When it comes to tax enforcement, data and technology aren’t always the answer

Governments around the world mobilise tax revenue to provide public goods, redistribution and social insurance. Unfortunately, efforts to raise revenue are often hampered by high levels of non-compliance. In response, tax administrations carry out audits on their taxpayers to uncover tax evasion and punish evaders. In rich countries, data from third-parties and algorithms are used successfully to detect fraud and guide audit selection. But would similar data-driven strategies fare well in other contexts?

Senegal’s tax administration engaged in a small revolution of its enforcement activities over the past years. It digitized and assembled all available data sources linked to the economic activity of firms and their taxes. It used this information to predict fraud by constructing risk indicators, which followed international best practices. For example, firms with abnormally high deduction, or firms with large deviation between self-reports and third-party reports were flagged as suspicious.

How taxpayers get selected for audits

The digitization efforts and risk indicators allowed the tax administration to select firms in a data-driven manner. This contrasted with the traditional discretionary audit selection: inspectors often picked the largest and most visible firms, and when selecting smaller firms, they used their experience and soft information. The discretionary method has its strengths and weaknesses: well-trained and experienced inspectors may perform better than algorithms in detecting potential tax evasion, since algorithms can only work with encoded data, but are prone to biases.

Starting in 2018, the tax administration ran part of its audit selection in a data-driven way, while the remaining part continued to be selected directly by tax inspectors. Over the past years, the tax administration expanded the number of tax centres that use the new audit selection method, and by 2021 all tax centres in Senegal selected half of their audit cases in this manner.

Comparing audits selected by technology vs. by tax inspectors

The two methods select very different taxpayers for audits: inspectors select firms with much larger reported sales. Although in theory inspectors are supposed to complete the entire audit program determined at the beginning of the year, in practice, they rarely undertake all cases. The algorithm’s cases were 18% less likely to be started compared to discretionary cases, reflecting inspectors' preferences for their own cases. However, the execution rates improved from 2018 to 2020, as the result of repeated information sessions about the algorithm.

Comparing the results for algorithm and discretionary cases actually conducted does not improve this picture, however. Despite presenting multiple risks, the algorithm-selected audits were slightly less likely to present any irregularity compared to the inspector-selected cases. For cases that presented irregularities, evaded amounts were equivalent. Thus, audits selected by the algorithm were less likely to be started, slightly less likely to show any irregularity, and when irregularities were found, showed a similar level of non-compliance.

Towards a long-term improvement in tax capacity?

The at-scale implementation of best international practice in audit selection by a tax administration is a stark example of tax capacity building. Senegal’s tax administration made important efforts to centralise and digitise relevant data sources, build risk indicators, and change their audit selection method. Unfortunately, choosing audits via a risk-based algorithm did not uncover more tax evasion, highlighting the challenges in using data to improve audit selection relative to a discretionary method based on inspectors’ experience and private knowledge.

In our view, the results reinforce what many practitioners already know: changing key institutional practices (and their power structure) with a top-down technological approach is not guaranteed to produce results in the short to medium-term. Technologies that might work successfully in rich countries' contexts cannot be easily replicated in developing countries. The successful implementation of data-driven audit selection hinges on high-quality data, fine-tuning indicators to local firm practices, and convincing inspectors of the importance of an objective audit selection mechanism. This takes time, and it is too early to tell how much current practices can be improved.

Senegal’s efforts in modernising its audit selection process is praiseworthy, and keeps on expanding: its data lab has become institutionalised and now covers more types of taxes and taxpayers; the data-driven selection of audits continues with more inputs by inspectors into the design of risk indicators; and as audit results data gets digitised, formal evaluations are underway. These evolutions could generate a virtuous cycle of cooperation between the data unit leading the technological change and tax inspectors, which will ensure that in the long-term data and algorithms are a productive part of the tax audit arsenal.

Authors:
Pierre Bachas is an Economist in the Macroeconomics and Growth Team in the Development Research Group at The World Bank.
Anne Brockmeyer is Research Director for Tax and Development at the Institute for Fiscal Studies, and a Senior Economist at the World Bank.
Alípio Ferreira Cantisani is a researcher at Toulouse School of Economics.

Bassirou Sarr is a Technical Advisor at the Ministry of Finance and Budget – Sénégal,  

All are also researchers on the Economic Development and Institutions collaborative research programme led by Oxford Policy Management. See the project details here.