AI WBSO pre-advice

Netherlands Enterprise Agency

This algorithm helps assess WBSO applications. It picks out applications with low financial risk, which can then be processed faster by RFO staff.

Last change on 20th of February 2025, at 10:42 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: DPIA
• Status: In use

General information

Theme

• Economy
• Public finance

Begin date

2023-01

Contact information

Via contactformulier: https://www.rvo.nl/onderwerpen/contact/formulier   

Responsible use

Goal and impact

The purpose of this algorithm is to assess WBSO applications faster. These are mainly simple applications with low financial risk. By processing these applications faster, R&D companies are more satisfied. It allows them to make use of their (tax) advantage sooner and to allocate their resources to innovation faster. The algorithm also helps to deploy technical experts within the Netherlands Enterprise Agency where they are needed most. In this way, the quality of work remains high and companies get more customisation.

Considerations

The use of the algorithm in WBSO applications ensures that less complex applications from entrepreneurs have a faster turnaround time. Less complex applications (with a low risk of rejection) are prioritised in the assessment of files by RFO staff. In this way, all applications get exactly the attention they need, increasing efficiency and productivity and, consequently, customer satisfaction.

Human intervention

The outcome of the algorithm is a supportive pre-advice, which will always be supervised by staff. The algorithm is supportive, and thus never makes the final decision in assessing the application itself. The final assessment is always done by a staff member, the algorithm itself does not make any decisions.

Risk management

There is constant monitoring of the algorithm to check that it still works without error messages. It also checks whether the algorithm might be biased, where different groups would be treated differently. It also checks whether the results of the algorithm still match the actual assessments of applications issued and thus whether the model is still good at predicting. Any changes to the AI model will again go through the extensive validation procedure. A check is made each year before the start of a new assessment year as to whether this procedure is necessary.

Legal basis

The WVA (Wet vermindering afdracht loonbelasting en premie voor de volksverzekeringen) regulates that an entrepreneur can get tax deductions for research and development work. The algorithm is used when assessing applications.

Links to legal bases

Wet vermindering afdracht loonbelasting en premie voor de volksverzekeringen (WVA): https://wetten.overheid.nl/BWBR0007746/2022-01-01

Elaboration on impact assessments

DPIA was carried out, outcome was that no personal data were processed. IAMA was not carried out, as there is minimal risk classification. IAMA only covers high-risk AI applications.

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

Data used in the algorithm include the number of (new) projects, number of hours, hourly wages, third-party presence, and any rejections or the outcome of a past check. The case number is used as an identification label. No personal data are used in the algorithm in any way.

Technical design

The risk is estimated by a self-learning Machine Learning model. This tries to make the best possible prediction with as few variables as possible in the simplest way possible. Then, applications with an estimated low risk are given priority in the prioritisation of the assessment.