Navigating the Risks of AI in Management Consulting - A Qualitative and Experimental Study of AI Hallucinations, Task Suitability, and Mitigation Strategies

Abstract

Large Language Models (LLMs) have been rapidly adopted not only by individual users, but also in professional settings, particularly within knowledge-intensive industries such as management consulting. The integration of these tools into professional practice has created new opportunities for efficiency, while also introducing significant risks. Most notably is the phenomenon of AI hallucinations, where models generate factually incorrect or unverifiable content with apparent confidence. While existing research has primarily examined hallucinations from a technical perspective, little attention has been given to how they manifest and are managed in real-world professional settings. Therefore, the purpose of this study is to explore how AI hallucinations affect small-sized management consulting firms, examining both the strategies consultants use to manage LLM-generated errors and which consulting tasks are most vulnerable to them. Using a qualitative, abductive research design, the study combines nine semi-structured interviews with management consultants and systematic experimental testing of Microsoft Copilot across four consulting tasks: exploratory research, client communication, summarization and interview synthesis. The findings reveal that error management in small consulting firms is largely informal, driven by individual professional judgement rather than formalized governance structures. Consultants rely on domain expertise, pattern recognition, cross-source verification and selective task allocation as their primary safeguards. Human oversight emerges as the central and non-negotiable control mechanism across all firms. The experimental component suggests that the task of interview synthesis was identified as the most vulnerable task, followed by summarization, exploratory research and client communication. Hallucinations were most severe and difficult to detect in tasks requiring interpretation, prioritization and narrative construction from ambiguous input data.