The Global Risks Alliance (GRA)
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.
Generative AI in finance · Model inventory · Committee 7
| MD-264 PD retail | Tier 1 | 12 months | 2 | Overdue |
| MD-251 LLM complaints triage | Tier 1 | never | 0 | Not validated |
| MD-297 IFRS 9 staging | Tier 1 | 5 months | 2 | Approved |
| MD-784 marketing propensity | Tier 3 | 6 months | 1 | Approved |
Fig.Model risk lifecycle. Design wireframe with placeholder names and illustrative values, not a record or a live system.
Generative AI risk in financial services is the risk from large language models and similar tools used to draft, answer, summarise and code, which can produce fluent but wrong output, leak data or be manipulated through inputs. Boards, model risk and security functions, conduct supervisors and legal teams hold the decisions.
For generative ai in finance, in GRA's spatial-finance reading, the map opens on night lights and population density with night lights tonight from NASA GIBS, framed on California; 4 further NASA layers wait in the list.
The interactive map needs JavaScript; the layer list and sources are below.
Layers 2
Connected activity, for this page
Keyboard, with the map focused
Hold Ctrl or ⌘ and scroll to zoom
Latest available
NASA GIBS / EOSDIS · ASTER GDEM (METI, NASA) · Blue Marble (NASA) · © OpenStreetMap contributors (ODbL) · Natural Earth
Loading the latest NASA imagery.
Drag to move. Ctrl or ⌘ and scroll, pinch or the buttons to zoom; hover for the values under the pointer; reset returns to the region.
For generative ai in finance, in GRA's spatial-finance reading, the map opens on night lights and population density with night lights tonight from NASA GIBS, framed on California; 4 further NASA layers wait in the list.Satellite and model products as NASA publishes them, each with its own date; indicators, not assessments of any place; claim ceiling INTEGRATED.
Layers and dates
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.
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.
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.
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.
Provider updates change model behaviour without the firm's code changing. Validation done on one version may not hold for the next.
Each term as this page uses it. Where a standard-setting body defines the term, the wording follows that body.
| Term | Definition |
|---|---|
| Hallucination | Plausible but false or unsupported content generated by a language model. |
| Prompt injection | Instructions hidden in inputs that make a language model ignore intended instructions or act unexpectedly. |
| Grounding | Constraining model answers to retrieved authoritative sources and checking they are supported. |
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.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Generative AI use cases in production | count by risk tier | quarterly | Use case register | DECLARED |
| Factual error rate on evaluation sets | share of responses | per evaluation | Internal evaluation | ESTIMATED |
| Prompt injection test failures | count per test cycle | per test | Red-team results | OBSERVED |
| Data loss prevention events involving AI tools | count | monthly | Security monitoring | OBSERVED |
| Customer complaints linked to AI-generated content | count | quarterly | Complaints data | REPORTED |
| Model version changes by providers | count | per year | Vendor notices | REPORTED |
The topic as a computational graph: named sources feed the methods, an evidence review gates the outputs, and the decisions stay with the authorities that hold them.
Fig. 02Nodes carry their type, two metric rows and a state; connectors show what feeds what; dashed frames name the owners; the timeline is the cadence. Design of the pipeline, not a record of a run.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| FSB report on the financial stability implications of AI | Vulnerabilities from AI adoption in finance | Free, FSB terms |
| IOSCO report on AI use cases, risks and challenges | AI use in capital markets | Free, IOSCO terms |
| NIST AI RMF generative AI profile | Risks and actions specific to generative AI | Public domain |
| OWASP Top 10 for large language model applications | Common security risks in LLM applications | Free, CC BY-SA |
| Bank of England and FCA AI surveys | Adoption and governance of AI in UK finance | Free, Bank of England and FCA terms |
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.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Boards and executive committees | Where generative AI may be used and with what controls | Use case risk tiers and evaluation results |
| Model risk and security functions | How generative AI is validated and secured | Evaluation, red-team and monitoring evidence |
| Conduct supervisors | Whether customer-facing AI meets conduct and disclosure duties | Complaints and outcome evidence |
| Procurement and legal teams | Which providers and data terms are acceptable | Contract terms and data handling evidence |
Generative AI questions start with the use case and its consequences.
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.
| Method or data | What it shows | Limits |
|---|---|---|
| Use-case risk assessment | Consequences of failure by application | Informal uses may not be captured |
| Evaluation against test sets | Accuracy and failure rates on known cases | Cannot cover all real inputs |
| Red-teaming and adversarial testing | Vulnerability to manipulation | Finds some weaknesses, not all |
| Output monitoring and logging | Errors and incidents in use | Detecting subtle errors is hard |
The current service level is R0: information and enquiries. 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.