Email has long been used as a secure and official mode of communication. In order to stop cyberattacks and crimes, we must examine the emails that individuals and organizations use to communicate. Any sort of inciting and unwanted digital communication transmitted in bulk that may contain objectionable content, such as viruses and cyberattacks, is considered a type of digital attack. More precise and powerful anti-spam filters based on artificial intelligence have to be developed due to the frequent increases in spam emails. The current keyword-based search methods and filters frequently produce irrelevant emails with few useful words. The bidirectional encoder representations from transformers (BERT) model is used to extract the text’s underlying features in order to get around the aforementioned limitation. We utilize this model to classify emails into many categories in order to improve accuracy. It divides emails into four categories: normal, harassing, suspicious, and fraudulent.