AI Won't Fix Your Regulatory Reporting Problems
By James Bowpitt — published 2025-11-10, last updated 2025-12-07.
Every banking technology conference now features the same pitch: artificial intelligence will transform X [insert your business line or use case]. Large language models, we're told, will finally solve the problem of mounting costs and increasingly complex reporting requirements. They won't. The enthusiasm for AI in regulatory reporting betrays a fundamental misunderstanding of what the job requires. Banks must submit numbers with a high degree of accuracy, fully explainable, and reproducible. AI models, particularly the fashionable large language models, deliver none of these things. Given the same input, they may produce different outputs. They cannot guarantee numerical accuracy. They operate as probabilistic systems in a domain that demands determinism. This doesn't mean AI is worthless in regulatory reporting. It means its application must be ruthlessly circumscribed. What the Job Requires Regulatory reporting translates financial and risk data into structured templates mandated by authorities like the PRA. Every figure derives from defined source data through traceable transformations. Regulators apply consistency checks through validation frameworks. Auditors demand that every number can be explained and reproduced. This environment has always favoured rule-based automation: SQL queries, Python scripts, deterministic logic encoded in reporting platforms like Suade, Regnology & AxiomSL. These tools provide transparency and auditability. They don't hallucinate. They don't approximate. They calculate. Any system introducing probabilistic outputs is fundamentally incompatible with submission processes. This rules out using AI to generate, validate, or directly influence reportable values. Where AI Actually Helps The tedious work surrounding those deterministic calculations, however, is another matter. Banks employ armies of analysts, both domestic and offshore, to investigate data quality issues, interpret validation failures, write reconciliation commentary, and explain variances to management. This is where AI, properly constrained, can add some real value. Data quality investigations : By looking at exceptions logs or dropped trade reports, LLMs can interpolate and infer corrections. "This looks like a corporate term deposit." Analysts then adjust, correct upstream booking issues or create deterministic rules in the ETL to correct the issue after verification. Validation failures: When validation rules fail, an LLM can interpret the technical error and explain it in plain language. "This failure occurs because subsidiary data for France appears incomplete in FINREP, causing consolidated totals to mismatch." The model reads logs and metadata. It doesn't touch the numbers themselves. Reconciliation commentary : LLMs can summarise why ledger balances don't match regulatory templates or suggest which account mappings might be causing problems. The deterministic matching remains programmatic. The interpretation becomes automated. Variance explanations : This is the safest application. The quantitative analysis identifying material movements remains entirely rule-based. The LLM simply drafts the narrative. "Deposits from customers decreased by £50 million, driven by corporate term deposit maturities." Analysts review and approve. In each case, the AI operates on by-products of reports: logs, exceptions and variance reports. It never modifies the reported figures themselves. It accelerates interpretation without compromising accuracy. The Neurosymbolic Compromise A more sophisticated approach combines neural networks with symbolic rule engines. The neural component identifies patterns in unstructured data or proposes classifications. The symbolic component (a formal rule engine encoding regulatory constraints) validates every proposal before acceptance. For instance, a neural model might infer from a free-text description that "Client escrow payable, UK" resembles a customer liability. The symbolic layer checks whether this…
Read more regulatory reporting commentary, browse our UK bank regulatory reporting courses, or check open regulatory reporting roles.
Source references include the Bank of England PRA, the European Banking Authority, and the Basel Committee on Banking Supervision.