International development institutions, United Nations agencies, and non-governmental organizations (NGOs) generate vast quantities of unstructured qualitative data, including narrative reports, field interviews, focus group summaries, and policy documents. Despite the significant financial investments in producing this data—NGOs alone disbursing over US23 billion annually—its potential remains underutilized due to the lack of scalable analytical methods. Traditional qualitative analysis approaches, such as content analysis, thematic analysis, narrative analysis, discourse analysis, and grounded theory, are resource-intensive, time-consuming, and often unsustainable at the scale required. Artificial Intelligence (AI) offers a promising alternative by enabling automated or semi-automated analysis of large datasets, enhancing efficiency, reliability, and the breadth of insights generated. This literature review examines the application of AI in analyzing unstructured qualitative data, focusing on its use in healthcare, social sciences, and policy research. AI techniques, including natural language processing (NLP), sentiment analysis, clustering, topic modeling, and pattern recognition, have demonstrated potential to replicate human-led content analysis outcomes in significantly reduced timeframes. For instance, in healthcare research, AI has been used to analyze chronic pain narratives and COVID-19-related qualitative datasets, often achieving results comparable to human analysis but with enhanced speed. Similarly, in education and social science contexts, AI tools such as Python’s SpaCy have processed thousands of open-ended survey responses, enabling meaningful interpretation at scale. The review identifies several advantages of AI-enabled analysis. These include substantial time and cost savings, reduced analyst fatigue, improved coding accuracy, and the ability to conduct more comprehensive literature reviews and thematic explorations. AI also supports iterative and reflexive analysis, expanding the possibilities of theory development and data visualization. Such capabilities make AI particularly attractive to institutions with limited analytical resources yet large volumes of qualitative data. Nonetheless, limitations persist. AI struggles with nuanced interpretation, including detecting sarcasm, cultural references, and subtle thematic variations. Ethical concerns—ranging from data privacy to algorithmic bias—remain critical considerations. Reliability issues, “hallucinations” in generative AI outputs, and reduced interpretive depth compared to traditional qualitative methods underscore the need for cautious application. Studies suggest that deductive AI setups may outperform inductive approaches, but human reflexivity, contextual knowledge, and interpretive judgment remain indispensable. A recurring theme across the literature is the advocacy for hybrid analytical models where AI augments rather than replaces human expertise. In such models, AI serves as a rapid pattern-detection and preliminary coding tool, allowing human analysts to focus on higher-order interpretation and critical inquiry. This approach can maximize efficiency while preserving the depth and contextual sensitivity that define rigorous qualitative research. The paper concludes that AI holds transformative potential for the analysis of unstructured qualitative data, particularly for organizations in development, health, and policy sectors. Realizing this potential requires sustained collaboration between institutions, academia, and the private sector to refine methodologies, test across diverse datasets, and address ethical and reliability challenges. Furthermore, generational differences in AI adoption—where younger researchers are more inclined toward AI use—suggest opportunities for targeted capacity-building initiatives. By strategically integrating AI into qualitative workflows, stakeholders can unlock greater value from existing data, inform decision-making, and enhance the inclusivity and scope of research insights.
Kalanda et al. (Fri,) studied this question.