Accurate time-series forecasting is essential for real-world applications such as predictive maintenance and feedback control. While deep neural networks have shown promise in recognizing complex patterns and predicting trends, their generalization capabilities are open to debate, and they typically do not perform well with limited data. In this paper, we examine the potential of time-series foundation models (TSFM) as a practical solution for addressing real-world (probabilistic) forecasting challenges. Our experiments using real building data demonstrate that, through fine-tuning TSFMs, we can achieve excellent predictions, even with limited data, and improve generalization in zero-shot prediction on unseen tasks.