AI-Assisted Regulatory Intelligence and Adaptive Compliance Verification System for Banking Supervision
Abstract Regulatory requirements issued by the Reserve Bank of India (RBI) fall into categories including numerical thresholds, deterministic conditions, temporal obligations, and qualitative standards such as ''adequate due diligence,'' each of which requires a different mode of verification. Bank employees and officers must first locate and interpret the appropriate requirement among an ever-growing corpus of Master Directions, circulars, and notifications, while the regulator must independently verify (at scale) whether submitted bank evidence satisfies the requirement. We present a two-module regulatory intelligence system that addresses both sides of this problem in a single pipeline. The Regulation Comprehension module ingests an RBI circular or Master Direction and segments and categorizes its clauses, extracts a typed and machine-readable rule representation, simplifies regulatory language for reading by humans, and enables natural-language question answering with clause-level sourcing. The main technical contribution is an adaptive reasoning-modality router that determines, for a given regulatory requirement and evidence set, whether a deterministic rule engine, a retrieval-based large language model (LLM), or human escalation should be used, governed by an explicit cost--accuracy--latency objective and a hard architectural override that directs fully specified deterministic rules and complete evidence to the rule engine unconditionally. This differs from the well-studied task of routing between LLMs of differing capability, since our router chooses among fundamentally different reasoning mechanisms rather than among models. We also contribute a rule-grounded synthetic RBI compliance benchmark across nine difficulty categories and a dedicated regulation-comprehension benchmark design (with extraction, question-answering, and simplification variants), such that the two modules can be evaluated with equal rigor rather than treating comprehension as a subordinate preprocessing step, yielding two distinct sets of metrics, baselines, and ablation points, each for one of the modules, across extraction accuracy, retrieval and answer-faithfulness, simplification quality, compliance accuracy, and routing efficiency. We also contribute an evidence-grounded explanation architecture that traces every such decision and comprehension answer to a specific rule, clause, and evidentiary or source field. Since operational bank data cannot be accessed for confidentiality and regulatory reasons, evaluation is performed on a synthetic, rule-grounded benchmark of 1,393 cases drawn from 215 rules. On the 294 held-out test cases, whose rules were never seen in training or tuning, the adaptive router reaches 92.2\% automated accuracy and 96.9\% end-to-end accuracy, compared with 85.4\% for a fixed hybrid router, 72.5\% for retrieval-augmented LLM-only reasoning, and 56.8\% for LLM-only reasoning, and ablations show that removing adaptive routing, the rule engine, or evidence grounding lowers accuracy. Adversarial and ambiguous cases remain the hardest. All results are on synthetic rather than operational bank data, the decisions of the LLM reasoning paths in these runs were produced by a deterministic surrogate of the LLM judge, and results of the regulation-comprehension benchmark are not reported in this paper.
Authors
- Pratiksha Deshmukh
- Shripad Kulkarni
- Pranav Bhoite
- Nisarg Wath
- Sofiya Rukadikar
Institutions
- COEP Technological University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11450785/v1
- Primary Topic
- Artificial Intelligence in Law
- Type
- preprint