What happened
The Belastingdienst, the Dutch tax and customs administration, uses a risk model to choose which VAT refund returns its inspectors check. The model, which replaced the old selection rules, estimates the chance of an error in every refund return. The Netherlands Court of Audit describes the models that it examined as trained and tested on returns with known check outcomes. Its report went to the House of Representatives on 11 June 2019. The Court found that the hit rate with the new model was higher than with the old rules, at 34% against 17%. It writes that the Belastingdienst can show this with a sample in which the selected returns were checked in full. Inspectors examined 32,000 returns a year in 2014 and 2015 and 40,000 a year with the model. The Belastingdienst estimates that the model brings in €60 million a year in extra revenue. The Court notes that the Belastingdienst had no results yet from measuring extra revenue in practice. It found that the income tax model for entrepreneurs was harder for inspectors to use.
Sources
-
Audit report, by authors with no tie to the system.
Algemene Rekenkamer (Netherlands Court of Audit). Datagedreven selectie van aangiften door de Belastingdienst, §2.1.3 Trefkans, p. 10, and Bijlage 1 'Van data naar risicomodel', pp. 28-29. Sent to the House of Representatives on 11 June 2019.
Wij hebben geconstateerd dat de trefkans bij btw-teruggave met het nieuwe risicomodel groter is dan met het oude model.
-
Press release, by authors with no tie to the system.
Netherlands Court of Audit. Tax and Customs Administration books good results using risk models to check returns. News item, 11 June 2019.
Most errors were detected by a relatively simple model to check VAT refunds. The new approach increased the proportion of returns detected with errors from 17% to 34%, good for additional tax receipts of approximately €60 million a year.
Details
AI system
- Workflow
-
Inspectors checked the VAT refund returns to which the model gave the highest chance of an error. The Belastingdienst trained the model on returns whose check outcomes were known and tested it on a separate file. Selected returns went to a desk check, followed by a books audit where the desk check called for one.
Money
- Kind of value
-
Extra revenue from the new VAT refund model, by the Belastingdienst's estimate
- Amount
-
Estimate: € 60 miljoen per jaar
- Recipients and payers
-
Dutch public purse
Follow-up
- Conditions for replication
-
The Belastingdienst must have returns with known check outcomes, from samples or recent checks, to train and test the model. Inspectors must have the capacity to check the selected returns, since capacity set how many returns were checked.
Entry details
- Sector
-
Tax compliance.
- Purpose
-
The Belastingdienst built its risk model for VAT refunds to estimate the chance of an error in every VAT refund return, so that inspectors check the returns with the highest risk.
37 more questions not yet checked
model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, currency and prices, period and unit, realization, gross or net, included costs, comparison, method, return and payback, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). Risicomodel van teruggave btw doubles the hit rate of VAT checks (AIBD-1029). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-1029.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)