Report
AI in regulatory agencies: the system finds, the agency decides
Key takeaways
- Risk-based targeting is public AI's best business case — if the hit rate is measured against a random sample and published.
- Collusion screening over public procurement data is the region's highest-return untapped application.
- The line: the system finds; the agency decides. No automated sanctioning, not even by drip.
- Bias against the small and formal is the specific distributive risk; the distribution metric exposes it.
- The internal generative rule: human verification of every fact and citation; no administrative act with generative content without substantive review.
Supervising thousands with teams of dozens is public AI's most promising use case — and its most delicate, because the end product is the sanction. Five families of application that pay off, and one line never to cross.
The full document is currently published in Spanish; an English edition is prepared when demand warrants it. The Spanish record carries the complete summary and contents.