ArgentinaPublic sector2017 – ongoingPublic Prosecution Service of the City of Buenos Aires
Prometea legal drafting system
A Buenos Aires prosecutors' office nearly quadrupled monthly output with its own AI, then moved most drafting to ChatGPT
Problem
The deputy attorney general's office for administrative and tax matters handled large volumes of repetitive cases, such as housing protection and public employment claims, where written opinions followed familiar patterns. Drafting them by hand took prosecutors' time away from harder cases.
Approach
In 2017 a team assembled by Juan Corvalán, then at the prosecution service, built Prometea with the University of Buenos Aires law faculty's AI lab. Trained on about 300,000 scanned case documents, it matched new cases with past ones, predicted the likely outcome and assembled a draft opinion from a few answers, which prosecutors reviewed. Because each new case category needed further training and development, the office began phasing Prometea out in early 2024 and, from May 2024, used ChatGPT to draft rulings in some public employment cases.
Results
Takeaway
Prometea greatly increased output for a few case types, but the cost of extending it led the office to a general-purpose model. Most performance figures came from the office and its developers and have not been independently verified.
Categories
- Size
- Small business
- Industry
- Public sector
- Business function
- Finance, risk and legalInternal work and knowledge management
- Innovation type
- Process innovation
- AI technology
- Predictive analytics and MLGenerative AI
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
The office said productivity rose by nearly 300% after Prometea launched in 2017, from about 130 to 490 cases a month. Every new crime category required more training, which led Corvalán to begin phasing it out in early 2024 in favour of ChatGPT. The office says ChatGPT cut drafting time from about an hour to ten minutes across 20 rulings, all reviewed by a lawyer, while experts raised concerns about bias, personal data and hallucinations.
Prometea was built with the University of Buenos Aires and trained on about 300,000 documents from 2016 and 2017. Reported time reductions included matters reaching trial falling from 167 to 38 days, and an internal review found 54 of 169 tasks could be fully automated. Most figures came from the office and its developers.
Full details are in the original sources.
