evidoria

← Back to browse

Good practice Imported

Algorithms and Bureaucrats: Senegal's DGID Tests Machine-Learning Tax Audit Selection Against Human Inspectors

Senegal · Dakar · See the Senegal profile · See the Dakar profile

Evidence: Quasi-experimental Top 47% 60/100 · Ask Evidence Copilot about this practice

Senegal's tax authority DGID and World Bank researchers split the national corporate audit program between inspectors and a transparent ML risk-scoring algorithm. Algorithm-picked audits detected 89% less evasion, showing human expertise still beat the algorithm here.

-89%
Evasion detected, algorithm- vs inspector-selected audits (study cycle)
-18 percentage points
Reduction in audit likelihood, algorithm-selected cases
~500
Full audits completed annually
Algorithms and Bureaucrats: Senegal's DGID Tests Machine-Learning Tax Audit Selection Against Human Inspectors

Details

Maturity
Pilot
Promoter
Direction Generale des Impots et des Domaines (DGID), with the World Bank Development Research Group
Period
2018-2025
Keywords
taxation, revenue administration, compliance analytics

Context

Senegal's tax authority DGID, with the World Bank's Development Research Group, ran an at-scale field experiment embedding a machine-learning algorithm directly into the national corporate tax audit programme.

Activities

Half of the audit slate was selected by experienced tax inspectors using their usual judgement; the other half by a transparent risk-scoring algorithm trained on prior filings and audit outcomes, compared across a full audit cycle.

Results

Algorithm-selected cases were 18 percentage points less likely to actually be audited, detected 89% less evasion than inspector-selected cases, and were less cost-effective; the algorithm did not reduce corruption risk.

Conclusions

Senegal completes only around 500 full audits a year, too little outcome data to train a model matching a veteran inspector's judgement — a cautionary result for other revenue authorities.

Implementation

Indicative cost
Medium (€50k–€500k) — Not disclosed; a research collaboration rather than a large capital deployment.
Time to results
Medium (1–3 years) — Comparison ran across one full annual audit cycle; published November 2025.
Staffing & skills
DGID tax inspectors, World Bank Development Research Group researchers

Conditions for success

  • Sufficient historical audit-outcome data to train a competitive model
  • Transparent, auditable risk-scoring methodology

Common failure modes

  • Data-constrained tax administration (~500 audits/year) limited model training quality
  • Algorithmic selection underperformed experienced human inspectors on every measured dimension

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful