Banking and Financial Services
Where model governance decides what is deployable
Banks have run quantitative models under supervision for decades, so the question is rarely whether a model can be built. It is whether the model can be governed, explained to a customer, defended to a supervisor, and monitored for drift by people who are accountable for it. The assessment weights those constraints heavily.
Where AI is most often proposed
These are the use-case families the pack recognises. Appearing on this list is not an endorsement — several of them resolve to a no-AI outcome more often than their sponsors expect.
- Credit decisioning, affordability and limit management
- Financial crime detection, sanctions screening and alert triage
- Customer service automation and complaint handling
- Collections prioritisation and forbearance support
- Document processing across onboarding and lending
- Trading, treasury and liquidity analytics
What actually constrains deployment here
Decisions that materially affect a customer
Credit refusal, pricing, limit reduction and account closure attract explainability and fair-treatment obligations. A model that cannot produce a reason a customer would accept is not deployable, whatever its accuracy.
Model risk management
Independent validation, documented assumptions, performance monitoring and defined ownership are treated as prerequisites rather than as later additions.
Discrimination and proxy variables
Features that stand in for protected characteristics create exposure even where the characteristic itself is excluded. The assessment asks how this is tested, by whom, and how often.
Operational resilience and concentration
Dependence on a single provider for a capability supporting a critical business service raises questions about exit, substitutability and continuity.
Outcomes that come up more often than expected
- Approved with conditions, where independent validation and monitoring must be in place first
- Hybrid human and AI, where the model triages alerts but a person makes the determination
- Restricted pilot only, where the population is bounded and every outcome is reviewed
These reflect what the framework tends to conclude given the constraints above. They are not statistics drawn from completed engagements, and no such figures are published because none have been gathered.
Content status
| Property | Value |
|---|---|
| Version | 1.0.0 |
| Maturity | Foundation |
| Validation status | Unvalidated |
| Validated by | Not validated by a named specialist |
| Functions | 17 |
| Processes | 32 |
| Use Case Families | 32 |
| Questions | 20 |
| Hard Gates | 10 |
| Risk Patterns | 12 |
| Controls | 0 |
| Alternatives | 0 |
| Kpis | 30 |
| Financial Templates | 0 |
| Implementation Patterns | 0 |
| Assessment Templates | 5 |
Internally authored by DUN-AI and not externally validated. No named specialist has reviewed this content.
A financial services regulatory specialist and a model risk function should review this pack before its regulatory content informs a decision.