The Necessity for a Specialist Layer in Legal Research
The legal AI market is evolving as general purpose AI and legal workflow platforms vie to become the primary interface for legal professionals. However, substantive legal research requires a distinct layer of specialist intelligence, which is essential for grounding findings in authoritative sources, domain specific retrieval, and auditable legal reasoning.
In the realm of legal research, the most concerning AI generated response is not necessarily the blatantly incorrect one, but rather the answer that is nearly accurate. This phenomenon, often referred to as hallucination, has been extensively documented, with various repositories cataloguing instances of AI inaccuracies, including the AI Hallucination Cases database.
More subtle failures of AI systems present a significant challenge. For example, when drafting a legal research memorandum, an AI might generate a coherent summary of relevant cases but struggle with the simpler task of quoting a judgment accurately. Legal practitioners are accustomed to identifying familiar risks in work produced by junior colleagues, such as incomplete analysis or weak reasoning. However, AI failures can manifest unpredictably, leading to errors in sophisticated analysis on one occasion and basic transcription on another.
This unpredictability forces practitioners into a difficult position. They must either accept an output that appears broadly correct but may contain critical errors or verify each proposition and quotation meticulously, undermining the efficiency that AI is intended to enhance.
To address these challenges, it is crucial to differentiate between the two layers of the emerging legal AI ecosystem. While professional review is necessary across all cases, the capabilities inherent in a system significantly influence the precision of its answers, the ease of verifying outputs, and ultimately, the efficiency delivered by AI.
Traditionally, legal AI has been categorised into three groups: general purpose systems like Claude and ChatGPT, legal AI platforms such as Harvey and Legora, and specialist legal AI designed for specific fields or jurisdictions. However, this framework is becoming outdated as general purpose AI providers increasingly enter the legal domain, with notable examples including Anthropic's legal plugins for Claude and Google's Gemini Enterprise for Legal.
When evaluating AI systems for legal research, it is more beneficial to move beyond these categories and focus on two distinct layers. The roles of these layers become particularly evident in substantive legal research, which is one of the most complex challenges for legal AI.
Access to legal authorities is merely the starting point. A reliable answer must not only identify applicable sources but also differentiate between current and outdated law, understand the hierarchies and relationships among authorities, and substantiate each conclusion with verifiable sources. For instance, a general AI lacking specialist research capabilities may confidently paraphrase an obsolete statute without indicating its replacement. In contrast, specialist AI should be able to identify decisions under both the old and new legal frameworks, explain their differences, and cite the relevant authorities supporting each proposition.
These complexities are particularly pronounced in international law and arbitration, where sources are diverse, spanning multiple jurisdictions, languages, institutions, and legal traditions. Much of the pertinent material is fragmented or inconsistently structured, and general AI systems do not inherently acquire the methods necessary for effective arbitration research simply by processing more documents. Addressing these requirements necessitates specialist systems designed specifically for the sources, structures, and methodologies of the field.
For example, Jus Mundi’s arbitration agent, Jus AI, employs a highly specialised corpus alongside Tenet v5, a proprietary model developed specifically for international law and arbitration research. This model interprets legal text in conjunction with metadata and contextual information that determine an authority’s relevance, rather than relying solely on textual similarity. As a result, Jus AI can identify stronger authorities and provide more precise, verifiable citations. In an independent evaluation conducted by 20 leading arbitration experts, Jus AI achieved an average score of 4.28 out of 5 across various criteria, including reasoning, correctness, fidelity to sources, and completeness.
The effectiveness of such systems is not solely attributable to the model itself. A specialist agentic architecture is employed to define the research question, decompose it into specific legal research tasks, retrieve information from a curated corpus, validate evidence, and produce an answer grounded in the consulted sources. A dedicated quality assurance protocol further enhances the reliability of the output.
Collectively, the elements of structured data, domain specific retrieval, specialised architecture, citations, guardrails, and quality assurance constitute the intelligence layer that substantive legal research requires. These capabilities cannot be assumed to exist within a general AI platform and must be specifically developed for the relevant domain or accessed through a specialist partner.
The argument for specialist AI does not suggest that it is better suited for every legal task or that it will replace general AI platforms. The greatest value of the two layers lies in their integration, leveraging the strengths of both. Specialist AI can still be accessed independently, but the trend is increasingly towards integration with existing environments used by legal professionals.
One approach to achieving this integration is through partnerships and interoperability. Recently, Jus Mundi launched its MCP connector for Claude, following its agent to agent integration with Legora. This allows practitioners to access citation backed arbitration and international law intelligence without needing to switch contexts or compromise the specialist infrastructure behind the answers provided. As noted by Dipen Sabharwal KC, Partner at White & Case, the adoption of Jus AI enhances the efficiency of legal work, enabling lawyers to uncover deeper insights and deliver increased value to clients.
Another strategy involves incorporating specialist capabilities into general AI platforms. The acquisition of Qura by Legora exemplifies this trend, highlighting the growing recognition that general legal AI platforms require specialist research infrastructure. However, this raises an important question: can specialist legal intelligence be seamlessly integrated into a general platform, or does its reliability hinge on maintaining the distinct data, methods, and quality controls that define its specialist nature?
Regardless of the integration model that emerges, the critical issue remains whether the integrity of the specialist layer is preserved. While interfaces may evolve, the foundation of reliable legal research will continue to rely on authoritative data, specialist retrieval methods, and sources that lawyers can verify.