The Open Radio Access Network (O-RAN) introduces modular, intelligent RAN control through the RAN Intelligent Controller (RIC), enabling real-time network optimization via xApps and rApps. This paper presents a case study on applying deep reinforcement learning (DRL) to the traffic steering use case within the O-RAN framework. We design a DRL-based agent that learns optimal user association policies to minimize end-to-end transmission latency. The training process is offline in a TS Training rApp deployed on the Non-RT RIC, leveraging historical data and model optimization tools. The trained actor network is then deployed in the Near-RT RIC as a TS xApp for real-time inference. Evaluation results show that the DRL-based xApp outperforms conventional strategies under environmental change, achieving near-optimal performance with significantly lower latency. These findings highlight the feasibility and advantages of integrating DRL into O-RAN to enable intelligent and adaptive RAN control.