As software systems become increasingly complex, automation in testing must evolve beyond execution to include test planning, analysis, design, and implementation. While effective, traditional testing methods often struggle with requirements intake, inefficient test case design and selection, script maintenance challenges, and result analysis mistakes. Could Large Language Models (LLMs), lately extended with the support of AI Agents, help bridge these gaps by automating tasks prone to human error and enhancing the efficiency of software testing activities? LLMs introduce new possibilities, enabling us to transform vague requirements into reasonable models, generate optimized test cases, and adapt test scripts to evolving conditions. As AI advances, its capabilities and limitations force us to rethink established testing methodologies. However, innovations are usually followed by challenges—questions of trust, correctness, interpretability, and ethical concerns remain key obstacles to widespread adoption. This talk will delve into the real-world impact of LLMs in software testing, cutting through the hype to critically examine their strengths and limitations. While LLMs offer intriguing possibilities, they also introduce uncertainties: How reliable are their outputs? Is new technology mature enough to be applied with low-level risks? Can they truly replace human intuition in critical testing scenarios? Let us explore several use cases for blending LLMs into testing process, highlighting successes and unexpected failures, and discuss the risks of over-reliance on LLMs.