Policy
Exceptions
Detect potential lending policy violations using applicability and effective-date metadata, not just semantic similarity.
Key capabilities
What reviewers get
The policy version that actually applied is used
Applicability filters remove irrelevant rules
Exceptions cited to the exact policy clause
Severity classified for triage
Operating principle
AI proposes. Deterministic services calculate. Source documents prove. Humans decide.
How it works
A grounded pipeline that keeps calculation, policy, and human judgment clearly separated.
Ingest source documents
Credit memos, financial statements, and supporting files are parsed into structured data with full source provenance preserved.
Calculate deterministically
A deterministic calculation engine computes figures and comparisons outside the language model, so numbers are reproducible and auditable.
Reason with grounded AI
AI proposes findings only from extracted evidence and the applicable, effective-dated policy — never from unverifiable assumptions.
Route to the reviewer
Every finding arrives with its source citation and severity so a human reviewer can confirm, adjust, or dismiss it — and decide.
An illustrative finding
IllustrativeA fictional example showing how a single issue flows from source data to a reviewer decision. Figures are illustrative only.
Scenario
A credit was underwritten in August 2025 and must be checked against the policy in effect on that date.
Extracted data
Credit dated Aug 2025. Two policy versions exist: v4.1 (effective Jan 1) and v4.2 (effective Jul 1, 2025).
Deterministic check
Effective-date logic selects v4.2 as the applicable version for an August 2025 credit.
Applicable policy
Commercial Lending Policy v4.2 §5.3.2 tightened the minimum current ratio guideline.
AI finding
AI proposes a potential exception evaluated against v4.2 — not the superseded v4.1.
Reviewer action
Reviewer confirms the applicable version and clause, then dispositions the exception.