Improving finance analyst efficiency with AI-driven resolutions (AIM107)

Here is a detailed summary of the video transcription in markdown format, with sections for better readability and single-level bullet points:

Introduction and Background

• The presenter, Jane, leads the data and analytics team at Cognizant and has experience in implementing transformative programs for AI, analytics, and data governance across various industries.
• The session will focus on an AI solution developed by the team to assist finance analysts in their day-to-day tasks.

Key Use Cases

1. Financial Governance and Taxation:

   • Addressing finance analysts' challenges in responding to governance queries related to accounting procedures, financial reporting, and business policies.
   • Assisting finance analysts in interpreting and applying complex tax legislation and regulations, such as those from the OECD.

2. Automation of UNSPSC Coding:

   • Automating the manual mapping of product descriptions to UNSPSC (United Nations Standard Products and Services Code) codes, which are required for accurate tax computation.

The Proposed Solution

• The solution follows a Retrieval Augmented Generation (RAG) architecture, with a pre-processing engine at its core.
• The pre-processing engine curates the available data (internal and external documents) and provides the relevant context to the language model (LLM) for generating the responses.
• The solution also includes audit-enabled logging, human-in-the-loop validation mechanisms, and confidence scores to ensure the quality and reliability of the responses.

Key Design Principles

1. Security: Addressing security threats, such as prompt attacks and disclosure of sensitive information.
2. Scalability: Designing the solution to be easily expandable and reusable across different use cases.
3. Stability: Ensuring consistent and grounded responses through techniques like data chunking and embedding.
4. Simplicity: Providing traceability and explanations for the AI's decision-making process.
5. Social Awareness: Incorporating factors like neutrality, content filtering, and harm mitigation.
6. Surveillance: Monitoring for hallucination and incorporating human-in-the-loop feedback for continuous improvement.

Architecture Overview

1. Pre-processing Engine:

   • Leverages Amazon OpenSearch and Titan for data embedding and retrieval.
   • Customizes data chunking and embedding techniques based on the specific use case.

2. Core AI Solution:

   • Includes modules for user interaction, prompt engineering, citation management, and output validation.
   • Designed to be scalable and reusable across different use cases.

Implementation Approach

1. Data Understanding and Pre-processing Engine Development:

   • Iterative process to understand the data and build the custom pre-processing engine.
   • Prototyping and validation with subject matter experts.

2. Prompt Engineering and Solution Stabilization:

   • Developing prompts for summarization and response generation.
   • Incorporating human-in-the-loop validation to improve the solution's consistency and accuracy.

3. Guardrails and Evaluation Framework:

   • Applying measures to mitigate hallucination and ensure responsible AI principles.
   • Implementing a validation framework to periodically assess the solution's performance.

Benefits and Challenges

Benefits:

• Cost savings by preventing incorrect responses and misclassifications.
• Improved operational efficiency and quality through automation and consistent responses.
• Compliance with regulatory requirements.

Challenges:

• Handling different data chunking and embedding techniques for various use cases.
• Maintaining context and versioning of documents.
• Leveraging historical data to complement the available information.
• Developing effective confidence score and human-in-the-loop validation mechanisms.

Future Use Cases

1. Financial Reporting:

   • Developing an AI-driven content authoring workbench for regulatory submissions and financial reporting.

2. Forecasting and Business Planning:

   • Leveraging the AI solution to provide alternative scenarios and recommendations for business planning.

3. Fraud and Anomaly Detection:

   • Applying the AI solution to identify patterns and detect anomalies in financial transactions.

4. Compliance:

   • Extending the solution to address compliance-related challenges across different industries.

Conclusion

The presented AI solution demonstrates the potential for leveraging LLMs to address various challenges in the finance domain and beyond. The key is to follow the design principles, implement a robust pre-processing engine, and incorporate human-in-the-loop validation to ensure the solution's reliability and effectiveness.