KYC Review Agent: AI Compliance Review Copilot

Flipping the perspective: instead of helping the user submit, help the compliance team review at 10x speed.

What It Does

KYC Review Agent is an internal AI tool for compliance teams. Instead of a human agent manually comparing each document to the applicant's profile and writing rejection emails, the AI does the first-pass review in seconds, with confidence scores, specific evidence, regulatory citations, and pre-drafted rejection emails.

The Before/After

	Before (Manual)	After (AI Review Agent)
Time per review	~12 minutes	~8 seconds
Reviews per agent/day	~40	~400+
Rejection email	Written from scratch	AI-drafted, agent reviews & sends
Audit trail	Manual notes	Auto-generated with confidence scores
Edge cases	All handled manually	Escalated to senior agent

Decision Framework

Submission Arrives
      │
      ▼
AI Reviews Document  (claude-opus-4-5 Vision)
      │
      ├─ Overall Confidence > 95% + No flags ──▶ AUTO_APPROVE
      │
      ├─ Confidence 70-95% + No errors ────────▶ RECOMMEND_APPROVE
      │                                          (agent confirms)
      ├─ Any error flags detected ──────────────▶ RECOMMEND_REJECT
      │                                          (agent reviews draft email)
      ├─ Confidence < 50% ──────────────────────▶ ESCALATE
      │                                          (senior agent)
      └─ Fraud indicators detected ─────────────▶ FRAUD_FLAG
                                                  (compliance + legal)

Architecture

                         ┌─────────────────────────────────────┐
                         │        Compliance Dashboard           │
                         │  (React + real-time WebSocket queue) │
                         └──────────────┬──────────────────────┘
                                        │
                         ┌──────────────▼──────────────────────┐
                         │       FastAPI Backend (Port 8001)    │
                         │  POST /review                        │
                         │  GET  /queue                         │
                         │  POST /queue/{id}/decision           │
                         │  GET  /analytics                     │
                         └──────────────┬──────────────────────┘
                                        │
                         ┌──────────────▼──────────────────────┐
                         │         KYCReviewAgent               │
                         │  • Claude Vision API                 │
                         │  • Business rule engine              │
                         │  • Decision framework                │
                         │  • Email draft generator             │
                         └──────────────┬──────────────────────┘
                                        │
                    ┌───────────────────┼────────────────────┐
                    ▼                   ▼                    ▼
              PostgreSQL           Redis Cache           S3 Storage
          (submission audit)   (queue/real-time)    (document archive)

Quick Start

git clone https://github.com/your-org/kyc-review-agent
cd kyc-review-agent

cp .env.example .env
# Add ANTHROPIC_API_KEY

docker-compose up

# Submit a document for review
curl -X POST "http://localhost:8001/review" \
  -F "document_type=photo_id" \
  -F "file=@applicant_id.jpg" \
  -F "full_name=John Doe" \
  -F "date_of_birth=1999-06-15" \
  -F "address=456 Oak Ave" \
  -F "city=Vancouver" \
  -F "province=BC" \
  -F "postal_code=V6B1A1" \
  -F "previous_attempts=2" \
  -F "previous_rejection_reasons=Blurry image,DOB mismatch"

Sample Review Packet Output

{
  "submission_id": "uuid-...",
  "dashboard_entry": {
    "ai_decision": "recommend_reject",
    "confidence": "88%",
    "status_color": "orange",
    "flag_count": 2
  },
  "flags": [
    {
      "issue_id": "F001",
      "category": "data_mismatch",
      "description": "Date of birth on document (1985-06-15) does not match profile (1985-06-20)",
      "confidence": 0.94,
      "evidence": "Section showing DOB field clearly states June 15",
      "regulatory_ref": "FINTRAC PCMLTFR s.64(1)(b)"
    }
  ],
  "draft_rejection_email": "Subject: Action Required...",
  "processing_time_seconds": 7.3
}

Infrastructure (AWS)

SQS Queue → Lambda Trigger → ECS Task (Review Agent)
                           → RDS PostgreSQL (audit log)
                           → ElastiCache (real-time dashboard)
                           → S3 (document storage, 90-day retention)
                           → SNS → Email (fraud alerts)

Cost Estimate (AWS, 50K reviews/month - enterprise scale)

Service	Monthly Cost
ECS Fargate (4 vCPU, 8GB)	~$116
Claude API (claude-opus-4-5, ~2K tokens/review)	~$750
RDS PostgreSQL (db.t3.medium)	~$50
ElastiCache Redis	~$30
S3 (document storage)	~$12
SQS + SNS	~$5
CloudWatch + Alerts	~$10
Total	~$973/month
Cost per review	~$0.019
Manual cost per review	~$4.00 (agent salary)
Savings	99.5% cost reduction

KYC Review Agent: assisting compliance teams