CrediSense — GenAI Credit Policy Assistant
A Microsoft Copilot Studio-based conversational assistant enabling banking credit users to query complex policy documents, retrieve structured guidance, and navigate approval logic — grounded in official policy sources with full governance controls.
Executive Summary
Credit policy documents in banking are extensive, frequently updated, and difficult to navigate consistently. CrediSense is a GenAI-powered assistant that allows credit users to ask policy questions in natural language and receive structured, policy-grounded answers — reducing lookup time and improving decision consistency.
Business Problem
Large credit policy manuals span hundreds of pages across multiple documents, covering product types, borrower categories, approval authorities, exceptions, and review conditions. Inconsistency in policy interpretation is a risk in itself — when different users interpret the same policy differently, it creates decision variability affecting credit quality and compliance.
My Role
I led the design and development from problem definition through deployment: conversation architecture, topic routing logic, adaptive card design, policy grounding configuration, guardrails, and stakeholder testing. I worked closely with credit policy and risk stakeholders throughout.
Architecture
The routing layer handles multiple policy topic areas independently, each with its own input logic, policy sources, and answer structure — keeping the assistant focused and accurate rather than attempting to answer everything from a single undifferentiated prompt.
Key Capabilities
Returns the relevant authority matrix based on facility type and exposure amount
Routes to borrower-specific policy rules based on entity type
Covers renewal conditions, review triggers, and documentation requirements
Surfaces conditions requiring board or committee escalation
Governance & Controls
Business Impact
Lessons Learned
Confidentiality Note: Due to employer obligations, code, raw data, proprietary models, and internal investigation details are not disclosed. This case study presents architecture, methodology, and business impact only.