BNY Mellon implemented an LLM-based virtual assistant to help their 50,000 employees efficiently access internal information and policies across the organization. Starting with small pilot deployments in specific departments, they scaled the solution enterprise-wide using Google's Vertex AI platform, while addressing challenges in document processing, chunking strategies, and context-awareness for location-specific policies.
# BNY Mellon's Enterprise Virtual Assistant Implementation
## Background and Context
BNY Mellon, a 250-year-old global financial services institution, faced the challenge of making vast amounts of internal information easily accessible to their 50,000 employees worldwide. The organization manages, moves, and safeguards financial assets, dealing with complex procedures and policies that vary across countries and departments.
## Initial Challenge and Solution Approach
### Problem Statement
- Traditional search methods required employees to sift through lengthy documents (often 100+ pages)
- Information was scattered across multiple sources and departments
- Different regions had different policy requirements
- Existing systems were limited to basic help desk functions (password resets, laptop orders, etc.)
### Solution Evolution
- Built upon existing conversational AI experience within the bank
- Leveraged existing relationship with Google Cloud and Vertex AI
- Implemented gradual rollout strategy starting with specific departments
## Technical Implementation Details
### Platform Selection
- Chose Google Cloud's Vertex AI ecosystem
- Decision influenced by:
### Document Processing Challenges
### Chunking Strategy Evolution
- Initial approach used basic document chunking
- Discovered limitations when dealing with diverse document types
- Required multiple iterations of chunking strategy
- Developed specialized approaches for different document formats
### Special Document Handling
- Encountered challenges with complex documents like cafeteria menus
- Implemented Google Document AI for specialized document processing
- Developed custom processing pipelines for different document types
### Context-Aware Information Delivery
### Location-Based Intelligence
- Implemented system to understand user context (location, role)
- Developed capability to serve region-specific policies automatically
- Example: System automatically serves New York policies to NY-based employees and UK policies to UK-based employees
### Knowledge Management Integration
### Content Optimization
- Provided feedback to knowledge management teams
- Implemented improved document tagging systems
- Added metadata requirements for better context understanding
- Enhanced content structure for LLM compatibility
## Monitoring and Improvement Systems
### Current Feedback Mechanisms
- Manual review of questions and responses
- Performance monitoring of virtual assistant
- Analysis of user interactions
- Content effectiveness tracking
### Planned Enhancements
- AI-powered automated performance analysis
- Bias detection systems
- Context accuracy monitoring
- Automated metrics generation
- Knowledge source effectiveness tracking
## Lessons Learned and Best Practices
### Content Management
- Documents need specific structuring for LLM consumption
- Metadata and tagging crucial for context awareness
- Regular content curation and updates necessary
- Importance of backwards compatibility with existing systems
### Implementation Strategy
- Start small with pilot programs
- Gradual expansion across departments
- Continuous feedback collection and implementation
- Regular strategy adjustment based on user needs
### Technical Considerations
- Need for flexible chunking strategies
- Importance of context-aware processing
- Value of specialized document processing tools
- Balance between automation and human oversight
## Future Development Plans
### Planned Enhancements
- Automated performance monitoring using AI
- Extended coverage to structured data sources
- Enhanced metrics and visibility tools
- Model tuning and optimization
- Expanded knowledge base integration
### Ongoing Initiatives
- Content improvement programs
- Enhanced metadata implementation
- Improved tagging systems
- Better regional context handling
## Impact and Results
### Current State
- Successfully deployed to all 50,000 employees
- Handles diverse information requests across departments
- Provides context-aware responses based on user location
- Integrated with multiple knowledge sources
- Improved information accessibility and efficiency
### Ongoing Benefits
- Reduced time to find information
- Improved accuracy in policy compliance
- Enhanced employee experience
- Streamlined information access
- Better knowledge management practices
Start your new ML Project today with ZenML Pro
Join 1,000s of members already deploying models with ZenML.