Direct naar content

You are here:

  1. Home
  2. Algorithms

Please note: The algorithm descriptions in English have been automatically translated. Errors may have been introduced in this process. For the original descriptions, go to the Dutch version of the Algorithm Register.

Additional Fixed Fee: audit NAM files

Groningen Mining Damage Institute

An AI system that can extract the claims paid from NAM files to determine whether the claimant from that file is eligible for the Additional Fixed Compensation (AVV) scheme.

Last change on 23rd of January 2026, at 11:07 (CET) | Publication Standard 1.0
Publication category
Impactful algorithms
Impact assessment
DPIA
Status
In use

General information

Theme

  • Nature and Environment
  • Law

Begin date

2024-12

Contact information

privacy@schadedoormijnbouw.nl

Responsible use

Goal and impact

Purpose: It is important for the implementation of the AVV scheme that the IMG works with the correct damage amounts paid out. In the past, it was found that the correct damage amounts from the files of the Netherlands Petroleum Company (NAM) were not always copied into the IMG's files. The AI system can extract the correct damage amount from the NAM files. These NAM damage amounts are then used to calculate the additional fixed compensation.


Impact: The AI system affects people who have received a damage amount from NAM and have now applied for the AVV. The outcome of the AI system calculation is used to assess whether a NAM claim qualifies for the AVV.

Considerations

Advantages:

  • Automated verification of disbursed claims from NAM files is faster than manual verification. This allows applications to be processed faster.
  • The algorithm finds the same amount listed in the database in 70% of cases. This saves the same amount of time. Without this time saving, the wide opening was not possible.


Disadvantages:

  • There is a chance that the results may not match the amount from the database. If this is the case, the result is checked by a human.
  • The model is quite heavy and the analysis takes quite a long time if automated. It averages 23 hours for 1,000 files.

Human intervention

The AI system can give two outcomes:

  1. The amount found by the system matches the amount previously recorded from the NAM file in the IMG file, or,
  2. The amount does not match.



  1. If an amount matches, it is automatically used for the AVV. No human intervention is then required. However, the AI system is regularly checked by a human for performance.
  2. The amount does not match. When the amount does not match, a file is always completely reassessed by a human.

Risk management

The GTC is part of the Lump-sum Scheme. A Data Protection Impact Assessment (DPIA) was carried out on the Lump-sum Scheme, which proposed measures to minimise the risks of, among other things, the AVV. In addition, a comprehensive study was conducted from the IMG on the legal qualification of the algorithm under the European AI Regulation.

Legal basis

Art. 2(3) Groningen Temporary Act and Art. 40 UAVG.

Links to legal bases

  • Groningen temporary law: https://wetten.overheid.nl/jci1.3:c:BWBR0043252&hoofdstuk=2&artikel=2&z=2025-07-17&g=2025-07-17
  • General Data Protection Regulation Implementation Act: https://wetten.overheid.nl/jci1.3:c:BWBR0040940&hoofdstuk=4&artikel=40&z=2021-07-01&g=2021-07-01

Elaboration on impact assessments

A DPIA was conducted on the Lump-Sum Scheme, of which the AVV is a part. In addition, the IMG conducted a comprehensive legal qualification of the AI system under the AVG and the AI Regulation, in the form of an advisory.

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

Training dates:

  • None. The model used concerns GPT 4.0-turbo. This is a trained LLM (large language model) capable of executing instructions (prompts) based on text.


Input data:

  • NAM claim files. These are the Word and PDF documents.

Technical design

Data processing is done in three steps, namely: reading out, anonymising and extracting.


  1. In reading out, Word and PDF documents are converted into plain text on a page-by-page basis. Each page is passed to the second step.
  2. In this step, each page is anonymised using the Wegstreepn package, a prgramma used within the IMG to anonymise documents. The anonymisation is required before the data is processed using the GPT model in the third step.
  3. In this step, all pages of all documents are sent to the GPT model together with a prompt. The model then returns the said claim amount (document name and page number). If additional charges are also involved, these are also sent. The location is also fed back.

External provider

The AI system will be built on OpenAI's GPT-4, within the IMG's Microsoft Azure environment.

Link to code base

https://dev.azure.com.mcas.ms/dictupoc/LandingZone_img_img_dp_017/_git/nam-textmining

Similar algorithm descriptions

  • Weighing and determining treatment of corporate income tax (Vpb) returns

    Dutch Tax and Customs Administration

    Algorithm that helps assess whether a corporate tax return should be processed automatically or manually.

    Last change on 13th of November 2024, at 12:11 (CET) | Publication Standard 1.0
    Publication category
    Impactful algorithms
    Impact assessment
    Field not filled in.
    Status
    In use

  • Benefits Automatic Income Settlement

    Municipality of Utrecht

    By applying the OIB-AIV process, we can process data on income of a resident receiving supplementary welfare benefits better and faster in the benefit administration. The effect is that the resident will no longer need to provide such income and related supporting documents, and any settlement or supplementary payment will be made faster and more unambiguously

    Last change on 31st of October 2025, at 13:29 (CET) | Publication Standard 1.0
    Publication category
    Impactful algorithms
    Impact assessment
    Uthiek
    Status
    In use

  • AiBezwaar (Taxation)

    Municipality of Aalten

    AiBezwaar supports the processing of objections by reducing the administrative burden. Incoming e-mails are analysed automatically, identifying relevant data such as the submitter, main parties involved and reasons for objection. The algorithm then suggests a text to answer the objection.

    Last change on 7th of October 2025, at 13:05 (CET) | Publication Standard 1.0
    Publication category
    Other algorithms
    Impact assessment
    Field not filled in.
    Status
    In use

  • AiBezwaar (Taxation)

    Municipality of Aa and Hunze

    AiBezwaar supports the processing of objections by reducing the administrative burden. Incoming e-mails are analysed automatically, identifying relevant data such as the submitter, main parties involved and reasons for objection. The algorithm then suggests a text to answer the objection.

    Last change on 6th of June 2025, at 12:21 (CET) | Publication Standard 1.0
    Publication category
    Other algorithms
    Impact assessment
    Field not filled in.
    Status
    In use

  • Signal model Transfer Tax for returns (OVB)

    Dutch Tax and Customs Administration

    This is a signal model for the Overdrachtsbelasting (OVB). The model checks whether claims for the OVB start-up exemption or the low OVB rate are justified. The throw-out reasons follow (in)directly from the law. If a return meets one of the throw-out reasons, there is a chance that the claim (exemption/low rate) is not justified. The processing should then reveal what is going on.

    Last change on 10th of September 2025, at 14:59 (CET) | Publication Standard 1.0
    Publication category
    Impactful algorithms
    Impact assessment
    Field not filled in.
    Status
    In use