A Structured Review of AI-Assisted Requirements Engineering: Capabilities, Limitations, and Research Gaps

Abstract

The software industry considers that Large Language Models (LLMs) will soon automate the core tasks of Requirements Engineering (RE). Current literature proves this assumption is premature. This paper reviews 25 recent studies to document exactly what AI can and cannot do during the software specification phase. Current studies split the engineering workflow into elicitation, analysis, generation, and traceability. The reviewed data suggests that AI is highly effective at basic pattern recognition and drafting initial text. this literature shows that these same tools consistently fail when processing ambiguous stakeholder language, lose project context during extended interactions, and frequently hallucinate invalid system requirements. The evidence we discovered shows that AI currently functions strictly as a supplementary tool in RE, not an autonomous replacement for human engineers. The paper concludes that future research must prioritize human-in-the-loop validation frameworks rather than focusing purely on automation.