Generative AI based virtual assistant for reconciliation research

Timely and accurate reconciliation of the company’s finan-cial information is an important internal control over the company’s financial reporting to support quarterly and annual external financial compliance activities. However, due to the complexity of typical end-to-end business system integrations, the process to research and investigate reconciliation items can be manual and time consuming. This paper proposes a virtual assistant to enhance the control process by providing end users with tools that will expedite the reconciliation, research and validation process to improve accountants’ productivity during financial reconciliation. We employ a large language model (LLM) to enable conversational interactions using natural language. Users can pose queries related to investigating unreconciled transactions, and receive relevant explanations and recommended actions. To map user questions into executable SQL, we use a retrieve-and-refine strategy with retrieval augmented generation (RAG). User questions are encoded into vector embeddings and indexed. Given a new question, relevant examples are retrieved and few-shot prompting constructs an SQL generation prompt for the LLM. The generated SQL is executed to retrieve in-formation from databases. The LLM can invoke additional agents and tools to perform more complex operations over the queried dataset, such as generating and executing an adjustment. We optimize prompt engineering to steer the LLM’s behavior, such as providing accounting terminology, instructions, and table schema details. During evaluation, the proposed architecture achieves 95% accuracy in generating correct SQL queries for real-world user questions related to account reconciliation.