The FSB’s October 2026 report on responsible AI adoption emphasises risk-based technology selection, explainability, performance monitoring and human oversight. While the report principally addresses financial institutions, these principles can also inform supervisory technology choices. A comparative assessment of supervisory practices, based on FSB insights:
Table 1. Comparative assessment
| Supervisory analytics dimension | Conventional automation | AI-based analytics | Preferred approach |
|---|---|---|---|
| 1. Regulatory returns validation | Validates prescribed formats, thresholds, reconciliations and mathematical rules | Detects unfamiliar patterns and potentially anomalous submissions | Automation for routine validation |
| 2. Capital adequacy monitoring | Calculates CAR, leverage ratios, regulatory capital and RWA using prescribed rules | Identifies emerging deterioration and unusual capital trends | Automation for regulatory calculations; AI for supplementary analysis |
| 3. Liquidity risk monitoring | Calculates LCR, NSFR, maturity gaps and predefined stress indicators | Detects complex funding patterns and emerging liquidity vulnerabilities | Hybrid |
| 4. Early warning indicators | Uses predefined ratios, thresholds, trends and composite scores | Detects nonlinear relationships and complex combinations of risk indicators | Hybrid, with AI where incremental predictive value is demonstrated |
| 5. Risk-based supervisory ratings | Applies approved risk and impact-scoring methodologies consistently | Identifies potential inconsistencies and suggests risk classifications | Automation for formal scoring; AI as decision support |
| 6. Inspection planning | Ranks institutions using established supervisory priorities, risk scores and inspection cycles | Identifies emerging risk clusters and suggests inspection focus areas | Hybrid |
| 7. Historical inspection observations | Searches structured databases using rules, keywords, regulatory references and taxonomies | Uses semantic similarity to identify related observations and suggest classifications | AI for semantic retrieval; automation for exact rule matching |
| 8. Regulatory breaches detection | Identifies breaches of explicit regulatory limits and requirements | Flags potential noncompliance in complex transactions or unstructured narratives | Automation for definitive rule-based breaches |
| 9. Credit portfolio analysis | Monitors NPLs, concentrations, arrears, sector limits and migration matrices | Identifies complex borrower relationships, deterioration patterns and emerging vulnerabilities | Hybrid |
| 10. Stress testing | Executes approved scenarios, deterministic formulas and established statistical models | Supports nonlinear modelling, scenario exploration and pattern identification | Conventional models for approved calculations; AI where validated |
| 11. Market conduct supervision | Flags transactions or practices breaching predefined parameters | Detects unusual conduct patterns, potential misconduct and suspicious relationships | Hybrid, with AI for complex pattern detection |
| 12. Complaints analytics | Calculates volumes, resolution periods, overdue cases and complaint categories | Classifies free-text complaints, detects emerging themes and identifies sentiment patterns | AI for text analysis; automation for quantitative monitoring |
| 13. Financial statement analysis | Computes ratios, variance analysis, peer comparisons and trend indicators | Extracts narrative information and identifies inconsistencies across disclosures | Hybrid |
| 14. Regulatory reporting dashboards | Produces repeatable KPIs, charts, reconciled data and scheduled management reports | Generates narrative interpretations and answers natural-language questions | Automation for official dashboards; AI for supplementary explanations |
| 15. Supervisory decision support | Applies documented decision rules and workflow controls | Generates interpretations, possible explanations and recommendations | Human-led hybrid, with documented supervisory accountability |
Table 2. Operational strengths and weaknesses
| Evaluation criterion | Conventional automation | AI-based analytics |
|---|---|---|
| Accuracy | High for correctly specified deterministic tasks | Variable; requires testing and validation |
| Explainability | Generally high for documented rules | Can be limited for complex models |
| Auditability | Usually straightforward | Requires additional traceability and controls |
| Implementation cost | Often lower for stable, structured tasks | Can be higher due to model, infrastructure and governance costs |
| Maintenance | Rule updates and data pipeline maintenance | Model monitoring, retraining, drift testing and version controls |
| Unstructured data | Limited without specialised tools | Strong potential for extracting and classifying text |
| Predictive capability | Effective with traditional statistical models | Potential advantage for complex nonlinear relationships |
| Reproducibility | Generally deterministic | May vary depending on model and configuration |
| Data requirements | Structured and well-defined | Often needs extensive, representative data |
| Supervisory accountability | Rules and decisions readily attributable | Requires explicit human accountability and safeguards |
Conventional automation should generally be the default for deterministic regulatory calculations, compliance checks, reconciliations and formal supervisory scoring. AI should be considered where it produces demonstrable incremental value, particularly in unstructured information analysis, anomaly detection and complex risk-pattern identification.
The strongest supervisory architecture will often combine conventional data pipelines, statistical models, rules engines and carefully validated AI applications rather than attempt to replace established supervisory systems.
Reference: Financial Stability Board. (2026, October 15). Sound practices for responsible adoption of artificial intelligence (AI): Final report. FSB report



Leave a Reply