that leaves when the person does.
The machinery is the same retrieval-augmented setup stress testing uses: the model answers only from documents fetched for the question rather than from memory. A retrieval failure is more dangerous here, and more silent, because a plausible answer assembled from the wrong paragraph looks exactly like a right one. When the system retrieves an outdated or mismatched document, it still writes a fluent, source-cited answer that reads as authoritative and is wrong. In a compliance setting, that answer can reach a decision before anyone checks the source.
That’s why regulatory search adds a third requirement to stress testing’s two. Retrieval precision and document currency remain the baseline. Compliance adds boundary detection: the system should say when it can’t find enough context to answer reliably, rather than answer confidently regardless. No current model does that dependably, so the working control is human review: an answer is checked against its cited source before it informs a decision. The system should be set up to make that check quick: every answer carrying a citation precise enough to take the reviewer straight to the source passage.
Where the governance frameworks fall short
Credible use cases exist. The governance question is whether the frameworks around them are adequate, and the answer is incomplete. The UK’s model risk guidance is the Prudential Regulation Authority’s Supervisory Statement SS1/23, which binds UK banks, building societies and PRA-designated investment firms holding internal model approval, and not credit unions, insurers or third-country branches. It says nothing about generative AI at all: the words appear nowhere in it.3 That is silence rather than a carve-out, and it leaves the question open rather than settling it. It isn’t a signal that governance obligations are reduced. Our companion piece covers SS1/23 in full.4
If anything, it marks how far the rules still have to travel: regulation will need to catch up with how these models are actually used, and getting there will take greater understanding on all sides, supervisors, firms, and the model providers between them.
The properties that make these tools operationally attractive are precisely those that complicate validation. A risk team’s validation toolkit was built for models like the PD (probability of default) model: fixed inputs, a deterministic output function, behaviour reproducible across runs. The tools in this article share none of those properties. An agentic system’s behaviour emerges from the interaction of the language model, tool responses, and accumulated context state, and that interaction isn’t fully enumerable in advance. Two identical prompts can produce different tool-call sequences and different outputs. Pre-deployment testing gives weaker guarantees than it does for bounded statistical models, because the input space is effectively open-ended.
This is the Bank of England and FCA survey’s understanding gap made concrete. An agentic system in production, without added instrumentation, is exactly the kind of AI a firm deploys but can’t fully characterise.
The practical governance response involves three structural requirements.
- Continuous monitoring: Treat each inference cycle as a quality data point, not a one-time pre-deployment test but ongoing assurance across every production run.
- Structured logging: Record inputs, intermediate reasoning steps, tool calls, and outputs so that every decision can be audited after the fact.
- Alignment monitoring: Catch drift triggered by environment changes, a revised policy document, an updated data schema, a new retrieval corpus, not only by changes to the model itself.
Pull the three use cases together and the demands on a risk team converge: know the system is retrieving the right document and a current one, log what it does so every run can be audited, and keep a human decision at each point that matters. None of that is exotic. It is model risk discipline, applied to a tool the existing rulebook wasn’t written for.
Return to the validator and the quarterly pack. What changes is where the hours go: less of the cycle spent assembling information, more of it on the judgement the documents are meant to contain. But the systems that free up that time don’t run themselves. Risk teams using generative AI in capital, stress testing or regulatory workflows need to extend their model risk governance to cover it, whether or not a supervisor has asked. What a supervisor’s silence does not do is limit what a function chooses to govern itself, and the industry frameworks filling the gap carry no binding force.5
The upside is real, whatever the hype around it. Governance still has to catch up, and the teams that build theirs now will be ready when it does.
Frequently asked questions
Which generative AI use cases hold up for a risk team?
Three survive governance scrutiny, and each solves a different bottleneck. Model development documentation relieves the drafting load around a model change. Stress-testing intelligence keeps scenario inputs current as the macro picture moves. Regulatory search makes a large library of supervisory material queryable in natural language. The value in a risk context is narrower than the capability claims suggest, and each of the three carries its own validation demand.
What does the Bank of England and FCA survey show about AI in UK firms?
The 2024 joint survey covered 118 UK-regulated firms and found 75% using AI, with foundation models, the category behind generative AI, accounting for 17% of reported use cases. Governance scaffolding has followed adoption: 84% of firms have a designated accountable person for AI. The more revealing figure is 34%, the proportion reporting a complete understanding of the AI they deploy. For the remaining two-thirds, AI runs in production while the firm cannot fully characterise its behaviour.
Why is the understanding gap hard to close?
It is concentrated in third-party models, where training data and methodology are not open to independent inspection. With frontier models the gap runs deeper than access: even with the training data in hand, systems at that scale cannot be comprehensively understood. That is unlikely to change soon, so a governance framework has to work with the gap rather than assume it away.
How does generative AI help with model development documentation?
By taking the assembly work rather than the judgement work. An internal ratings-based model change generates hundreds of pages: data quality reports, methodology write-ups, parameter narratives, submission packs. Drafting them consumes qualified modeller time without requiring the analytical judgement that justifies the cost. An agentic system that drafts documentation, checks it against a template and routes it for human review handles the procedural load without touching the analytical substance. The distinction that matters is that these systems surface work for a human decision rather than completing it autonomously.
What is retrieval-augmented generation, and why does it suit stress testing?
Retrieval-augmented generation looks up the current version of a relevant document at the moment a question is asked, and answers from that rather than from training data that may be months old. Scenario design is a good fit, because the bottleneck there is staying current: turning a picture of the world that changes weekly into structured quantitative inputs on a fixed cycle. When a new policy statement lands, the knowledge base absorbs it the same day and nothing needs retraining.
What changes in validation when a system uses retrieval?
The object of verification moves. What needs checking is not whether the generated text is coherent, but whether the retrieved document is the right one and whether the knowledge base is current. Retrieval precision and document currency are the two properties to specify before deployment, and neither is tested by reading the output.
Why is regulatory search riskier than it looks?
Because a retrieval failure there is both more dangerous and more silent. When the system retrieves an outdated or mismatched document it still writes a fluent, source-cited answer that reads as authoritative and is wrong, and in a compliance setting that answer can reach a decision before anyone checks the source. This use case therefore adds a third requirement to the other two: boundary detection, meaning the system should say when it cannot find enough context to answer reliably. No current model does that dependably, so the working control is human review of an answer against its cited source, with citations precise enough to take the reviewer straight to the passage.
Why doesn't conventional validation work for these tools?
A risk team's validation toolkit was built for models like a probability of default model: fixed inputs, a deterministic output function, and behaviour reproducible across runs. These tools share none of those properties. An agentic system's behaviour emerges from the interaction of the language model, tool responses and accumulated context state, and that interaction is not fully enumerable in advance. Two identical prompts can produce different tool-call sequences and different outputs, so pre-deployment testing gives weaker guarantees than it does for bounded statistical models.
What governance does a risk team need to add?
Three structural requirements. Structured logging, recording inputs, intermediate reasoning steps, tool calls and outputs so every decision can be audited after the fact. Continuous monitoring, treating each inference cycle as a quality data point rather than relying on a one-off pre-deployment test. And alignment monitoring, catching drift triggered by environment changes such as a revised policy document, an updated data schema or a new retrieval corpus, not only by changes to the model itself. An agentic system in production without that instrumentation is exactly the kind of AI a firm deploys and cannot characterise.
Does SS1/23 cover generative AI?
Not squarely, and the boundary is less explicit on the face of the statement than commentary often implies, so a firm should establish and document its own position on scope rather than assume one. What follows either way is that a supervisory boundary is not a limit on what the function governs internally. A risk team using generative AI in capital, stress testing or regulatory workflows should extend its own model risk discipline to cover those tools: know the system is retrieving the right document and a current one, log what it does so every run can be audited, and keep a human decision at each point that matters.
Sources
- 1 Bank of England and FCA. Artificial intelligence in UK financial services 2024 View source ↗
- 2 Gini. AI adoption has outrun understanding View source ↗
- 3 PRA. SS1/23: Model risk management principles for banks View source ↗
- 4 Gini. SS1/23: model risk management for UK banks View source ↗
- 5 Deloitte. From Hype to Control, Validating Agentic AI View source ↗
