Supreme Audit Institutions (SAIs) are crucial for transparency in public financial management but face major challenges when manually processing large volumes of multilingual, semi-structured, and heterogeneous financial documents. This paper reviews recent advances in multimodal Transformers and OCR-free models for document artificial intelligence and proposes a modular and reproducible pipeline for financial document analysis in African SAI contexts. The pipeline integrates layout-aware models (LayoutLMv3, DocFormer) and OCR-free architectures (Donut) and is validated on public benchmarks and a proprietary Francophone dataset (NOVAFRIQ). Experimental results show up to 93.7% F1-score for information extraction and 95.2% accuracy for document classification. Pilot simulations indicate potential processing time reductions of 60–75% and error decreases of approximately 40%. The study demonstrates a promising approach for building robust, explainable, and multilingual AI-assisted auditing tools.