Enterprise GenAI Virtual Assistant for Operations and Underwriting Knowledge Access

radian · Finance · year n/a · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Mortgage operations and underwriting knowledge access

Radian needed operations and underwriting teams to access thousands of internal policies and knowledge sources quickly while preserving trust, accuracy, governance, security, observability, and role-based permissions in a regulated financial-services environment.

Constraints:

• Knowledge lived across SharePoint, Confluence, and other internal systems.
• The assistant needed to support multiple business lines rather than becoming separate bots for each system.
• Trust and answer accuracy were the main adoption requirements for operations users.
• Responses had to respect user roles and permissions.
• Security, observability, governance, and cross-functional buy-in had to be built early.

Architecture

Radian built Radian Virtual Assistant as an enterprise knowledge platform rather than a single chatbot. It uses AWS Bedrock Knowledge Base to connect trusted content sources, retrieve verified content, generate accurate and traceable answers, and enforce enterprise-grade security through role and permission boundaries.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  radian-virtual-assistant["Radian Virtual Assistant"] --> unified-knowledge-layer["Unified knowledge layer"]
  radian-virtual-assistant --> bedrock-knowledge-base["Bedrock Knowledge Base"]
  unified-knowledge-layer --> bedrock-knowledge-base
  bedrock-knowledge-base --> answer-generation["Answer generation"]
  answer-generation --> role-and-permission-controls["Role and permission controls"]
  role-and-permission-controls --> observability-and-feedback["Observability and feedback"]

Components

Data flow

Users ask natural-language questions; the assistant searches unified knowledge sources such as SharePoint and Confluence; Bedrock Knowledge Base retrieves verified content; the system builds accurate, relevant, traceable answers constrained by user roles and permissions.

Why this architecture

Radian chose an enterprise assistant and unified knowledge layer to avoid bot sprawl, preserve governance/auditability, and make thousands of policies instantly searchable across operational units.

Models used: AWS Bedrock · Amazon Bedrock Knowledge Base · GenAI

Scale

• The assistant provides instant access to thousands and thousands of policies across operational units.
  “instant access to thousands and thousands of policies across operational units”transcript, stated
• The target use cases span operations, vendor management, title business, MI operations, and support teams.
  “whether it's MI operations, whether it's our vendor management kind of like business line or title business”transcript, stated

Alternatives considered (roads not taken)

Build separate bots for separate systems.

Rejected because: Radian wanted an enterprise application with unified knowledge, governance, control, auditability, and scalability.

Let teams keep searching countless files manually.

Rejected because: Bedrock Knowledge Base connects directly to trusted content and retrieves verified pieces for the answer.

Patterns

rag-retrieval (retrieval)

Radian uses Bedrock Knowledge Base to retrieve verified internal content before generating responses.

privacy-preserving-processing (governance)

Responses are defined by user roles and permissions in the regulated environment.

observability-tracing (operations)

Radian invested early in observability and security hooks for operationalization.

Tradeoffs

Build an enterprise platform rather than a chatbot.

Chose Trusted scalable assistant across enterprise use cases. over Single-purpose chatbot..

Rationale: Radian needed governance, auditability, centralized control, and multiple business-line support.

Use Bedrock Knowledge Base over manual search.

Chose Verified retrieval-backed answer generation. over Teams searching countless files themselves..

Rationale: The knowledge base retrieves trusted pieces and builds traceable answers.

Gotchas (what bit them)

Bot sprawl recreates fragmented knowledge access.

Root cause: Separate bots for disparate systems would mirror the existing silos.

Resolution: Build a unified enterprise assistant and knowledge layer.

Users will not adopt answers they do not trust.

Root cause: Operations teams need accurate and traceable information for regulated work.

Resolution: Prioritize trust, accuracy, verified content, and feedback metrics.

Governance cannot be bolted on after launch.

Root cause: Regulated industries require security, observability, and cross-functional buy-in.

Resolution: Include operations, legal, IT, and security early.

Evaluation

Radian evaluates through answer accuracy, trust, user feedback, ground-truth data, observability, and security hooks.

Metrics: Accuracy of answers · User feedback · Ground-truth data · Observability signals · Security hooks

Outcomes

• Radian Virtual Assistant provides unified access to internal policies and knowledge sources.
• The assistant supports multiple enterprise use cases beyond operations and underwriting.
• The architecture emphasizes trust, traceability, governance, and role-based access.

The clearest scale signal is thousands of policies across multiple operational units and business lines.