The automation of tasks via Graphical User Interface (GUI) agents represents a major breakthrough in artificial intelligence. However, most current approaches rely on general-purpose Large Vision-Language Models (LVLMs), which are costly and often oversized for the fundamental task of visual perception. This work explores an alternative, adopting the viewpoint that smaller, specialized models (SLMs) represent a promising direction for agentic AI. We propose to tackle the challenge of visual instruction grounding—locating a UI element from a natural language command and a screenshot—through two specialized approaches. The first, SICocr, is a modular architecture combining an optimized object detector, Optical Character Recognition (OCR), and an LLM for semantic interpretation. The second, SICdirect, is an end-to-end perception model specifically fine-tuned for this task. To evaluate these approaches, two new dedicated datasets were constructed. Experimental results indicate that the end-to-end specialized approach, SICdirect, achieves higher performance than the modular approach and SeeClick, which serves as the reference state-of-the-art method.