State & Algorithm · No. 01
AI in Costa Rica's Public Sector: 21 projects, 7 institutions, and the questions no one has answered yet
Analysis of the first systematic inventory of AI projects in Costa Rica's public sector: what exists, what returns value, what is stalled, and what lacks coordination.
Executive Summary
- → 21 AI projects across 7 Costa Rican public institutions. 16 in production.
- → Poder Judicial (₡5,245M) and Hacienda (₡8,000M) have publicly documented financial returns. Both operate on pre-existing digital infrastructure.
- → Three CCSS models for detecting cancer, pulmonary disease, and acute coronary syndrome are stalled for ₡390M (less than 0.02% of the CCSS budget).
- → The National AI Center of Excellence, promised in ENIA 2024-2027, has no documented public operation.
- → There is no institutional obligation to share learnings or coordinate between entities.
The Inventory: What Was Measured and How
The first systematic inventory of the Observatorio IA Costa Rica documented twenty-one AI projects in production or with verified status across seven Costa Rican public institutions. Twenty-one is a floor, not a ceiling: the inventory is ongoing.
Inclusion Criteria
- Verifiable public source: official statement, institutional declaration, or press coverage with confirmed data.
- Costa Rican public sector institution: central government, autonomous, or semi-autonomous.
- System that operates or has documented status: the announcement alone is insufficient; evidence of operation or current status is required.
Deliberately excluded: pilot-phase projects without public data, private sector initiatives, and university projects without public entity agreements.
Documented Financial Returns
Of the twenty-one projects, two have the highest level of evidence: financial returns publicly documented by the institution itself. Both share a structural condition: they operate on data infrastructure that already existed before the AI project.
Poder Judicial: Seven Years Under the Radar
The citizen service chatbot has been running since 2018, built internally without an external vendor. The budget prediction model for judicial debt collection (2019) expanded to more than 60 management centers with accumulated savings exceeding ₡100 million.
In 2024, the Poder Judicial processed ₡5,245 million in judicial collections without manual case-by-case review. The system classifies, prioritizes by recovery probability, and generates reports automatically.
transparencia.poder-judicial.go.cr · pj.poder-judicial.go.cr · observador.cr
Ministry of Finance: Built on E-Invoicing Infrastructure
In 2025, an anomaly detector applied to electronic invoice flows detected ₡8,000 million linked to simulated invoices. Costa Rica's e-invoicing system processes approximately 3 million receipts per day. That pre-existing infrastructure was the enabling condition for the detector.
“The return on government AI depends less on the model used and more on the quality of the data and systems that precede it. Institutions without that foundation cannot skip the step.”
The Most Urgent Case: CCSS
If Poder Judicial and Hacienda represent what already works, CCSS represents both what works and what could work but is stalled.
The bot that cross-references EDUS with the Civil Registry cleans surgical waitlists: deceased patients, those already treated, and duplicates. Between 2023 and Q1 2026, it resolved 367,403 patients and removed 136,774 cases. The cleanup rate dropped from 31.2% to 18.2%, indicating more accurate lists, not just shorter ones.
LIDIA: the working model and the three that are stalled
LIDIA is a machine learning model developed at Clínica Clorito Picado on more than one million EDUS records. It identifies at-risk type 2 diabetes patients with documented high accuracy in the pilot, enabling preventive intervention. The model costs ₡130 million (approx. USD 250,000).
Three additional models designed to detect breast cancer, pulmonary diseases, and acute coronary syndrome on the same dataset are stalled for lack of budget: ₡390 million in total (3 × ₡130M). That figure represents less than 0.02% of CCSS's ordinary budget.
Pending policy decision: If the cost of activating these models is verifiably lower than the cost of a preventable complication per late-detected case, the justification for inaction requires explicit argumentation. That argument does not exist in the public domain.
Timeline: AI in Government (2018–2026)
The documented projects span eight years. The most mature institutions—Poder Judicial and CCSS—started before a national strategy existed.
The Institutional Gap
The MICITT's National AI Strategy (ENIA) 2024-2027 promised a National AI Center of Excellence. Costa Rica was the first country in Central America to adopt a national AI policy. The center has no documented public operation.
Meanwhile, UCR conducts AI ethics research with Erasmus+ funding and CeNAT proposes LaNIA as an applied research platform. Each institution builds its own systems without a formal obligation to coordinate with MICITT or share learnings.
Poder Judicial created its own governance framework because no one else did: it is the only institution to have formally published guidelines for internal generative AI use, according to available public information. The practical result is that learnings from each institution are not available to others. Each starts from zero.
Catalog of Identified Projects
The table includes the ten main projects documented with verifiable public data. The complete inventory contains twenty-one projects across seven institutions.
| Institution ↑ | Project ↕ | Year ↕ | Status ↕ | Impact |
|---|---|---|---|---|
| CCSS | Waitlist cleanup bot | 2023 | Active | 367,403 patients resolved |
| CCSS | LIDIA (type 2 diabetes) | 2025 | Pilot | High pilot accuracy, 1M+ records |
| CCSS | Breast cancer model | — | Stalled | ₡130M required |
| CCSS | Pulmonary disease model | — | Stalled | ₡130M required |
| CCSS | Acute coronary syndrome model | — | Stalled | ₡130M required |
| Hacienda | Tax evasion detector | 2025 | Active | ₡8,000M recovered |
| Poder Judicial | Citizen chatbot | 2018 | Active | 24/7 service |
| Poder Judicial | Budget prediction | 2019 | Active | 60+ centers, ₡100M+ saved |
| Poder Judicial | AI judicial collections | 2024 | Active | ₡5,245M processed |
| Poder Judicial | Generative AI governance framework | 2024 | Active | Only public guideline |
Selection of projects with verified public data. The full Observatory inventory includes 21 projects.
What's Next at the Observatory
This first edition of the inventory is a starting point, not a final state. The open questions that will guide future editions:
- What is the total cost of the active projects? Returns are documented for PJ and Hacienda. Investment costs lack the same transparency.
- Which institutions have the data infrastructure for the next project? MEP, SUTEL, and municipalities have data. The question is whether they have the structure to use it.
- When will the three LIDIA models be activated? The budget decision has a deadline. So does the follow-up.
The inventory accepts corrections and additions. If you work at a public institution with an AI project that is not documented, the channel is open.