Buenos Aires courts cut ruling time by moving from Prometea to ChatGPT

In 2017, Buenos Aires’s public prosecution service built Prometea to predict outcomes and automate routine judicial work. It trained the system on more than 300,000 rulings and case files, while legal staff continued to review its recommendations.

Prometea made standardized cases much faster: the office’s productivity rose nearly 300%, from about 130 to 490 cases a month, and the system could predict outcomes with 90% accuracy in under 20 seconds. But every new crime category needed more training and algorithm development.

In 2024, the office began phasing Prometea out for ChatGPT in some public-employment cases. Staff upload documents, receive a preliminary classification and draft, then a lawyer reviews every sentence and the deputy attorney approves it. Twenty sentences had been drafted, cutting preparation time from about an hour to 10 minutes, and the office was expanding the program while working on anonymization and keeping Prometea for sensitive cases.

ChatGPT cut drafting time from about one hour to 10 minutes; Prometea had already lifted monthly output from about 130 to 490 cases, nearly 300%.

The office had already mapped repeatable legal work into templates, giving a generative tool a narrow operating lane.

A human editor remained accountable for the final text instead of treating the model’s draft as a ruling.

The switch reduced the maintenance burden of building a separate model for each new case category.

The team treated privacy, bias, and hallucinations as operating constraints and kept the older system for more sensitive matters.

Automate repeatable legal work, but keep human review at the decision boundary: speed helps only when sensitive inputs and outputs remain governed.

The shift was still an experiment: only 20 sentences had been drafted with ChatGPT, and staff were still developing an anonymization process for sensitive case data. The office was expanding the tool to procurement complaints, while Prometea remained in use for public-employment rulings and research involving child sexual-abuse material. The old system’s footprint therefore became a governance guide even as the newer tool handled more routine work.