Cameroon-Built AI Claims 97.2% Precision in Tax Fraud Detection, but Revenue Impact Remains Unclear

A team of researchers says an artificial intelligence system designed for Cameroon can identify tax fraud with 97.2% precision, potentially offering the tax administration a new tool for risk-based audits. But publicly available information does not yet establish how much additional revenue the system could actually generate.

The system was developed by researchers affiliated with the universities of Ngaoundéré and Dschang and France’s ESSCA School of Management. According to the authors, it was evaluated using 3.2 million transactions and later tested at three tax centers, where it reportedly reduced processing time by 67% and generated additional revenue.

The amount of additional revenue was not disclosed. The publicly available summary also does not identify the three tax centers, specify how long the pilot lasted or state how many cases were actually reviewed. As of Aug. 25, 2026, no publicly available document from Cameroon’s Directorate General of Taxation (DGI) identified during this review corroborated the specific pilot.

The findings were presented by Moïse Demlegue Frina, Jean Robert Kala Kamdjoug and Donald Richie Kuete Menou in a WorldCIST 2026 conference chapter titled Optimizing Tax Revenue Through Machine Learning-Driven Fraud Detection: An Explainable Approach for Cameroon. Springer Nature published the chapter online on Aug. 2, 2026, although the publisher lists the proceedings volume as 2027.

What the 97.2% figure actually measures

According to the abstract, the researchers used 3.2 million records described as tax, banking and customs transactions collected between May and November 2025.

Their system combines supervised machine-learning models, including XGBoost, Random Forest and support vector machines, with unsupervised methods such as Isolation Forest and autoencoders. Bayesian optimization is used to tune the models.

The best ensemble model reportedly achieved 97.2% precision, 94.8% recall and an F1 score of 96%. The figures are mathematically consistent, with the harmonic mean of precision and recall reaching 95.985%, or 96% when rounded.

In a standard binary classification model, precision measures the proportion of transactions flagged as fraudulent that are actually fraudulent. A precision rate of 97.2% would therefore imply that 2.8% of the system’s alerts are false positives. Recall, meanwhile, measures how much of the fraud present in the sample the model successfully identifies.

The 97.2% figure should not, however, be interpreted as an overall success rate for tax audits. Its significance depends on factors including the prevalence of fraud in the dataset, how fraudulent cases were labeled, the threshold used to trigger alerts and how training and test data were separated.

The public abstract provides no confusion matrix, fraud prevalence rate or description of testing on an entirely new period or group of taxpayers. Because the full chapter is behind a subscription, these limitations apply to the publicly available information and do not establish that such details are absent from the complete study.

The 97.2% precision figure also cannot be directly compared with the 34% false-positive rate that the authors attribute to existing DGI procedures. If the terms are being used in their conventional statistical sense, they have different denominators. A valid comparison would require the full confusion matrix and fraud prevalence for both systems.

The researchers also say their system can process more than 10,000 transactions per second with latency below 100 milliseconds and generate a risk score in under two seconds. The abstract does not specify the infrastructure or testing conditions behind those results.

It also reports a 67% increase in acceptance of the system after explanations were generated using SHAP and LIME, without specifying the number of participants, the starting level of acceptance or the evaluation protocol.

Pilot reported, but financial return remains unproven

The most consequential claim concerns the operational pilot. The researchers say deployment at three tax centers cut processing time by 67% and generated additional revenue.

The public summary, however, does not identify the centers, provide the dates of the pilot, disclose the number of cases examined or quantify the additional tax assessments and collections. It also does not specify the baseline used to calculate the reduction in processing time.

Those gaps make it impossible to assess the system’s economic return from the information currently available. Such an assessment would require comparing revenue actually collected with the costs of infrastructure, database integration, security, staff training, maintenance and handling disputes. Tax assessments would also need to be distinguished from amounts ultimately collected after appeals.

The project nevertheless addresses a need already identified by Cameroon’s tax administration.

In its 2023-2025 tax system modernization plan, the DGI said it lacked a centralized tool for large-scale data cross-checking. Its plans included acquiring the necessary infrastructure, developing Big Data skills and connecting a cross-checking tool to FUSION for risk analysis and audit selection.

The DGI’s 2026-2028 strategic plan shows that 45% of audits were being programmed through FUSION as of Dec. 15, 2025, compared with a 90% target. The administration still reported the absence of a centralized large-scale cross-checking tool, while work continued to optimize FUSION and develop a data lake.

The same document nevertheless identifies large-scale automated cross-checking among the strengths highlighted by the 2025 TADAT assessment. The absence of a centralized architecture therefore does not mean that no automation existed, nor does it rule out a local pilot. It indicates that administration-wide integration remained incomplete.

For 2026-2028, the DGI plans to put artificial intelligence, Business Intelligence and predictive systems into effective use and establish a unit dedicated to large-scale cross-checking of internal and external data.

Those plans are consistent with the type of system presented by the researchers, but they do not confirm the pilot described in the study.

Baudouin Enama