A virtual assistant (VA) is an artificial intelligence (AI) powered software agent that facilitates assistance to a user by performing a task or answering questions through text or voice interactions. As we have virtual assistants to answer us, the necessity of accurate IC and NER has increased so that it can reply according to the query posted by the user. To improve the performance of VA, here is a proposed hybrid model that combines both bidirectional encoder representation from transformers (BERT) and long short-term memory (LSTM). Natural language understanding is important feature of VA which requires IC and NER, and the hybrid architecture integrates the strengths of both BERT and LSTM to enhance these functions. To capture the semantic meaning of incoming text from user, BERT offers deep embedded contexts, whereas LSTM assist in describing long-term temporal patterns and sequential dependencies. Proposed hybrid model performs better compared to individual BERT and LSTM models, having higher accuracy in IC and NER tasks. BERT-LSTM model achieves an IC accuracy of 99.5% with a precision of 99.7% and an NER accuracy of 95.07% with a precision of 97.8% when trained and tested on a dataset consisting of real and synthesized user queries. These outcomes show the potential of model to handle complex user inputs more accurately compared to individual BERT and LSTM models. This research provides valuable insights into the potential of mixing transformer-based and recurrent neural network architectures to significantly improve VA capabilities, with exciting prospects for applications in real-world intelligent systems.