2.0 The new risk environment and AI's role in credit risk management

2.1 Credit risk model enhancement

Machine learning models are materially enhancing traditional credit risk approaches, particularly in SME and retail portfolios. By expanding antinovelty models from 10 to around 50 to 60 behavioural and transactional variables, banks can construct a robust proxy for SME cash flows that outperforms standard financial statements and delivers a 3 to 8% uplift in accuracy ratios.

These performance gains translate into meaningfully higher levels of safe automation. Risk leaders spoke about automated retail credit decisions rising from 83% to 91%, and from 19% to 25% in SME lending.

This combination of better discrimination and fewer type 2 errors gives risk functions greater confidence to scale digital origination while preserving credit quality.

Artificial intelligence is now core to the credit value chain, improving risk differentiation, automation and portfolio oversight. However, institutions that succeed are those that treat AI not as a standalone technology project, but as a strategic transformation of credit decisioning, portfolio management and risk control.

2.2 From static portfolio control to dynamic resilience

With today’s risk environment increasingly defined by heightened volatility and interconnected risk channels, stress testing is being repositioned from a regulatory compliance exercise to a central tool for building organisational resilience. This is designed to minimise portfolio surprises and help inform enterprise-wide decision making.

To cope with the complexity of inter-connectedness in credit portfolios, banks are moving away from static, annual risk appetite limits towards continuous, dynamic portfolio optimisation. Risk leaders view integrating portfolio optimisation with frontline underwriting as a priority for closing historical silos so that origination, pricing and portfolio steering can operate on a shared, real time view of risk-adjusted return.

To achieve this, generative AI and broader AI techniques are being explored to alleviate bandwidth constraints and enable real time scenario analysis and playbook execution.

Use cases include automated generation of complex, multifactor stress scenarios that combine human judgement with data driven insight, and the deployment of predefined “scenario-to-action” playbooks that trigger portfolio or capital actions when thresholds are breached.

The challenge for us is we need to improve our organisational resilience. And also, we need to identify and bear in mind specific risk types, and visualise them to avoid creating surprises. All of the inter-connectedness needs more imagination to come up with reasonable scenario testing.

Head of Corporate Risk Management, Asian Pacific banking group

Discussions during RiskMinds International highlighted a structural disconnect: namely that more than 90% of large organisations are already using AI yet fewer than 10% have robust governance frameworks in place to manage associated risks.

Automated data lineage

Centralised governance workflows

Continuous model validation

2.3 Governance & validation gaps

Effective AI governance requires shared ownership across risk, compliance, IT and business units if financial institutions wish to keep pace with rapid AI adoption.

Three focus areas were suggested:

One challenge for model validation is that we can use standard techniques for model validation, but AI brings in other risks that we haven't thought about. Such as the explainability of the output: what is the output telling us? Are the outputs ethical?

Head of Model Risk and Validation & Chief Model Risk Officer, Scandinavian bank

Traditional model validation practices are no longer adequate for dynamic, multi-use AI models. Validation must evolve from point-in-time assessments of individual models to continuous, use-case-centric oversight, focusing on how models are configured, integrated, and deployed in production.

Additionally, there is a growing recognition that capital and pricing infrastructures must evolve to keep pace with AI-driven front-end decisions. Risk leaders have expressed ambitions to leverage technology for more granular, real-time calculations of economic and regulatory capital at the deal level, ensuring that pricing and structuring accurately reflect true marginal risk rather than relying on portfolio averages.

The dynamic and multi-use nature of AI demands ongoing, flexible validation approaches that can adapt more rapidly than current regulatory frameworks mandate.

2.4 Build the foundations

before scaling AI

Organisations must prioritise building governance, validation, and organisational capability alongside AI investments. AI and machine learning already demonstrate clear, quantifiable benefits (improved accuracy, higher automation rates, better early warning and richer scenario analysis). However, in the absence of strong foundations, AI implementations become liabilities rather than competitive advantages.

Those that move fastest to close the governance and validation gap will be able to deploy AI at scale with confidence, turning current pilot level improvements in credit risk performance into a durable competitive advantage.

Institutions are urged to act now by:

Implementing continuous validation and monitoring frameworks, with automated data lineage and centralised workflows, capable of keeping pace with rapid model deployment and change.

Prioritising high-impact, proprietary data use cases and scaling AI literacy so that staff can safely exploit new tools while understanding their limits.

Establishing enterprise AI governance that clarifies ownership, embeds ethical and fairness standards, and links AI use cases explicitly to risk appetite and regulatory expectations.