The Arabic language is widely used worldwide by hundreds of millions of users. This research presents a novel approach to Arabic handwritten word recognition and classification leveraging the power of autoencoder deep learning architectures. Our study is underpinned by creating a proprietary database yielding a substantial corpus of 3000 handwritten instances. Our investigation aims to assess the efficacy of autoencoder networks in accurately deciphering and categorizing handwritten Arabic script, a task renowned for its intricacies and challenges. Experimental results demonstrate the ability to attain an exceptional recognition and classification accuracy of 96.56% on our dataset. This achievement underscores the potency of autoencoder deep learning models in tackling Arabic handwriting recognition tasks, reaffirming their applicability in diverse real-world scenarios, such as document analysis, authentication systems, and linguistic research.