Skip to content

GRA · AI and Financial Model Risk

Spatial financeWhere capital meets hazard, exposure and transition

Generative AI in finance

Financial firms use large language models to draft documents, answer customers, summarise research, write code and support analysts. Generative AI produces fluent output that can be wrong, can leak data and can be manipulated through its inputs. This page sets out how generative AI risks arise in financial services, how they are measured and controlled, and who decides where it may be used.

01Canvas

How the work runs.

Systems · around this subjectIllustrative

Model risk lifecycle

Generative AI in finance · Model inventory · Committee 7

3 validations overdueGRA · AI and Financial Model Risk

Lifecycle

  1. Inventoryregistered
  2. Developdocumented
  3. Validateindependent
  4. Approvecommittee
  5. Implementcontrols
  6. Monitorthresholds
  7. Reviewannual
  8. Retiredecommission

Models

MD-264 PD retailTier 112 months2Overdue
MD-251 LLM complaints triageTier 1never0Not validated
MD-297 IFRS 9 stagingTier 15 months2Approved
MD-784 marketing propensityTier 36 months1Approved

Validation scope

Conceptual soundness
In progress
Outcomes analysis
Pending
Benchmarking
Pending
Use limits
human review required
Model risk lifecycle: the screen’s structure, terms and states for Generative AI in finance.Example names and illustrative values; a design, not a record or a live system.

Fig.Model risk lifecycle: example names and illustrative values, not a record or a live system.

02The question

Generative AI in finance: how the risk develops.

Use cases and failure controls. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.

Mechanism

Fluent but wrong

Models can generate plausible statements that are false, citations that do not exist or calculations that are wrong. In finance, an error in a disclosure, advice or code change can harm customers or the firm.

Mechanism

Inputs can be weaponised

Prompt injection through documents, emails or web pages can make a model ignore instructions, reveal data or take unintended actions, especially when it is connected to tools.

Mechanism

Data leaks across boundaries

Confidential information entered into external services or retrieved into answers can reach people not entitled to it. Data handling terms and access controls decide the exposure.

Mechanism

Behaviour changes with versions

Provider updates change model behaviour without the firm’s code changing. Validation done on one version may not hold for the next.

03Exposure, loss and capital

What Generative AI in finance means for exposure, loss and capital.

The exposure, loss, scenario and capital views a risk officer, supervisor or ministry of finance would ask for, with the frameworks they report against. Structures follow supervisory and actuarial practice. Values are illustrative: never prices, ratings or advice.

Transmission channels

How a shock in Generative AI in finance reaches credit, market, underwriting, liquidity and operational risk, and feeds back.

Channel structure after the Basel Committee’s work on climate-related risk drivers and transmission channels; links shown are typical paths, not estimated exposures.
NGFS climate scenarios · Network for Greening the Financial SystemIFRS S2 Climate-related Disclosures · ISSBTCFD recommendations · Financial Stability Board TCFDPrinciples for the effective management of climate-related financial risks · Basel Committee on Banking Supervision
04Aanwysers

What is measured, and how well.

Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.

IndicatorUnitCadenceSource typeLabel
Generative AI use cases in productioncount by risk tierquarterlyUse case registerDECLARED
Factual error rate on evaluation setsshare of responsesper evaluationInternal evaluationESTIMATED
Prompt injection test failurescount per test cycleper testRed-team resultsOBSERVED
Data loss prevention events involving AI toolscountmonthlySecurity monitoringOBSERVED
Customer complaints linked to AI-generated contentcountquarterlyComplaints dataREPORTED
Model version changes by providerscountper yearVendor noticesREPORTED
05Bronne

Where the evidence comes from.

Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.

BronWhat it givesLicence class
FSB report on the financial stability implications of AIVulnerabilities from AI adoption in financeFree, FSB terms
IOSCO report on AI use cases, risks and challengesAI use in capital marketsFree, IOSCO terms
NIST AI RMF generative AI profileRisks and actions specific to generative AIOpenbare domein
OWASP Top 10 for large language model applicationsCommon security risks in LLM applicationsFree, CC BY-SA
Bank of England and FCA AI surveysAdoption and governance of AI in UK financeFree, Bank of England and FCA terms
06Decisions

Who decides, and what the evidence serves.

AI and Financial Model Risk prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

Wie besluitThe questionEvidence this page serves
Boards and executive committeesWhere generative AI may be used and with what controlsUse case risk tiers and evaluation results
Model risk and security functionsHow generative AI is validated and securedEvaluation, red-team and monitoring evidence
Conduct supervisorsWhether customer-facing AI meets conduct and disclosure dutiesComplaints and outcome evidence
Procurement and legal teamsWhich providers and data terms are acceptableContract terms and data handling evidence

Outputs this topic produces

  • Guide to generative AI failure modes in financial services
  • Method note on evaluating factual accuracy and grounding
  • Brief on prompt injection and tool-connected assistants
  • Questions for boards on generative AI use cases
07Core questions

What experts ask.

Generative AI questions start with the use case and its consequences.

  • What is the tool being used for, and what happens if its output is wrong?
  • How is the accuracy of output verified before it affects customers or decisions?
  • What data is sent to the model, and could confidential information leak?
  • How are prompts, versions and outputs logged for accountability?
  • How dependent is the institution on a single provider, and what happens if the service changes or fails?
08Mechanisms

Mechanisms and drivers.

Large language models generate text by predicting likely continuations. They can produce fluent, plausible output that is false, which matters greatly where accuracy is required. Their behaviour can change when providers update models, and they can be manipulated by crafted inputs. Their outputs are hard to test exhaustively because the space of possible inputs is vast.

Risk depends strongly on the use. Drafting an internal summary that a professional reviews carries different risk from answering customers directly or informing credit decisions. Controls therefore scale with consequence: human review, restricted data, retrieval from approved sources, testing against known cases and monitoring of outputs.

09Methods and data

How it is studied, and where the methods stop.

Method or dataWhat it showsLimits
Use-case risk assessmentConsequences of failure by applicationInformal uses may not be captured
Evaluation against test setsAccuracy and failure rates on known casesCannot cover all real inputs
Red-teaming and adversarial testingVulnerability to manipulationFinds some weaknesses, not all
Output monitoring and loggingErrors and incidents in useDetecting subtle errors is hard
10Metodes

How it is studied, and what the methods cannot say.

11Vrae

Questions and answers.

Is a large language model a model under model risk rules?
Supervisors increasingly treat AI tools that inform decisions as models or as part of model risk, and expect firms to govern them proportionately to their use.
What is prompt injection?
An attack in which instructions hidden in inputs, such as a document or web page, cause a language model to ignore its intended instructions or take unintended actions.
What is grounding?
Constraining a model’s answers to retrieved, authoritative sources and checking that statements are supported by them, to reduce fabricated content.
Why do language models produce false information?
Because they generate statistically likely text rather than retrieving verified facts. Without grounding and checks, they can produce plausible but incorrect output.
Should generative AI make credit or claims decisions?
That is for institutions and regulators to decide. Where the consequences for people are significant, strong human oversight and testing are widely expected.
Does the platform evaluate specific AI products?
No. GRA does not evaluate, rate or endorse products or vendors.
VerslaeDra by

Dra bewyse vanaf hierdie bladsy by.

Reports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

Global Risks AllianceAI and Financial Model RiskMethods and TechnicalExponential Technology
KI- en finansiële modelrisiko

Bring an exposure or capital question to the Alliance.

Information and enquiries are open today. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.