How Venture Capital Could Use Large Language Models to Screen Sustainability Impact Startups

Abstract: This study investigates the potential of large language models (LLMs), such as ChatGPT, to aid venture capitalists in the screening of startups that maximize sustainability impact. To determine the scope that maximizes impact for venture capitalists' and to identify effective screening criteria, the study utilized theoretical research and interviews. The thesis suggests that the ideal investment space is investments into high-risk, software-centric companies contributing to a sustainable system change that maximizes outcome impact instead of optimizing for environmental, social and governance metrics. This investment space along with other defined critical success factors were then deployed in an effort to test LLMs' efficacy in targeting companies maximizing impact. Two prompting techniques were trialed, one question-based prompt where questions on critical startup success factors were asked, and another using a comparative method where the characteristics of screened startups were matched with investor profile preferences. In both versions of the model, the provision of context proved indispensable to analyze relevant startups, given GPT-4’s knowledge cut-off in 2021. Without context, the LLM often could not provide an answer or provided an imaginary one, especially for younger startups. The question-based prompting could accurately address some specific questions, while the investor profile prompt showed the most promising results by being able to efficiently summarize and present relevant output text on the given areas of interest. It was also found that the quality of the data input in the model directly affects its efficacy and it is therefore necessary to pick data carefully to avoid biases and greenwashing. This was especially true for question-based prompting, since the investor profile prompt was better at conducting an overall assessment of the companies with scarce information, but did still struggle to produce insightful ratings. In terms of the specific screening for impact startups, the model shows potential for targeting the ideal investment scope suggested by the thesis. The paper concludes by suggesting an immediate use case for the investor profile prompting technique in ChatGPT, supplemented by future use cases for automated systems to conduct outbound and inbound screening at scale.