Using AI Agents in Software Requirements

Requirements engineering (RE) plays a pivotal role in the software development lifecycle, providing the foundation for all downstream design, implementation, and validation ac-tivities. However, RE is inherently challenging due to the sheer volume and complexity of elicitation data. Fatigue, selective attention, and unconscious bias undermine results by human agents. We present a training method for Artificial Intelligence (AI) agents that systematically transforms raw elicitation data into high quality requirements without susceptibility to these limitations. With a corpus of 28 stakeholder interviews, require-ments were produced by human analysts, baseline (untrained) AI agents, prompt engi-neered AI agents, and agents trained with our method. An independent panel of experts rated completeness, consistency, traceability, and stakeholder alignment. Trained agents significantly outperformed both humans and untrained agents across all criteria, deliver-ing broader coverage and fewer conflicts while maintaining format fidelity. These results indicate that purpose trained AI agents can raise RE quality and throughput beyond cur-rent practice, offering a scalable, auditable pipeline from elicitation to specification.