Micro, Small, and Medium Enterprises (MSMEs) frequently encounter excessive compliance challenges stemming from disjointed recordkeeping, changing tax laws, and restricted access to professional expertise. This study introduces a Smart Tax Filing Assistant, a web-based AI solution aimed at automating document ingestion, tax entity extraction, and compliance verification, incorporating human oversight in the review process. The system architecture combines OCR-based preprocessing, transformer-based Named Entity Recognition (NER), classification for expense categories, and a deterministic rule engine that connects to a machine-readable tax knowledge base. We used a dataset of 10,200 fake and anonymous financial documents, such as invoices, receipts, and bank statements, to do the evaluation. The results showed that extraction F1 scores were over 92% for important fields and that the average processing time per document was 72% shorter than manual filing. Compliance Confidence Scores (CCS) made reliability even better by combining model probability, rule coverage, and checks for consistency. The results show that using explainable AI with secure data sharing makes it much easier, more accurate, and more open for MSMEs to pay their taxes. This framework provides a scalable model for AI-enabled tax administration, with the potential for wider implementation in emerging economies.