From December onwards, the (National Council of Justice) will test an AI () that can shorten the average time for analyzing opinions in legal actions by up to ten times.
But even with national expansion planned until 2027, the model — created by with support from Web Services (AWS) — has no budget forecast for continuity after the end of big tech’s support for the pilot phase of the project.
The company provided around US$350,000 (equivalent to R$1.9 million at current exchange rates) in computing and infrastructure credits, used in training and initial testing of the tool.
The agreement was signed between the CNJ and InovaHC, the innovation center at USP’s Hospital das Clínicas, which in turn delegated responsibility for developing the model to the Institute of Mathematics and Statistics (IME-USP).
The partnership’s cooperation agreement envisages automating up to 80% of the screening of health actions, reducing manual administrative tasks by 80% and centralizing 80% of legal demands on a single platform by August 2027.
In an interface similar to a chat, the judge can ask, for example, whether a certain medicine is indicated for a disease. He will receive the technical and legal information available on the topic.
Today, this type of analysis takes, on average, 20 days. The goal is to reduce the deadline to up to 48 hours, according to IME professor João Eduardo Ferreira.
“The two days would be for the most conflictual, complex cases. Otherwise, the expectation is something almost immediate”, he says.
The AI will use data from e-NatJus, a CNJ platform that gathers technical notes from the SUS to support judicial decisions.
Data from the CNJ’s Justice in Numbers Panel show that the volume of new health-related lawsuits has been growing steadily in recent years. There were 352 thousand cases in 2020, 406 thousand in 2021, 470 thousand in 2022 and 577 thousand in 2023. In 2024, the total reached 690 thousand cases. By September 2025, there were already more than 513 thousand new processes.
The CNJ chose the Santa Catarina Court of Justice to test the new AI. “We made the selection based on the magistrates who already work with us in analyzing cases and participate in the e-NatJus management committee”, says counselor Daiane Lira.
There is still no definition regarding the selection of participating judges, nor the form of training. The work plan envisages “training 100% of the public involved remotely”.
The project does not generate costs for the CNJ, according to Lira. She highlights the experimental nature of the partnership.
“There is no obligation for the CNJ to assume any burden in relation to the storage, for example, of this system. Our commitment to InovaHC is that it cannot have operational costs”, he says.
“In the end, if the product involves a cost, the CNJ will make a decision, but this is not the subject of the agreement”, he states.
But the national expansion foreseen in the cooperation agreement between CNJ and InovaHC will require new resources and financial planning. Validation of the pilot will be decisive in determining whether the council will adopt the model on a large scale, according to Giovanni Cerri, president of the InovaHC board.
“If, after validating this algorithm, the CNJ finds that the pilot brings benefits, any costs of this technology will obviously fall on the CNJ. But this has nothing to do with AWS or any private company”, he says.
The director for the public sector at AWS Brazil, Paulo Cunha, defines big tech’s participation in the pilot as a contribution whose return lies in “market acceleration” and the diffusion of AI technologies.
“With a project like this, you train dozens of professionals and show that generative artificial intelligence can penetrate any environment. It’s a long-term investment,” he says.
Brazil has institutions that could support the continued operation of the model. Examples include Serpro (Federal Data Processing Service), DataSUS (Department of Information and IT of the Unified Health System) and Prodesp (São Paulo Data Processing Company), according to Ferreira, from IME.
“The model will need to be constantly updated. Maybe not daily, but at least with monthly versions, because neither medicine nor law stops”, says the professor.
AWS says that, if the resources given run out before the pilot is completed, “the possibility of new credit contributions is planned to ensure the continuity of operations.”
The company also states that, at the end of the project, teams from IME and AWS will be able to define migration strategies for other infrastructures, such as those in the public sector, for policy continuity.
Projects that use generative AI in critical areas, such as health and justice, often raise fears about possible “hallucinations” — when the system creates false information or information without a factual basis.
The developers claim that the model does not create new analyses, it only synthesizes existing technical opinions, displaying the source of each section consulted.
To achieve this, it combines two complementary technologies.
The first is the use of an SLM (small language model), aimed exclusively at information about health actions. Because it is leaner and more specialized, it tends to make fewer mistakes and operate at a lower cost.
The second is RAG (retrieval-augmented generation, or generation with integrated search), a mechanism that makes the AI consult official documents — such as the e-NatJus technical notes — before putting together a response. Together, they promise only evidence-based answers, without creating new content.
“If I specialize the model, I offer what matters and reduce the noise”, says the IME professor.
The researchers clarify that the project works in accordance with the LGPD (General Data Protection Law). When there is personal data, the material is anonymized, with the removal of name, address and other identifiers.
