Aurora (“Dawn”) is an AI ecosystem developed by the Public Prosecution Service of Rio Grande do Sul (Brazil) to transform justice administration. Active in judicial and extrajudicial domains, it supports prosecutors protect vulnerable groups and collective rights by analyzing large volumes of documents and audiovisual records and integrating AI into legal workflows. With a human-in-the-loop approach, it enhances institutional capacity while preserving accountability and professional judgment.
Initiative overview
Aurora emerged as an institutional effort to mark a "new Dawn" in legal work, treating technology as a strategic teammate rather than a standalone tool. It was developed to address increasing procedural complexity, data fragmentation, and cognitive overload within the justice system of Rio Grande do Sul. Prosecutors must analyse large volumes of documents, reports, and recorded hearings across multiple systems, which limits timely action in areas such as criminal prosecution, child protection, environmental enforcement, and public integrity. The initiative seeks to strengthen institutional capacity by embedding AI as a structured and accountable capability within core justice functions.
Rather than focusing solely on document automation, Aurora operates as a cognitive support infrastructure. It enables natural language interaction with procedural records, automated transcription and summarisation of audiovisual evidence, cross-referencing of testimonies across different hearings, and identification of potential inconsistencies. Integration with existing judicial back-office systems (including eProc and SAJ) reduces workflow disruption and eliminates mechanical bottlenecks from case intake to judicial submission. Aurora evolved through an internal prototyping phase in which prosecutors tested and refined prompts and transcription workflows before full deployment. This ensured alignment with operational needs and encouraged voluntary adoption. Governance is anchored in institutional policy and ICT oversight, including human-in-the-loop review, audit trails, and structured prompt curation. A collaborative prompt library enables users to share and evaluate analytical instructions, which are subsequently curated by a central body that may assign an official “Approved” designation. This transforms individual expertise into institutional knowledge while preserving intellectual autonomy.
As of February 2026, Aurora has processed 13,012 hours of audio and video and supported the production of 37,028 legal documents. On February 11, 2026, it reached a historic milestone of 1,000 documents produced in a single day. System analytics indicate 1,432 unique users, reflecting strong voluntary uptake. Results are monitored through usage data and workflow indicators. In time-sensitive cases, such as sexual abuse testimonies, faster evidence analysis reduces delays and strengthens safeguards. Future impacts include expanded integration and analytical capacity to improve access to justice.
Other relevant details
Key challenges included integrating AI capabilities with legacy justice systems, ensuring data protection and legal accountability, and overcoming resistance from previous low-adoption automation initiatives. Aurora addressed these risks by prioritising usability, transparency and voluntary adoption, and by embedding AI tools within existing workflows rather than replacing them. Progressive integration with core back-office systems and human-in-the-loop review strengthened trust and auditability. Avoiding excessive standardisation of legal reasoning was mitigated through a user-governed prompt framework that enables shared learning while preserving professional autonomy. The main lesson is that successful AI adoption in justice institutions depends on trust, governance, policy alignment and demonstrable value, supported by simple UX and organic uptake rather than mandatory deployment.
Aurora’s governance is anchored in a public procurement process that evaluated over 600 technical, legal and ethical criteria. To ensure institutional trust from the outset, only anonymised datasets were used during vendor selection and proof-of-concept stages. The system employs grounding techniques to ensure that AI outputs are strictly derived from case-specific documents and metadata, minimising hallucinations and supporting traceability. A central governance body validates shared prompts through an institutional approval process, while safety settings filter sensitive criminal evidence in accordance with legal and ethical standards. All outputs remain subject to human review, ensuring that final determinations stay under prosecutorial authority. Data is stored in isolated secure environments in full compliance with Brazil’s General Data Protection Law (LGPD). Aurora’s future roadmap focuses on transitioning from task-based support to a more agentic architecture, enabling AI agents to assist with procedural triage, scheduling, and certifications within clearly defined governance parameters. This evolution is intended to reduce administrative burden and allow professionals to focus on legal strategy and the protection of collective rights. While the collaborative prompt library is internal, the functional prototypes that originated it can be adapted by other public institutions seeking to adopt accountable and scalable AI solutions. This approach supports a sustainable pathway for institutional AI adoption aligned with public sector values and operational constraints.
