Enhancing Enterprise Data Collection through Interactive Template-Filling Agents with Localized Language Models

Structured user data gathering in enterprise environments is critical but sometimes error-prone and boring when using traditional static forms. We propose an innovative interactive template-filling agent solution based on localized compact language models (LLMs). These agents guide users through dynamic prompts, context recall, and instant feedback to complete filled-in data form templates. We show a proof-of-concept deployment of this system with a Django-based front end and a locally run LLM (e.g., Flan-T5 or LLaMA2). Comparative evaluation with static ones on an HR onboarding task revealed statistically significant gains in data completeness (95% vs. 75%), accuracy (6% vs. 18% error rate), and user satisfaction (4.4/5 vs. 3.2/5). Our solution facilitates effective, privacy-aware data gathering, applicable in HR, legal, and customer onboarding scenarios.