In a sector as regulated as finance, conversational artificial intelligence represents both a major opportunity and a considerable challenge. While public AI assistants like ChatGPT or Claude offer impressive capabilities, their use poses unacceptable risks for financial institutions handling sensitive data. Faced with this dilemma, a leading Swiss private bank approached Novatix to design a custom solution: an internal chatbot hosted in a secure environment.

The Challenge: Reconciling AI Innovation with Absolute Security

For this private bank specializing in wealth management, generative artificial intelligence represented a considerable optimization opportunity. However, market solutions posed several fundamental problems.

Regulatory and Security Constraints

The strict regulatory obligations of the Swiss banking sector impose drastic requirements regarding data protection. The bank had to absolutely guarantee that no client data ever leaves the institution's secure environment, that all information processing remains under total control, that complete traceability of interactions is ensured, and that the most demanding security standards are respected.

Identified Operational Needs

Despite these constraints, the bank wanted to offer its employees the advantages of generative AI for:

• Accelerating searches in voluminous internal documentation (procedures, directives, regulations)
• Facilitating the generation and revision of financial analysis reports
• Automating certain repetitive writing tasks
• Enabling analysis of complex financial documents

Bank employees had identified the potential of public AI tools to gain efficiency, but any use of an external solution was categorically excluded for confidentiality and regulatory compliance reasons. A radically different approach was necessary.

Nova Search: An Integrated On-Premise AI Infrastructure

After thorough analysis of needs and constraints, Novatix proposed an ambitious approach: deploying a generative AI system entirely hosted in the bank's infrastructure, without any external connection.

Secure Architecture

The Novatix team designed a robust infrastructure ensuring complete isolation. The system was deployed on a dedicated environment isolated from the external network. The infrastructure was hosted in the bank's secure datacenters in Switzerland, with a total absence of data transfer to external servers. A clustering architecture was implemented to ensure high availability and system resilience.

The architecture was validated by the cybersecurity team and compliance officers before development even began. Each component was selected to guarantee total isolation.

Adapted Technologies and Models

The solution relies on a carefully selected technology stack:

• Open source language model adapted to the bank's needs
• Additional training on internal documentation to integrate specific business context
• Vectorization and indexing system for internal documents for RAG (Retrieval Augmented Generation)
• Intuitive user interface integrated into the existing intranet

Key Architecture Components

The architecture developed by Novatix is structured around five main layers. The frontend layer offers a conversational interface accessible from secure workstations. The application layer manages query orchestration, context, and access controls. The AI layer hosts optimized language models and inference infrastructures. The data layer includes vectorized knowledge bases for contextual enrichment. Finally, the audit layer ensures comprehensive logging of all interactions.

Advanced Customization for Financial Needs

Beyond infrastructure, Novatix developed capabilities specific to the banking context:

• Integration of Swiss financial regulations into the knowledge base
• Financial document analysis functionalities (annual reports, prospectuses, etc.)
• Writing capabilities adapted to financial communication requirements
• Automatic compliance verification algorithms for generated responses

Deployment: A Controlled Phased Approach

Deploying such a solution represented a major project requiring a methodical approach.

Phase 1: Prototype and Proof of Concept

Novatix first developed a functional prototype at reduced scale to:

• Validate the technical feasibility of the on-premise approach
• Test model performance in the constrained environment
• Evaluate interface usability with a pilot group
• Confirm compliance with security requirements

Phase 2: Development and Integration

After concept validation, the team proceeded with complete development:

• Deployment of scaled production infrastructure
• Knowledge base enrichment with internal documentation
• Integration with authentication and rights management systems
• Implementation of audit and traceability mechanisms

Phase 3: Controlled Deployment

User deployment followed a progressive approach:

1. Initial deployment to a pilot group of expert users
2. Adjustment period based on user feedback
3. Extension to a larger group of employees
4. Final deployment to the entire organization

The progressive approach allowed continuous system refinement before large-scale deployment. Each phase brought valuable lessons for tool improvement.

Results: Security and Efficiency Without Compromise

Six months after complete deployment, the AI solution demonstrated its value, with both quantitative and qualitative benefits.

Productivity Impacts

The chatbot deployment generated measurable concrete results. Significant reduction in time spent on documentary research and acceleration in analysis report production were observed. Requests to the documentary support service decreased, freeing time for employees who can focus on higher value-added tasks.

Qualitative Benefits

Beyond quantitative aspects, significant qualitative improvements were observed. Produced document quality increased thanks to writing assistance. Regulatory compliance improved through facilitated access to legal information. New employee onboarding accelerated, and user feedback is predominantly positive.

The power of generative AI is now accessible to employees with absolute guarantee that sensitive data remains entirely under the institution's control.

Validated Security and Compliance

Beyond efficiency gains, the solution successfully passed the most rigorous audits. It obtained complete validation by the internal cybersecurity team and underwent external audit by an independent specialized firm. Its compliance with FINMA requirements for data protection was verified, and no security incident or data leak has been observed since deployment.

Technical Challenges Overcome

Developing the AI solution required solving several major technical challenges.

Performance Optimization

On-premise infrastructure constraints imposed limits in terms of available computing power. The Novatix team overcame this challenge through several innovative approaches. Thorough model optimization was performed, notably through quantization and distillation techniques. Specialized inference algorithms were developed specifically for this environment. An intelligent distributed architecture was designed to efficiently distribute computing load. Finally, advanced caching mechanisms were implemented for frequent queries.

Adaptation to Financial Specificities

Generic open AI models required extensive adaptation to the banking context. Fine-tuning on a specialized financial corpus was performed to improve response relevance. Specific guardrails were developed to avoid generating inappropriate content. Factual verification mechanisms for financial information were integrated. Finally, compliance rules were directly incorporated into the content generation process.

Adapting open-source models to the specific financial context was a major challenge. The Novatix team developed proprietary techniques to ensure relevance and accuracy of responses in this highly specialized domain.

Key Success Factors

The success of this ambitious project relies on several determining factors.

Collaborative Approach

Close collaboration between Novatix and the bank's teams was maintained throughout the project. A mixed steering committee was formed with representatives from all key departments. End users actively participated from the design phase. Business experts were integrated into the content validation process. Permanent coordination was ensured with security and compliance teams.

Training and Adoption

A comprehensive support program facilitated tool adoption. Training sessions adapted to different user profiles were organized. Complete documentation and contextual user guides were created. An internal ambassador network was established to ensure proximity support. Feedback mechanisms were implemented to allow continuous system improvement.

Continuous Improvement

The system was designed to evolve permanently. Regular analysis of usage patterns identifies optimization axes. The knowledge base is constantly enriched with new relevant data. Underlying models undergo regular updates. New functionalities are developed iteratively based on user feedback.

Evolution Perspectives

Building on this initial success, the bank and Novatix have defined an ambitious roadmap for the next steps.

Planned Functional Extensions

Building on this initial success, the bank and Novatix have defined a roadmap for the system's next evolutions:

• Multilingual assistant: extending capabilities to German and Italian to serve the entire Swiss clientele
• External document analysis: capability to analyze client-provided documents while maintaining security
• Visualization generation: automatic creation of graphs and tables from financial data
• Business tool integration: connectors with CRM and wealth management systems

Strategic Vision

The bank's ambition goes beyond a simple internal chatbot. The institution is progressively building a secure AI ecosystem that will transform its working methods while scrupulously respecting its regulatory obligations.

Conclusion

This project illustrates the importance of a custom AI approach. Rather than adopting generic market solutions, organizations handling sensitive data benefit from developing systems perfectly adapted to their specific constraints.

The collaboration between this financial institution and Novatix establishes a precedent in the Swiss banking sector: AI is no longer the domain of lightly regulated technology companies, but becomes a strategic tool compatible with the highest standards of confidentiality and security. This advancement paves the way for responsible digital transformation of the financial sector, where technological innovation reinforces rather than compromises client trust.