The long-term performance deterioration of bridges reduces their load-bearing capacity, leading to accidents, making maintenance essential for safe operation. However, government reports from the USA and China indicate significant investment shortfalls in bridge maintenance, each exceeding hundreds of billion. Balancing structural risk and minimizing maintenance costs is a major challenge. Traditional bridge management relies on expert judgment based on current inspection reports, which fails to account for historical data and future deterioration. To address this, a more intelligent agent is needed to optimize inspection timing, preventive maintenance, and rehabilitation using infrastructural information. This study employs the Actor-Critic algorithm from deep reinforcement learning to process structural data and generate maintenance actions. Training the intelligent agent requires experience in various scenarios to provide optimal maintenance actions for each situation. This involves creating a virtual environment based on digital twins’ technology. The agent's training leverages the Proximal Policy Optimization (PPO) method to interact with the virtual environment and learn effective management policies. A practical application of this approach is demonstrated through a comparative study using an actual steel pipe arch bridge. The study benchmarks the performance of the intelligent agent against traditional maintenance strategies. The findings reveal that both policies can ensure the safety of the bridge, but the intelligent agent can reduce inspection costs by 75% and rehabilitation costs by 15%. The major functionalities of the digital twins are displayed in the accompanying video: https://www.youtube.com/watch?v=0PGvA9ELwj0