A Systematic Literature Review of AI-Powered Legal and Financial Assistance Systems: Approaches, Challenges, and Future Directions

Abstract The rapid evolution of Artificial Intelligence (AI), particularly Large Language Models (LLMs) and transformer-based neural architectures, has significantly transformed knowledge intensive domains such as law and finance. Recent developments in domain-adapted language models, Retrieval-Augmented Generation (RAG), and hybrid reasoning frameworks have enabled intelligent assistance systems capable of supporting legal research, financial literacy, compliance analysis, and customer interaction. This systematic literature review analyzes research papers published between 2020 and 2025, following PRISMA guidelines, to examine the current landscape of AI-powered legal and financial assistance systems. The review categorizes major technical methodologies including domain-specific fine-tuning, retrieval-enhanced architectures, and explain ability mechanisms, while identifying persistent limitations such as hallucination, weak numerical reasoning, multilingual challenges, and trust deficits that hinder large-scale deployment. Special emphasis is placed on multilingual and dynamically evolving regulatory environments such as India. The findings indicate that while significant progress has been achieved in factual grounding and conversational intelligence, future research must prioritize transparency, accountability, and secure deployment in high-stakes domains Keywords: Legal AI, Financial Chat bots, Retrieval-Augmented Generation, Explainable AI (XAI), Large Language Models, Domain-Specific NLP, Security, Indian Legal System 1. Introduction The integration of Artificial Intelligence into highly specialized professional domains such as law and finance has emerged as one of the most significant technological advancements of the past decade. Legal and financial systems involve complex terminology, extensive document citation, regulatory dependencies, and high accountability requirements, making them ideal yet challenging candidates for intelligent automation. Traditional software systems and rule-based expert systems often lacked adaptability and struggled to interpret context-dependent queries, limiting their practical usability in real-world professional environments. The introduction of transformer-based architectures, particularly models such as GPT-3, GPT-4, BERT, and domain-specific adaptations like LEGAL-BERT [15] and FinBERT, has substantially improved the ability of machines to understand and generate domain-specific text. These models demonstrate strong contextual learning capabilities and have enabled applications including legal document summarization, judgment prediction [7], contract analysis, financial advisory systems [13], fraud detection, and automated customer support [11]. Despite these advancements, significant challenges remain unresolved. Large Language Models frequently generate hallucinated responses [2], misinterpret legal statutes, or produce incorrect numerical outputs in financial reasoning tasks. In countries such as India, additional complications arise due to multilingual communication, inconsistent legal formatting, rapidly changing laws, and limited access to authoritative digital datasets [3]. Consequently, ensuring reliability, transparency, explainability, and security has become a major focus of current research efforts [12, 10]. The growing dependence on AI systems in high-stakes environments necessitates rigorous evaluation of their technical capabilities, ethical implications, and deployment readiness [1]. This review therefore aims to provide a comprehensive understanding of existing approaches, limitations, and emerging research directions in AI-powered legal and financial assistance systems. 2. Literature Survey This systematic literature review focuses on analyzing research contributions published between 2020 and 2025 concerning AI systems developed for legal and financial assistance applications. The review specifically investigates transformer-based architectures, retrieval enhanced systems, explainability frameworks, and security-oriented deployment mechanisms. Methodologies Algorithms Dataset Limitations Human-Centered Design, Design Thinking, Legal Design [1] None No data set Conceptual only; lacks empirical validation Empirical evaluation of LLM outputs [2] ChatGPT-4, PaLM2,LLaMA 2 U.S. federal court cases High hallucination and unreliable out puts RAG,OCR,NER, semanti cretrieval [3] FAISS, DeepSeek-R1-5B Indian Constitution and statutes Issues with multijurisdiction queries NLP,ML,Informa tionRetrieval [4] TF-IDF, Naive Bayes Indian legal judgments Limited scalability and ambiguity handling NLP and chatbot frameworks [5] Seq2Seq, LSTM, Word2Vec Indian legal FAQs Cannot provide personalized advice Conversational AI and speech processing [6] Rule-based matching, GPT-4 API Legal documents and DOJ records Dialect and response-length issues Machine Learning and NLP [7] Random Forest, CNN, BiLSTM Chinese civil case judgments Limited to Chinese jurisdiction Survey and literature review [8] Transformers, LSTMs, CNNs Git Hubre positories Noempiricalval idation Benchmark creation and evaluation [9] GPT-4, Claude, Falcon Legal Bench dataset English-only benchmark Legal and regulatory analysis [10] AI Act, GDPR frameworks Conceptual legal study Lacks empirical validation Responsible AIen gineering [11] Responsible AI patterns IBM Watson datasets Limited quantitativee valuation NLP and risk analysis [12] LSTM, CNN, Random Forest Financial advice documents Subjective regulatory interpretation Comparative study of financial LLMs [13] GPT-4,Bard Financial advisory queries Poor numerical reasoning Few-shot learning and RAG [14] QLoRA, Retrieval Augmentation Real-world credit datasets Optimization challenges Transformer retrieval Limited multilingual coverage and fine-tuning [15] Llama 3.1 8B, RAG, FAISS Lawyer-GPT dataset Limited multilingual coverage 3. Methods There are various methods of approaches that combine quantitative analysis with qualitative investigation which offer AI’s effects. Further studies incorporate predominant, longitudinal research track of AI’s evolution. over time, allowing for more robust conclusions about its long-term effects on the financial sector. Some of the methodologies adopted are mentioned below. 3.1. PRISMA-Based Review Protocol The systematic review was conducted using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology to ensure transparency, reproducibility, and methodological consistency. PRISMA provides a structured framework for identifying, screening, and synthesizing relevant research studies. The review process involved multiple stages including database search, duplicate removal, title and abstract screening, full-text evaluation, and final inclusion analysis. The primary objective was to collect high-quality evidence related to transformer-based legal and financial AI systems developed in recent years. 3.2. Search Strategy and Data Sources Research papers were collected from major academic databases and digital libraries including IEEE Xplore, ACM Digital Library, Springer Link, Elsevier Science Direct, and arXiv. Search keywords included terms such as “Legal AI”, “Financial Chatbot”, “Retrieval-Augmented Generation”, “Explainable AI”, “Indian Legal NLP”, and “Transformer-based Legal Systems”. The search period was restricted to publications between 2020 and 2025 because this period corresponds to the widespread adoption of transformer architectures and Large Language Models. 3.3. Inclusion and Exclusion Criteria Studies included in this review satisfied the following conditions: • Published in English language • Focused on legal or financial AI applications • Used transformer-based, LLM-based, or hybrid AI methods • Included empirical evaluation or implementation details Studies were excluded if they: • Focused only on generic conversational chatbots • Used outdated pre-transformer methods • Lacked evaluation metrics or methodological clarity 3.4. Search Findings The final review included thematic areas including Legal AI Systems, Financial AI Systems, Explain ability Frameworks, Retrieval-Augmented Generation architectures, and Security-Oriented AI deployment mechanisms. Publication trends revealed a gradual shift in research focus over time. Early studies between 2020 and 2021 concentrated primarily on transformer baselines and domain adaptation strategies [8, 1]. During 2022 and 2023, Retrieval-Augmented Generation emerged as a dominant architecture for reducing hallucination and improving factual grounding [4, 11, 13, 7]. Recent works from 2024 and 2025 increasingly emphasize explain ability, governance, multilingual adaptation, and trust-aware deployment [2, 9, 5, 10, 3, 6, 14]. The review also identified strong geographical concentration in research publications, with most studies focusing on the United States, Europe, and China [7, 9]. Comparatively fewer studies addressed multilingual and jurisdiction-specific challenges relevant to India [3, 4, 5], indicating a substantial research gap 3.5. Legal AI Systems Legal AI systems represent one of the most rapidly developing areas within applied Natural Language Processing. These systems aim to automate or assist tasks such as legal research, statute retrieval, case summarization, judgment prediction [7], and legal question answering [9]. Domain-specific transformer models such as LEGAL-BERT [15-18] and Indian Legal BERT have demonstrated improved performance by training on legal corpora instead of general-purpose text. These systems capture legal terminology, procedural language, and citation structures more effectively than standard language models. However, legal reasoning r