PROMETEA: how Argentina is using AI to draft routine legal opinions and clear case backlogs

The problem

The Public Prosecution Service of the City of Buenos Aires handles a high volume of cases that follow similar patterns. Matters such as housing protection, labour and public-employment claims, enforcement of unpaid fines, and drunk-driving probation tend to involve standardised steps and largely predictable outcomes. Even so, each case required legal staff to read the file, assess the facts, and draft a ruling from the beginning.

The deputy attorney general's office for contentious administrative and tax matters was handling roughly 130 cases a month, and matters that reached trial took around 167 days to settle. So much of that effort went into routine drafting for cases whose result was already plain that there was little room left for the harder matters that needed real legal analysis.

The solution

In 2017, a team led by Juan Corvalán built an artificial intelligence system, PROMETEA, to handle the office's most repetitive drafting tasks. It was developed within the prosecution service alongside the University of Buenos Aires law faculty's innovation and AI lab, and the build is credited to ZTZ Tech Group. PROMETEA was trained on the office's own records, around 300,000 court documents scanned from 2016 and 2017, of which roughly 2,000 were rulings.

When PROMETEA is used, what it helps draft is a legal opinion, the prosecutor's formal written opinion on the case. The system prepares the first draft, and the decision on the case stays with the prosecutor. A prosecutor enters a case, sometimes using only its case number; PROMETEA reads it and compares it with earlier cases, looking for those whose facts most closely match. Based on how those comparable cases were resolved, it predicts the likely outcome and yields the same result for the case at hand. The prosecutor then answers a few questions about the matter, and the system uses the answers to assemble a first draft of the opinion, which appears on screen ready to be edited. The prosecutor reviews the draft, completes it, and makes the decision.

Two machine-learning methods sit behind this. Supervised learning, where the system is trained on examples that staff have labelled by hand, enables it to recognise the type of case in front of it. Clustering, which groups documents by shared features, helps pull together relevant material even when files are written differently. Staff work from a single screen, putting questions to the system in plain language or through a chatbot.

When the office mapped its own work, an internal task analysis found that of 169 activities, 54 could be handled by the system on its own, 41 in part, and 74 not at all. According to a more recent account, PROMETEA can handle around half of the office's routine work.

The case types it handles are the office's most standardised. Most often, prosecutors use it for traffic matters and routine legal opinions, freeing staff to spend more time on the harder cases. Prosecutors are required to review its output before anything is used, and the decision stays with them. In the office's own proof-of-concept testing, the proposed outcome matched what a human expert would provide about 96% of the time.

The impact

1. Reported productivity rose

The office reports monthly output climbing from about 130 to roughly 490 cases, a gain it puts at close to 300%. At the point of reporting, 33 of its draft opinions had been approved by prosecutors and put forward, and the system was in use on 84 open cases.

2. Case-resolution times fell

Matters that reached trial, previously around 167 days, were being settled in 38 days, a 77% reduction. In the office's measured examples, 1,000 housing rulings were completed in 45 days rather than 174, 1,000 labour rulings in 5 days rather than 83, and drunk-driving probation decisions in 26 days rather than 110.

3. Other public bodies took up the system

Beyond the prosecution service, PROMETEA was used at the Constitutional Court of Colombia, where the task of sifting urgent cases from a daily flood of filings dropped from 96 days to a couple of minutes; at the Inter-American Court of Human Rights, to help prepare documents; and at the Buenos Aires Civil Registry, which cleared about 6,000 administrative corrections in two months rather than eight. By Corvalán's account, more than 60 bodies have worked with PROMETEA, among them the United Nations, the Organisation of American States, and the Universities of Oxford and Sorbonne.

Lessons learned

• Standardised, high-volume cases are where the measurable returns appeared. PROMETEA was applied to case types that follow set patterns, such as housing protection or drunk-driving probation. For a government considering a similar tool, this indicates the conditions under which such a tool tends to pay off: a high volume of repetitive matters with a consistent structure, and a clear view of which tasks fall outside automation.

• Human review was built into the workflow, and system use is not recorded separately. Every draft is reviewed by a prosecutor before it is used, and decisions remain with people, with the system producing a first draft rather than a final ruling. For PROMETEA, staff are not required to report when they have used it, but are required to review any work it produces. Whether and how tool use is logged is a design choice that shapes later transparency and audit, and is one worth settling at the outset of a comparable project.

• A system that drafts decisions carries oversight considerations alongside its design for transparency. PROMETEA's creators describe it as traceable and open to inspection rather than a closed system whose reasoning is hidden. Assessments of AI in justice settings raise considerations that apply here as well: how automated conclusions are justified, where accountability sits when developers and officials rely on the system, and whether the data a model learns from carries earlier patterns into its outputs. Because PROMETEA was trained on the office's past rulings, those patterns form part of what it now proposes, which places these questions in the realm of governance as much as technology.

This case study was written with assistance from artificial intelligence.