Signal model office test OB (OBN)

Dutch Tax and Customs Administration

This page contains information about the 'OB Negative' algorithm. This algorithm supports the tax inspector in determining which turnover tax returns with a negative balance (a tax refund within the small and medium-sized enterprise (SME) target group) should be assessed manually.

Last change on 8th of April 2025, at 9:11 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Public finance

Begin date

Field not filled in.

Contact information

[e-mail]

Link to publication website

https://over-ons.belastingdienst.nl/onderwerpen/omgaan-met-gegevens/algoritmeregister/

Responsible use

Goal and impact

Most businesses file sales tax (also known as VAT) returns. A negative sales tax return means that a business has paid more VAT in the return period than it received and therefore requests a refund from the tax authorities. This may be the case, for example, in the case of high investments or low turnover in that period. The Tax and Customs Administration receives some 2.6 million negative returns every year. Some of these negative returns are assessed manually. This means that an employee checks whether the return has been filled in correctly on the basis of supporting documents. If this reveals inaccuracies, a correction is made.

• What is the algorithm used for?

Not all negative declarations can be assessed manually. The number is too large for that. Also, the Tax and Customs Administration does not want to burden entrepreneurs unnecessarily with providing supporting documents. That is why the manual assessment focuses mainly on the declarations that are estimated in advance to have a higher chance of being incorrect. The algorithm OB-negative is used to detect these potentially incorrect SME returns. Based on the results of the algorithm, the returns are selected for manual assessment.

Considerations

By using OB-negative, potentially incorrect returns can be better spotted. The algorithm makes optimal use of the relevant data, such as the amounts entered in the return. Compared to manual selection, the algorithm detects more incorrect returns. It also selects fewer declarations that do not contain inaccuracies.

Selection by the algorithm always takes place in the same way. This ensures efficient and consistent selection. By using the algorithm, the employee can fully focus on the manual assessment.

Human intervention

Human intervention in the Inland Revenue context implies that a competent and knowledgeable employee plays a substantial role in decision-making.

The algorithm determines whether a refund decision can be issued automatically for the negative return, or whether it is selected for manual assessment. In that case, an employee assesses the correctness of the return after investigation. Any correction follows only after this manual assessment. This involves substantial human intervention in decision-making when there are legal consequences.

Risk management

The General Administrative Law Act (Awb) requires the government's actions to be transparent and lawful. The Tax and Customs Administration observes the general principles of good governance when applying and developing algorithms.

Conditions, a quality framework, have been drawn up by the Tax and Customs Administration for the development of algorithms. This contains rules and agreements that are followed during algorithm development. The conditions of the National Audit Authority are leading in this respect. At set moments, the Tax and Customs Administration checks whether the algorithm still meets the quality requirements.

The algorithm uses data collected under various tax laws. As required by the AVG, no more data is used than necessary.

The algorithm is reviewed periodically and adjusted if necessary to remain compliant with laws and regulations.

Legal basis

• General Administrative Law Act
• General State Taxes Act
• Citizen Service Number (General Provisions) Act
• Turnover Tax Act 1968
• Collection Act 1990
• 2003 Tax and Customs Administration Implementation Regulations
• Archives Act 1995
• Regulation implementing the General State Tax Act 1994
• Directive 2006/112/EC on the common system of value added tax
• Chapter IX of the National Budget
• Regulation (EU) No 904/2010 on administrative cooperation and combating fraud in the field of value added tax

Links to legal bases

• Algemene wet Bestuursrecht: https://wetten.overheid.nl/BWBR0005537/
• Algemene wet inzake rijksbelastingen : https://wetten.overheid.nl/BWBR0002320/
• Wet algemene bepalingen burgerservicenummer: https://wetten.overheid.nl/BWBR0022428/
• Wet op de omzetbelasting 1968: https://wetten.overheid.nl/BWBR0002629/
• Invorderingswet 1990: https://wetten.overheid.nl/BWBR0004770/
• Uitvoeringregeling Belastingdienst 2003: https://wetten.overheid.nl/BWBR0014506/
• Archiefwet 1995: https://wetten.overheid.nl/BWBR0007376/
• Uitvoeringsregeling Algemene wet inzake rijksbelastingen 1994: https://wetten.overheid.nl/BWBR0006736
• Richtlijn 2006/112/EG betreffende het gemeenschappelijke stelsel van belasting over de toegevoegde waarde: https://eur-lex.europa.eu/legal-content/nl/TXT/?uri=CELEX%3A32006L0112
• Hoofdstuk IX van de Rijksbegroting: https://www.rijksfinancien.nl/hoofdstuk/IX/2023
• Verordening (EU) nr. 904/2010 betreffende de administratieve samenwerking en de bestrijding van fraude op het gebied van de belasting over de toegevoegde waarde: https://eur-lex.europa.eu/legal-content/NL/ALL/?uri=celex%3A32010R0904

Elaboration on impact assessments

The algorithm's data processing operations are tested against the AVG. If risks are found, measures are taken to ensure privacy and security.

The AVG prescribes that no more data should be used than necessary. This is called data minimisation. The Tax Authority regularly examines whether the data used is still necessary and therefore may be used.

• Equality and non-discrimination

The algorithm is tested against non-discrimination legislation. No special personal data, such as ethnic origin, are processed by the algorithm. Only necessary personal data are processed. This reduces the risk of discrimination. Employees involved in developing and managing the algorithms receive training on data protection and prejudice.

Operations

Data

• Declaration details
• Payment details
• Company details
• Previous Manual Sales Tax Assessments

Links to data sources

• Aangiftegegevens: Belastingdienst
• Betalingsgegevens: Belastingdienst
• Bedrijfsgegevens: KvK (Kamer van Koophandel) en Belastingdienst
• Eerdere Handmatige Beoordelingen Omzetbelasting: Belastingdienst

Technical design

The algorithm is not self-learning. This means that it does not adapt itself to better recognise incorrect returns. The statistical model has been trained by staff based on past data to detect correlations. The algorithm is evaluated periodically and updated when necessary. For an update, the algorithm is re-trained.

The algorithm OB-negative consists of two parts, a statistical model and number of business rules:

1. The statistical model gives a score to each negative return within the SME target group. This score is a combination of the estimated probability that the return is possibly incorrect and the size of the possible deviation (the possible correction amount). The declarations with a high score are assessed manually.
2. The business rules are fixed rules that determine whether manual assessment should take place. These rules are determined by experts based on legislation and known risks.

External provider

The algorithm was developed by Tax Administration staff and is maintained internally.