OBJECTIVES: This study aims to develop a virtual assistant interface specifically designed for a bespoke Excel-based cost-effectiveness (CE) model. This interface will leverage the capabilities of a large language model (LLM), Claude-3, to perform various tasks within the model and customize it for different markets.

METHODS: A user interface (UI) was developed utilizing Claude-3-Opus model (with a backend in Python 3.11.5 and a front end using Jinja2 templates) to facilitate the retrieval and updating of data in an Excel model, leveraging generative AI capabilities. Users can input natural language prompts and select model parameters from a dynamically generated dropdown menu. The LLM processes these inputs with system prompts and Excel column headers, generating logic codes without transmitting Excel data, ensuring privacy. These codes are executed in Python interpreter to retrieve or modify the Excel model, with results displayed on the user interface. A comprehensive set of 30 prompts were designed to test the actual model updates: 10 for data retrieval and 20 for updates in the model. The interface includes multi-parameter change functionality for country adaptations using an input sheet and has been tested with 20 distinct input sheets.

RESULTS: The AI interface correctly processed the pre-defined set of prompts: 10/10 for data retrieval and 20/20 for data updates in the Excel model. Additionally, the multi-parameter change functionality successfully updated the Excel model with the new values from the uploaded input sheets for all cases.

CONCLUSIONS: This study confirms that generative AI can be used to effectively adapt Excel models for country-specific needs, enhancing data management operations and ensuring accuracy. The Assistant UI introduces a user-friendly interface, simplifying exploration of complex Excel models and making them more accessible to lay users. Future applications could see the Assistant UI providing a unified platform for accessing various Excel models, ensuring a consistent and streamlined user experience.