## System Overview
The AI agent system classifies client requests into two primary categories:
– **Category A**: Self-funded transactions (no bank financing required)
– **Category B**: Bank-financed transactions (credit/financing required)
## Architecture Flow
“`
Client Request → Initial Processing Agent → Classification Agent → Route Decision → Specialized Handler
↓ ↓ ↓ ↓
Extract Key Info Analyze Financial Scope Assign Priority Generate Response
“`
## Core Agents Configuration
### 1. Initial Processing Agent
**Purpose**: Extract and structure key information from client requests
**Relevance.ai Configuration**:
– **Tool**: Text Processing + Entity Extraction
– **LLM**: Claude 3.5 Sonnet (for nuanced financial language understanding)
**Prompt Template**:
“`
You are a private banking intake specialist. Extract the following information from the client request:
INPUT: {client_request}
Extract and structure:
1. TRANSACTION TYPE: [Investment, Real Estate, Business Acquisition, Personal Loan, etc.]
2. AMOUNT: [Numerical value and currency]
3. TIMELINE: [Urgency and proposed timeline]
4. CLIENT ASSETS MENTIONED: [Any reference to existing assets, liquidity]
5. FINANCING INDICATORS: [Keywords suggesting need for credit/financing]
6. COMPLEXITY LEVEL: [Simple/Medium/Complex]
7. RELATIONSHIP CONTEXT: [New client, existing client, portfolio size hints]
Format as JSON for downstream processing.
“`
### 2. Classification Agent
**Purpose**: Determine if request requires bank financing
**Relevance.ai Configuration**:
– **Tool**: Decision Tree + Classification Model
– **LLM**: GPT-4 (for logical reasoning and decision-making)
**Prompt Template**:
“`
You are a senior private banker making routing decisions. Based on the structured client information, classify the request.
INPUT DATA: {processed_client_data}
CLASSIFICATION CRITERIA:
CATEGORY A (Self-Funded – No Bank Credit):
– Client has sufficient disclosed liquidity
– Investment management requests
– Asset allocation changes
– Trust and estate planning
– Portfolio rebalancing
– Financial planning consultations
– Existing asset transfers
CATEGORY B (Bank Financing Required):
– Loan applications (personal, business, real estate)
– Credit facilities and lines of credit
– Leveraged investments
– Margin trading requests
– Bridge financing
– Structured financing solutions
– Insufficient liquidity for stated transaction
DECISION FRAMEWORK:
1. Does the transaction amount exceed likely client liquidity?
2. Are there explicit financing keywords (loan, credit, leverage, borrow)?
3. Is this a wealth management vs. credit product request?
OUTPUT FORMAT:
{
“category”: “A” or “B”,
“confidence”: 0.0-1.0,
“reasoning”: “Brief explanation”,
“priority”: “High/Medium/Low”,
“estimated_amount”: “USD amount”,
“suggested_next_steps”: “Recommended action”
}
“`
### 3. Category A Handler (Self-Funded)
**Purpose**: Process wealth management and advisory requests
**Relevance.ai Configuration**:
– **Tool**: Knowledge Base + Response Generation
– **LLM**: Claude 3.5 Sonnet
**Prompt Template**:
“`
You are a private banking relationship manager handling self-funded client requests.
CLIENT REQUEST: {original_request}
CLASSIFICATION DATA: {classification_output}
Generate a professional response that:
1. Acknowledges the client’s request
2. Confirms no financing is needed
3. Outlines next steps for wealth management services
4. Suggests relevant additional services
5. Provides timeline expectations
Tone: Professional, consultative, relationship-focused
Length: 150-250 words
Include: Specific next steps and contact information
“`
### 4. Category B Handler (Bank Financing)
**Purpose**: Process credit and financing requests
**Relevance.ai Configuration**:
– **Tool**: Document Generation + Compliance Check
– **LLM**: GPT-4
**Prompt Template**:
“`
You are a private banking credit specialist handling financing requests.
CLIENT REQUEST: {original_request}
CLASSIFICATION DATA: {classification_output}
Generate a professional response that:
1. Acknowledges the financing request
2. Outlines required documentation
3. Explains preliminary credit assessment process
4. Sets expectation for timeline and next steps
5. Mentions compliance and regulatory requirements
Include standard documentation requirements:
– Financial statements (3 years)
– Asset verification
– Income documentation
– Credit history authorization
– Purpose of funds documentation
Tone: Professional, thorough, compliance-aware
Length: 200-300 words
“`
## Relevance.ai Workflow Setup
### Step 1: Create Knowledge Base
“`yaml
Knowledge Sources:
– Bank product catalog
– Regulatory compliance guidelines
– Client onboarding procedures
– Credit policy documents
– Wealth management service descriptions
“`
### Step 2: Configure Agent Chain
“`yaml
Agent_Chain:
1. Trigger: Email/Chat input
2. Process: Initial_Processing_Agent
3. Classify: Classification_Agent
4. Route: Decision based on Category A/B
5. Handle: Appropriate category handler
6. Output: Formatted response + internal routing
“`
### Step 3: Integration Points
“`yaml
Integrations:
– CRM System: Client data lookup
– Core Banking: Account verification
– Email System: Response delivery
– Ticketing System: Case creation
– Analytics: Performance tracking
“`
## Quality Assurance Framework
### Validation Rules
1. **Amount Threshold Check**: Transactions >$10M always route to human review
2. **Compliance Flag**: Regulatory sensitive requests get compliance review
3. **VIP Client Check**: High-net-worth clients get priority routing
4. **Confidence Threshold**: Classifications <80% confidence trigger human review
### Monitoring Metrics
– Classification accuracy rate
– False positive/negative rates
– Response time per category
– Client satisfaction scores
– Human override frequency
## Sample Test Cases
### Test Case 1: Self-Funded Request
**Input**: “I’d like to reallocate 30% of my portfolio from bonds to emerging market equities. Current portfolio value is $15M.”
**Expected**: Category A – Investment management
### Test Case 2: Financing Request
**Input**: “Looking to acquire a $5M commercial property. Need financing options and terms.”
**Expected**: Category B – Real estate financing
### Test Case 3: Ambiguous Request
**Input**: “Want to discuss expanding my art collection. Considering pieces in the $2M range.”
**Expected**: Requires clarification – could be either category
## Implementation Timeline
**Week 1-2**:
– Set up Relevance.ai workspace
– Configure initial processing agent
– Build knowledge base
**Week 3-4**:
– Implement classification logic
– Create decision tree rules
– Build category handlers
**Week 5-6**:
– Integration testing
– Calibrate confidence thresholds
– Train on historical data
**Week 7-8**:
– Pilot with limited client base
– Monitor and adjust
– Full deployment
## Success Metrics for MVP
– **Accuracy**: >85% correct classification rate
– **Speed**: <30 seconds average processing time
– **Coverage**: Handle 70% of requests without human intervention
– **Client Satisfaction**: >4.5/5 rating on automated responses
– **Cost Reduction**: 40% reduction in manual triage time
## Future Enhancements
1. **Predictive Analytics**: Anticipate client needs based on transaction patterns
2. **Multi-language Support**: Handle requests in multiple languages
3. **Voice Integration**: Process voice requests through speech-to-text
4. **Advanced Personalization**: Customize responses based on client history
5. **Risk Assessment**: Integrate real-time risk scoring for financing decisions
This MVP provides a solid foundation for automated client request classification while maintaining the high-touch service standards expected in private banking.
