Financial supervisors have an unusual technological advantage: they can require regulated entities to submit specified information in prescribed formats, at defined frequencies and subject to validation requirements. This means that SupTech does not need to become an AI-driven black box. Where reliable structured data are available, much of the supervisory process can—and generally should—remain deterministic, with AI concentrated where uncertainty, judgement and unstructured information matter most.
- Start with data architecture, not AI models. Supervisors can define reporting populations, taxonomies, validation rules, granularity, reconciliation requirements and resubmission procedures. The real foundation of SupTech is therefore regulatory data engineering. Better AI cannot compensate for poorly defined or inaccurate regulatory data.
- Keep the supervisory spine deterministic. Capital, liquidity and leverage calculations; large-exposure checks; reporting completeness; reconciliations; threshold breaches; peer ratios; and predefined escalation rules are naturally suited to rules engines. They provide reproducibility, audit trails and explainability—qualities essential when supervisory decisions must be internally reviewable and legally defensible.
- Distinguish calculation from judgement. Computing a capital ratio can be deterministic, but deciding whether counterparties are connected, whether an exposure is forborne or whether significant credit deterioration has occurred may involve judgement. Moreover, many supposedly “deterministic” regulatory outputs rely on probabilistic inputs such as IRB parameters, expected credit losses or behavioural assumptions.
- Use AI to expose what fixed rules may miss. Machine learning can identify anomalous classifications, unusual peer behaviour, networks, changing transaction patterns and evidence of firms managing positions around known regulatory thresholds. Such outputs should normally be investigative leads rather than supervisory findings.
- AI becomes more important when supervision looks forward. Tomorrow’s capital position, future liquidity stress or emerging credit deterioration cannot be calculated exactly because they have not yet occurred. Evidence from supervisory research shows that machine-learning techniques can improve bank-distress prediction, although rare events, short supervisory time series, regime changes and changing reporting definitions impose important constraints.
- Stress testing illustrates the ideal hybrid architecture. Accounting identities, capital mechanics and flow-through rules should remain deterministic, while behavioural parameters—such as default probabilities, deposit behaviour and funding reactions—may be estimated statistically or through machine learning. Structural constraints remain necessary because flexible models may perform poorly when extrapolating beyond historical experience.
- Generative AI fits naturally at the reporting end. It can summarise risk movements, compare examinations, draft risk narratives, connect findings with rules and prepare management briefings. Current supervisory adoption is indeed concentrated heavily in document processing, knowledge management and review rather than autonomous decision-making.
The strongest SupTech architecture is therefore not “AI everywhere.” It is a traceable deterministic core surrounded by carefully governed AI: rules for what can be calculated, statistical methods for what must be estimated, and human judgement for what must ultimately be decided.
Selected References
Bank for International Settlements. (2024). Artificial intelligence and the economy: Implications for central banks. Annual Economic Report 2024.
Beerman, K., Prenio, J., & Zamil, R. (2021). Suptech tools for prudential supervision and their use during the pandemic. FSI Insights No. 37.
Doerr, S., Gambacorta, L., & Serena, J. M. (2021). Big data and machine learning in central banking. BIS Working Papers No. 930.
Gambacorta, L., Lauridsen, N., Kiuhan-Vásquez, S., & Prenio, J. (2025). Making suptech work: Evidence on the key drivers of adoption. BIS Working Papers No. 1309.
Prenio, J. (2025). Starting with the basics: A stocktake of gen AI applications in supervision. FSI Briefs No. 26.
Suss, J., & Treitel, H. (2019). Predicting bank distress in the UK with machine learning. Bank of England Staff Working Paper No. 831.
Leave a Reply