The transformative AI-paralegal for better justice

The most critical factor in winning a legal case? Strong arguments based on the right legal sources. Historically, reaching a definitive conclusion required a manual, multi-step crawl: starting from legislation, moving to preparatory works, analyzing case law, and consulting legal literature. Even for straightforward issues, this process typically takes at least 30 minutes.

The Edilex service is a crucial tool for Finnish legal professionals, consolidating hundreds of thousands of legal documents. Still, the search phase remained a hurdle. Users still had to identify the exact right keywords, open multiple tabs, and manually vet voluminous text to find a single relevant passage – taking valuable time away from customer work. On top of this data-heavy research, professionals then had to manually draft summaries and citations for their clients, compare legal cases, and summarize past and future legislative reforms. To solve this, the Alma Media and Vincit teams worked together to research how AI could streamline the process.

Key business drivers:

• Transforming legal research: Shifting the profession from manual retrieval to an "ask-and-verify" model, trivializing time-consuming tasks to reclaim research time.
• Commercializing trust: Differentiating Edilex by pivoting from creative AI to a closed-environment AI-paralegal – monetizing precision and hallucination-proof reliability.
• Business growth: Leveraging the AI-paralegal as a high-value differentiator to drive subscription growth and justify the migration of users to premium service tiers.
   

International legal firms have spent millions training proprietary language models, often with mixed results. Our strategy was built on a bold provocation: Could we achieve elite-level legal AI through agility and data orchestration rather than raw model training? We built a proof of concept (PoC) to explore how Generative AI could be harnessed to maximize legal efficiency – and ultimately ensure better justice for Finnish citizens.

The initial pilot revealed that generic Large Language Models (LLMs) struggled with the nuances of Finnish legal terminology, leading to hallucinations that could jeopardize a lawyer’s work. This steered our strategy to verified retrieval in a closed environment. We chose a model-agnostic strategy that prioritized the structural integrity of the Finnish legal database over the "creativity" of the AI.

The PoC phase was a stress test for Finnish legal data: we validated that the AI could locate precise sources from a high-stakes subset of laws and Supreme Court decisions using custom embeddings and vector databases.

1. Semantic data chunking: We utilized the structured hierarchy of Edilex data, ensuring the AI understands the difference between a binding law and a guiding precedent.
2. Search precision vs. scale: As the document mass grows, so does the risk of "noise." Our architecture is optimized to maintain pinpoint search accuracy, which drives both user satisfaction and cost-efficiency for Alma.
3. Hallucination-proofing: The AI acts only as a synthesizer of retrieved text. If the answer isn't in the database, the system doesn’t return second guesses.
4. Model-agnostic agility: We built a flexible engine that can swap LLMs (GPT-4o, etc.) on the fly, ensuring the tool remains state-of-the-art as tech evolves, while keeping development on time and on budget.