Generative artificial intelligence is rapidly entering the humanities, yet its integration into established research infrastructures remains insufficiently examined. This paper presents a case study on embedding large language models (LLMs) into lexicographic workflows for semantic classification. The project implements an AI-supported system for assigning lexical meanings to a controlled semantic framework used in German dialect lexicography. In a controlled evaluation of 1,000 meanings, model-generated suggestions achieved an adoption rate of over 90%, including additional proposals judged equally valid by experienced editors of approximately 10%. On this basis, a web service was developed and integrated into the editorial environments of three dialect dictionary projects.

The paper moves beyond performance metrics to analyze what happens when generative AI (GenAI) becomes part of an operational humanities infrastructure by evaluating the LLM-based process in a ongoing study that will be finalized in summer. Within a hybrid human-AI workflow, the system generates up to three classification suggestions per entry, while editorial authority remains fully intact. The integration has measurably reduced search effort, stabilized decision-making, and enabled the scalable structuring of large legacy datasets comprising more than 40,000 meanings.

At the same time, the study addresses the epistemic consequences of algorithmic assistance. When model suggestions function as recurring reference points, subtle standardization effects and forms of emerging algorithmic authority may arise. A structured user study evaluating acceptance, efficiency gains, and perceived risks is currently underway, with full results to be presented at the conference. The findings suggest that sustainable AI integration in the Digital Humanities depends not on automation alone, but on carefully designed hybrid systems that combine efficiency gains with transparent governance and continuous critical oversight.