Fsi Genai Usecases

FSI Generative AI Use Cases

Workload Overview

This case study presents concrete use cases for applying generative AI in financial institutions (banks, insurers, securities firms). All examples are built around Amazon Bedrock and supporting AWS services, and each ships with a public sample implementation. The case study collects four use cases, each addressing a distinct business problem in financial services:

1. Document / Content Review (RAPID) — automates the large volume of document review work in financial institutions using a Human-in-the-Loop model. Guideline documents are turned into checklists automatically, target documents are reviewed against them by generative AI, and a human makes the final decision.
2. AI Sales Roleplay — a training system where a sales representative practices financial-product sales skills by conversing (via voice) with an AI-played customer. It evaluates emotion, compliance, video (facial expression / eye contact), and reference accuracy in real time.
3. Contract Assistant Agent — an AI assistant that comprehensively supports contract-related work (new contract drafting, existing-contract lookup, and answering contract questions) through a supervisor + specialist multi-agent configuration on Amazon Bedrock.
4. ATM Fraud Detection (elderly phone use) — analyzes surveillance-camera video with Amazon Nova (multimodal) to automatically detect behavior patterns that may indicate fraud (e.g. an elderly person operating an ATM while on a phone call), to prevent bank-transfer (振り込め) and special fraud damage before it occurs.

Purpose / significance in financial services. These use cases target recurring pain points in the sector: heavy manual document审查 burden, the need for consistent sales quality and compliance adherence (e.g. suitability principle, explanation obligations under financial instruments regulations), efficient handling of contract operations, and protecting customers — especially the elderly — from fraud. Common business drivers across all four are operational efficiency (24-hour operation, automation of routine work), quality standardization (removing per-person variation), cost reduction (labor savings), and transparency / accountability (AI decision rationale is made explicit, supporting explainability and audit response).

Best Practices and Key Components

Best Practices

• MUST: Keep a human in the loop for final judgment in document review — AI performs the first-pass review and a human makes the final decision and approval (RAPID's Human-in-the-Loop model).
• MUST: Make the AI's decision rationale explicit, so that the reason for a pass/fail judgment can be shown and accountability (説明責任) is preserved.
• MUST: Enforce compliance guardrails on generative AI output in customer-facing / sales-training scenarios — FinanceCompliance and GeneralCompliance guardrails are recommended for financial institutions, checking against financial instruments and banking regulations.
• MUST: Cover the legally required explanation items in sales roleplay (risk explanation, fee explanation, suitability confirmation) and prevent omission, to reduce legal risk from breach of the explanation obligation.
• MUST: Set required goals (必須ゴール) with explicit判定基準 for compliance-related and legally-required items in roleplay scenarios (assign priority 5 for compliance / legal requirements).
• SHOULD: Use a supervisor + specialist multi-agent configuration (Amazon Bedrock multi-agent collaboration) so a single interface can route diverse contract requests to the most suitable specialist agent automatically.
• SHOULD: Use direct browser-to-S3 upload via a time-limited presigned URL (5-minute expiry) to keep uploads secure and to avoid API Gateway payload limits for large video files.
• SHOULD: Apply S3 lifecycle policies to surveillance video and DynamoDB TTL to automatically delete old / unneeded data, for both cost and data management.
• SHOULD: Minimize false positives by correctly identifying normal customer behavior (short calls, calm demeanor, smooth operation) so business efficiency and customer experience are maintained.
• PREFER: A serverless architecture (API Gateway + Lambda + Amazon Bedrock + S3 + DynamoDB) for these GenAI workloads to balance scalability and operational efficiency through loosely-coupled layers.
• PREFER: Deploying via AWS CloudShell (no prerequisites) for the AI sales roleplay sample, running the provided bin.sh deploy script.
• MAY: Restrict self-sign-up and choose a specific Bedrock region for the roleplay sample via deploy options (--disable-self-register, --bedrock-region).
• MAY: Define custom guardrails (e.g. a 金融商取法 guardrail: no assertive judgments, material-fact explanation, suitability principle, no exaggerated advertising) when organization-specific compliance requirements exist.
• WILL: Detection of high-risk fraud immediately notifies store staff so they can speak to the customer and prevent damage before it occurs.
• WILL: The ATM fraud-detection Amazon Nova analysis performs integrated multimodal analysis of both video and audio to identify fraud patterns.
• WILL: The contract-assistant generated draft contracts are stored in S3 and provided to the user; sample templates include purchase, franchise, and time-and-material contracts.

Key Components

Amazon Bedrock (Generative AI Foundation)

• Component Overview: The central generative-AI engine across all four use cases. Hosts the LLMs used for document analysis and review (RAPID), roleplay dialogue and compliance checking (via Anthropic Claude), multi-agent contract processing, and multimodal video analysis (via Amazon Nova). Also provides multi-agent collaboration and guardrail capabilities.
• Assumed AWS Services: Amazon Bedrock, Amazon Nova, Anthropic Claude (on Bedrock), Bedrock Agents, Bedrock multi-agent collaboration, Bedrock Guardrails.

Multi-Agent Orchestration (Contract Assistant)

• Component Overview: A supervisor agent (Contract Agent) receives user contract inquiries, analyzes intent, and routes to the optimal specialist among three sub-agents — contract-drafting, existing-contract-lookup, and contract-question answering — then composes the final answer. Implemented with Amazon Bedrock multi-agent collaboration.
• Assumed AWS Services: Amazon Bedrock (Agents, multi-agent collaboration), AWS Lambda (existing-contract lookup), Amazon S3 (generated contracts / templates), Amazon OpenSearch Service (knowledge base for contract Q&A).

Serverless Processing Layer

• Component Overview: Event-driven, serverless compute that handles presigned-URL generation, analysis triggering, status checks, and general application logic across the use cases. In ATM fraud detection this is three Lambda functions (presigned-URL, auto-analysis, status). Loosely-coupled layers (frontend, API, processing, AI analysis, data) enable flexible scaling and maintenance.
• Assumed AWS Services: AWS Lambda, Amazon API Gateway.

Data & Storage Layer

• Component Overview: Stores documents, videos, scenario data, evaluation results, checklists, and analysis metadata. S3 holds documents (RAPID), generated contracts and templates (contract assistant), video files (fraud detection), and scenario / evaluation data (roleplay). DynamoDB manages checklists and review results (RAPID), analysis results / metadata (fraud detection), and user info / progress (roleplay).
• Assumed AWS Services: Amazon S3, Amazon DynamoDB.

Web Delivery & API Layer

• Component Overview: Delivers the browser-based UIs and REST APIs. CloudFront provides low-latency delivery of the web applications; API Gateway exposes REST API endpoints and manages authentication/authorization and rate limiting. The ATM fraud-detection API exposes a presigned-URL endpoint (GET /presigned-url) and a status endpoint (GET /status/{videoId}).
• Assumed AWS Services: Amazon CloudFront, Amazon API Gateway.

Multimodal & Media Analysis Services

• Component Overview: Services that process non-text modalities. Amazon Nova (on Bedrock) analyzes ATM surveillance video and audio for fraud patterns. In AI sales roleplay, Amazon Transcribe converts speech to text and Amazon Rekognition analyzes video for non-verbal communication (facial expression, eye contact, gestures).
• Assumed AWS Services: Amazon Nova (on Amazon Bedrock), Amazon Transcribe, Amazon Rekognition.

Operations & Monitoring

• Component Overview: Observability for the workloads (documented in detail for the ATM fraud-detection use case). CloudWatch Metrics/Logs collect metrics and aggregate logs; CloudWatch Alarms detect anomalies and notify the operations team via SNS, visualized through an integrated dashboard. Monitored items include Lambda duration and error rate, API Gateway latency/errors, DynamoDB read/write capacity, S3 storage usage, and Bedrock API call count and errors.
• Assumed AWS Services: Amazon CloudWatch (Metrics, Logs, Alarms), Amazon SNS.

FISC Compliance Summary

There is no dedicated FISC mapping file for this case study, and the source documents do not contain an explicit FISC safety-standards mapping. Therefore no references/fisc-mapping.md is created.

The source documents do, however, address financial regulatory / compliance concerns qualitatively, which are relevant to FISC-aligned governance:

• Compliance guardrails based on the Financial Instruments and Exchange Act (金融商取法) and Banking Act (FinanceCompliance / GeneralCompliance), including prohibition of assertive judgments, the material-fact explanation obligation, the suitability principle (適合性原則), and prohibition of exaggerated advertising.
• Explanation-obligation adherence and customer protection in sales activities (risk explanation, fee explanation, suitability confirmation).
• Transparency and accountability: making AI decision rationale explicit to support explainability and audit response; systematic management of detection history and response records for audits (ATM fraud detection).
• Document review against laws/regulatory guidelines (e.g. FSA supervisory guidelines, industry self-regulatory rules), internal compliance rules, and information-security policies (RAPID review-criteria categories).

No further FISC control mapping should be inferred beyond these source-stated points.

Reference Architecture

Each use case has its own architecture and workflow. Details, including diagrams and per-use-case flows, are in the Japanese reference files:

• references/architecture.md — case-study overview and cross-cutting architecture.
• references/contract-assistant-agent.md — multi-agent contract assistant.
• references/document-review.md — RAPID document / content review.
• references/cctv-fraud-detection.md — ATM fraud detection with Amazon Nova.
• references/ai-roleplay.md — AI sales roleplay training.
• references/ai-roleplay-scenario-guide.md — financial roleplay scenario creation guide (6 steps).

CDK Sample Overview

No CDK sample is provided for this case study. The individual use cases are deployed from their own public sample repositories (referenced in the Japanese reference files), for example via AWS CloudShell and a provided deploy script (bin.sh) for the AI sales roleplay sample. This case study asset itself contains no CDK code.