According to an Anthropic study from March 2026, AI models could theoretically perform nearly 90% of legal tasks, while actual observed usage remained at approximately 20%. A September 2026 survey by the International Legal Technology Association (ILTA) found that 94% of over 500 law firms were testing or using generative AI, up from 80% the previous year, though full-scale deployment remains rare.
In Russia, an Avito and Pravo.ru study from December 2025 involving over 500 legal professionals showed that 88% of respondents use neural networks. Key barriers include concerns over information accuracy for 84% of users and the necessity of manual verification for 80%. These challenges persist because controlling the reasoning path of AI models in autonomous workflows remains difficult.
Researchers from Fudan University, Ant Group, and the Beijing University of Posts and Telecommunications (BUPT) introduced the TG-RAG (Thought Guidance – Retrieval-Augmented Generation) architecture at the ICML 2026 conference. The method aims to mitigate cognitive drift in complex reasoning tasks by formalizing standard operating procedures into graphs.
The TG-RAG architecture separates declarative knowledge from procedural knowledge. It executes procedures step-by-step, interrupting generation at analytical boundaries to insert specific directives and relevant data. This approach ensures that the model follows a predefined expert protocol, which is critical for multi-step legal tasks like due diligence.
