It's Not the AI,... It's the Way You Use It: Making LLMs work in the Real World

Abstract

Pharmaceutical deviation and corrective and preventive action (CAPA) processes are broken—plagued by rushed closured, “human error” cliché and data silos. Can large language models (LLMs) fix this, or will they just automate bad habits? Based on expert interviews, this article reimagines LLMs not as decision makers but as devil’s advocates—tools that provoke better thinking, expose bias and deepen investigations. The proposed five-step roadmap demands more than tech: strong data governance, human oversight, iterative validation, small-scale pilots and a cultural shift toward learning. The EU’s draft Annex 22 on Artificial Intelligence (AI), released in July 2025 for public consultation, adds timely pressure—demanding traceability and accountability but risking innovation paralysis if applied too rigidly. The takeaway: LLMs could be transformative, but only if quality leaders have the courage to use them as catalysts for change—not as digital box-tickers in the same old broken system.