Top 47%
in this catalogue (1041 scored practices)
Scores cluster high, so position within the catalogue is often more telling than the number alone.
Evidence of impact / measured public value3/3
A randomized, at-scale field experiment embedded in Senegal's real national audit program, published as a World Bank working paper with quantified outcomes (18pp fewer audits completed, 89% less evasion detected), is as rigorous as evidence gets in this dataset.
Transparency, fairness & accountability2/3
The algorithm was explicitly built and described as a transparent risk-score, and the researchers published the negative findings openly rather than only positive results, though the underlying scoring formula was not made public to taxpayers.
Transferability / demonstrated replication1/3
The paper's core lesson is that a naive ML approach did not transfer well to a data-poor tax administration with only ~500 audits a year of training data, so the practice itself offers a caution rather than a proven blueprint to replicate.
Scalability beyond pilot1/3
The algorithm was tested at full program scale, not a small pilot, but the experiment concluded it was less cost-effective than the status quo, so scaling it further is explicitly not recommended by the study's own authors.
Governance, capability & sustainability2/3
DGID, Senegal's national tax authority, ran the trial jointly with an independent World Bank research team under a clear experimental design, but no new standing oversight body was created around the tool itself.
Evidoria. Algorithms and Bureaucrats: Senegal's DGID Tests Machine-Learning Tax Audit Selection Against Human Inspectors. Persistent ID: 94ce09ef-b6fa-496c-b9e0-08d6c4cc346d.