We introduce DoseAI, an online-updateable causal AI model for selecting optimal dynamic treatment regimes under the constraint of maximally tolerable side effects. DoseAI is trained on observational disease course data and accounts for time-dependent confounding, where time-varying variables influence treatment decisions and bias effect estimates. To address this, DoseAI minimizes the absolute Spearman correlation between predicted and observed future dosages as a proxy for distributional similarity. Using simulated lung cancer data, including chemotherapy and radiotherapy effects on tumor volume and body weight, DoseAI effectively reduces tumor size while typically adhering to toxicity limits.