Agentic AI in Process Analytical Technology: LLM Assistants for Chemometric Workflows

<jats:p>Modern manufacturing in chemical and pharmaceutical industries increasingly depends on real-time analytical insight to maintain quality, safety, and flexibility. Yet, turning process measurements into actionable analyses still depends on scarce chemometric expertise and manual workflows. Herein, we adapt agentic AI to the process analytical technology (PAT) domain by developing PAT agents that combine large language models (LLMs) with validated open-source and commercial chemometric software, including RamanSPy, SpectroChemPy and PEAXACT, under a multi-agent orchestrator. We benchmark the agents on nine real-world Raman data-analytics tasks of varying complexity, across three user personas, and several LLM backbones. The PAT agents with frontier LLMs solve analytical tasks at 93–96% pass rate, while numerical claims are grounded in actual tool outputs, and are thus physically consistent and reliable. The agents translate individual user requests into adaptive analytical workflows, halt and request clarification on inconsistent inputs, and remain reviewable through structured logs of every tool call and state transition. These results establish PAT agents as transparent digital coworkers that expand access to advanced process analytics and support more adaptive, data-rich manufacturing, while expert chemometric oversight remains critical for industrial deployment.</jats:p>