As the digital transformation of enterprises accelerates, traditional artificial intelligence algorithms face core problems such as insufficient model adaptability, low efficiency in processing unstructured data, and weak cross-departmental data fusion capabilities in dynamic business scenarios, resulting in unsatisfactory intelligent application effects and difficulty in quantifying ROI. This study aims to develop a dynamic optimization algorithm framework based on meta-learning and reinforcement learning, build a multi-dimensional application effectiveness evaluation system, and design a three-stage optimization mechanism of “dynamic parameter adjustment-heterogeneous data fusion-business closed-loop feedback” (First, a meta-learning algorithm is used to establish a priori knowledge base, and the business flow characteristics are captured in real time through the LSTM network. The hyperparameters are dynamically adjusted in combination with the Q-learning algorithm. Secondly, the improved Transformer architecture is used to process multimodal data such as text and images, and the graph neural network is introduced to realize cross-departmental data topology modeling. Finally, an evaluation model with 12 indicators including operating costs, customer satisfaction, and market response speed is established, and the analytic hierarchy process and entropy weight method are used to combine weighting). The measured data of 15 companies in the three major industries of manufacturing, finance, and retail show that the algorithm increases the accuracy of production forecasts to 92.1% (±1.2%), increases the supply chain response efficiency by 4.3 times, and shortens the customer complaint processing time to 1.32 h. The algorithm framework effectively solves the problems of dynamic environment adaptability and cross-domain data fusion.