Skip to content

AI engineering case study

Private Legal Review Copilot

Core stack

RAG · GDPR · Legal Tech · Semantic Search

Delivery approach

  1. Situation

    Legal teams working with Swiss documents spent time locating clauses, comparing language, summarizing obligations, and preparing first-pass answers. General chat tools lacked document boundaries, reliable citations, retention controls, and a clear distinction between source language and generated text. In legal review, an unsupported answer creates more work and risk than a stated lack of evidence.

  2. Task

    Design a private legal copilot that reduced routine contract-review time and tied every factual answer to the active document set. It needed GDPR-aligned data handling, isolation by tenant and matter, visible source context, and conservative behavior when evidence was missing. The product would support professional review, not provide unsupervised legal advice.

  3. Action

    The trust model came first: scoped document access, encryption in transit and at rest, explicit retention, auditable processing events, and separation between matters. Ingestion preserves page, section, clause, version, language, and effective date. Semantic and lexical retrieval run together so exact legal terms are not lost through embedding similarity. Requests are classified as extraction, comparison, summarization, question answering, or drafting, with a source policy for each mode. Quoted clauses remain visually separate from generated explanations, and citations open the relevant location. Unsupported questions return a limitation or ask for clarification. Draft language is labeled as a proposal and is never shown as source text. Reviewer-authored questions test clause completeness, citations, conflicting versions, cross-document comparison, and refusal behavior. Operations track retrieval gaps, citation opens, latency, access denials, and processing failures. Prompts are versioned and can be rolled back.

  4. Result

    Review turnaround fell by 45%. Lawyers and operations staff can open the clauses behind an explanation without losing the document trail. Source text, generated interpretation, and proposed drafting remain separate, so a reviewer can verify each part before applying legal judgment.

Technology stack

  • RAG
  • GDPR
  • Legal Tech
  • Semantic Search