Matching jobseekers with vacancies

Municipality of Delft

Matchplaats (AI) supports the employment consultant in matching jobseekers with vacancies. The tool generates a list of potential matches between jobseekers and vacancies based on their skills, characteristics, interests and work-related points of interest. The work consultant discusses the potential matches deemed suitable with the jobseeker. If interested, the jobseeker is introduced to the employer.

Last change on 13th of May 2026, at 7:05 (CET) | Publication Standard 1.0

• Publication category: High-Risk AI-system
• Impact assessment: DPIA, IAMA
• Status: In use

General information

Theme

Social Security

Begin date

2026-04

Contact information

[e-mail] (Werkse! is onderdeel van gemeente Delft)

Link to publication website

www.werkse.nl (Werkse! is onderdeel van gemeente Delft)

Responsible use

Goal and impact

The algorithm used in the AI tool supports two main goals with the following impact:

1. More efficient and better matching: The algorithm helps achieve faster and better matches by matching candidates with jobs that match their capabilities and competences, where their skills, attributes, interests and concerns related to work are the basis and not just work experience or qualifications.

2. Inclusion and fairness: The system is designed to minimise bias and provide equal opportunities, offering transparency and explainability, where in the current manual process it cannot always be made transparent (such as unconscious biases that may play a role).

The impact for a job seeker may also lead to exclusion or discrimination, this depends on how the algorithm is set up. To identify and prevent possible exclusion or discrimination, a number of measures have been put in place, such as bias monitoring, consultants are trained to investigate and recognise bias, consultants always apply manual validation, a feedback mechanism has been put in place, consultants are instructed how to pass on any bias detected and the functioning of the algorithm has been analysed by the FG and the CISO of Werkse!

Considerations

The use of the tool is an important means of implementing matching as effectively and fairly as possible.

Benefits for the jobseeker:

1. Increasing opportunities through scalability: The volume of both available vacancies and jobseekers is so large that it is virtually impossible for consultants to manually oversee all possibilities and identify the best matches. The tool makes it possible to objectively and accurately analyse a wide range of vacancies and job seekers within a very short period of time

2. Fairer treatment through bias monitoring: Unlike the manual matching process, in which unconscious biases of consultants can play a role, the tool is designed according to FAIR principles (transparent, objective and explainable). In addition, results are continuously monitored for possible bias which contributes to a more inclusive and equitable matching process

3. Personalised and high-quality counselling: By using the tool, the counsellor can focus on counselling the jobseeker and critically assessing the proposed matches instead of spending time searching for vacancies. This leads to more personalised and high-quality counselling

4. Optimisation through data-driven decisions: The tool supports counsellors with data-driven insights, making matches not only faster, but also more responsive to the jobseeker's skills, interests and needs. This increases the likelihood of long-term and successful placements

Jobseekers are entitled to protection of their personal data and fair, transparent treatment in the matching process. Their privacy is ensured through strict data minimisation, organisational and technical security measures and monitoring that minimise the risk of unauthorised access or misuse. Only data necessary to achieve a good match is used for matching job seekers to vacancies. Special personal data are not used for matching. The jobseeker's data are removed from the tool for matching purposes after a jobseeker's journey is concluded. Jobseekers can also indicate that they do not want to use Matchplaats. In that case, the 'opt-out' function is switched on so that the jobseeker's details do not appear in Matchplaats and matching takes place in the old manual way.

Human intervention

The work consultant assesses the matches proposed by Matchplaats. The work consultant discusses the potential matches found suitable with the jobseeker. If interested, the jobseeker is introduced to the employer. The AI tool supports consultants through automated processing, but does not make decisions. Thus, as in the manual process, the decision to start an application process is made together with the jobseeker.

Risk management

To reduce the risks of discrimination, exclusion and unlawful processing of personal data, technical and organisational measures have been taken to ensure a secure, accountable, fair and transparent matching process. The following organisational measures have been taken:

- Users of the AI tool have received AI literacy training and have attended user training Matchplace

- There is a process for authorisations to issue and revoke authorisations

- There is a process for secure development, version control and vulnerability management

- Periodic bias reports are prepared and discussed with management and the FG of Werkse!

- Users of the AI tool are trained to investigate and recognise bias and there is a process to pass on and record signals of bias

- Access to data is on the basis of individual accounts with role-based access control

- Only data necessary for achieving a good match is processed and the job seeker's data is deleted from the AI tool for matching purposes after a job seeker's journey is closed

The following technical measures are in place:

- The stored data is encrypted (AES-256) both in transit and at rest

- The data is processed within a secure network with end-to-end encryption

- The Matchplace application supports Single Sign-On (SSO) via Microsoft Entra ID

- Access and system actions are logged

- There is a backup and recovery procdure where data is stored in Azure Blob Storage and PostgreSQL databases (NextHuman's Azure Cloud within the EU). Daily backups are kept for 35 days. Backups are encrypted both during storage (AES-256) and transport (TLS). Only authorised persons have access to the backups. The backup and restore procedure is tested quarterly

- 2-factor authentication (2FA) is used to access the system and communication between systems is secured via a private VNet

- The AI tool is designed to be transparent to users and stakeholders. The tool includes an explainability component, allowing consultants and jobseekers to understand the reasoning behind the potential matches generated.

Legal basis

The municipality is legally obliged to support jobseekers, with a distance to the labour market, in finding suitable work on the basis of the Participation Act (chapter 2 art. 7, 8a, 9) and the AVG (art. 6.1)

Links to legal bases

• Participation Act: Hoofdstuk 2 art. 7, 8a, 9
• AVG: Art. 6.1

Elaboration on impact assessments

A DPIA (Data Protection Impact Assessment) has been carried out to identify the risks to job seekers' privacy. This DPIA will be updated at the end of the pilot and will then form a basis to take a go - no-go decision on commissioning the AI tool Matchplace. An IAMA (Human Rights Assessment) was also conducted to test for discrimination and equal opportunities. Key elements from the IAMA are also included in Chapter 3 of the DPIA. Key measures taken to mitigate risks in this context are:

- Man in the loop: the consultant always checks AI matching suggestions and discusses the suitable potential matches found with the job seeker. If interested, the jobseeker is introduced to the employer

- Explainability: for every match made, it is clear how it came about. This is clear to both the jobseeker and the consultant

- Bias monitoring: any bias in the matching process is continuously identified and reported at least quarterly and tightened where necessary

Impact assessment

• Data Protection Impact Assessment (DPIA): Matching kandidaten met vacatures met behulp van AI
• Human Rights and Algorithms Impact Assessment (IAMA): Matching kandidaten met vacatures met behulp van AI

Operations

Data

Only data necessary to achieve a good match is used for matching jobseekers to vacancies. In the AI tool, data such as:

- Name

- Sector and job preferences

- Language skills

- Courses followed

- Computer skills

- Explanation of social network

- CV

- Possession of driving licences & travel opportunities

- Specific guidance needs or workplace requirements

- Points of attention that affect workload capacity and work-fitness (described in functional terms in line with the AVG)

- Any indication of job match, day care or informal care

- Interview and observation reports.

Technical design

Matchplace uses 4 different algorithms: 1. From client info (Afas) → matching profile

- Input data: data from jobseeker's master card in source system, CVs, interview and observation reports, education, hobbies & interests, guidance needs, workload capacity

- Model: GPT-4o (transformer). It extracts relevant job-related skills, characteristics, interests and points of interest from unstructured text

- Output: a structured matching profile with objective characteristics that serves as input for matching

2. From vacancy (Afas, Indeed, Glassdoor, LinkedIn) → vacancy profile in Matchplaats

- Input data: vacancy texts and requirements, company profile

- Model: same stack - GPT-4o: recognises skills/requirements from the text

- Output: a job profile with objective features (skills, attributes, interests, points of interest)

3. Matching client ↔ vacancy in Matchplaats

- Input data: candidate's matching profile and vacancy profile

- Model: Word2Vec + BM25 (statistical scoring model): Word2Vec puts words in vector space (semantic similarity), BM25 weights keywords and reduces noise (words like "flexible" are given less weight), a matching score is calculated from a weighted average of these two

- Output: a ranking of vacancies (or candidates when mirroring). The explainability component shows the weight in the score for each criterion (e.g. "creativity", "no lifting")

4. Bias monitoring tool

- Input data: Results of all matches made and is more for bias measures around gender, age, language. Batch results from running the entire jobseeker database and (a sample of vacancies) and is more for false positive/negative detection

- Model: statistical monitoring: analyses of differences in match rates by age, gender, language and sector, as well as distributions of jobseekers or vacancies in the outcomes

- Output: alerts in case of significant deviations, e.g. if a group is structurally matched less often or if a group is pushed more often in a specific sector. GPT-4o (Transformer Architecture) and Word2Vec (embedding model) are self-learning algorithms and BM25 (Best Matching 25, information retrieval, specified formulas combined with statistics from the data) is a semi-self-learning algorithm.

External provider

NextHuman