Reporting Assistant Know Your Customer

Municipality of Amersfoort

The AI assistant processes recorded audio from a conversation into an anonymised report (pilot). The Ken je klant process is carried out by the WIZ department of Amersfoort municipality. An interview with residents of Amersfoort and Leusden receiving benefits under the Pw or IOAW takes place periodically.

Last change on 28th of October 2025, at 12:33 (CET) | Publication Standard 1.0

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

General information

Theme

Health and Healthcare

Begin date

2025-09

Contact information

[e-mail]

Responsible use

Goal and impact

The aim of deploying the AI assistant is to reduce the administration burden among employees. By deploying the tool, a time saving of 50-75% is aimed at processing interviews into a draft report. The AI assistant does not make independent decisions; an employee always checks the outcomes. Conclusions are always formulated by the employee. Every report, including those not prepared through the AI Assistant, is submitted to residents for approval. Participation is voluntary for residents and is explicitly requested prior to the interview. Use of the tool is separate from (access to) services.

Considerations

To make a careful consideration for the deployment of the AI assistant, a short DPIA was conducted, an internal values dialogue was held, and previous advice from the external sounding board group on digitisation and ethics (from pilot project reporting assistant at neighbourhood teams) was followed. Prior to the decision to start this pilot, other steps were taken to significantly improve administrative pressure for staff, but this did not yield sufficient results.

Human intervention

The (interim) results of the AI assistant can be read along in real time by the employee and resident. If the report is generated, the employee can modify, delete or add any generated text.

Risk management

Through the DPIA carried out, the values dialogue conducted and the advice of the Digitalisation and Ethics Sounding Board Group, data protection and ethics risks have been identified. Among the measures taken to manage risks are the following: employees have been trained to work with the AI Assistant. There is an explicit focus on being able to properly explain to residents what this AI assistant does, and guidelines have been drawn up for its use. Furthermore, agreements have been made with the supplier about the processing of personal data, the right level of security has been applied to the system and this has been independently tested. Furthermore, only the data needed for the process is kept. There is constant monitoring of the AI assistant's performance.

Legal basis

The processing of personal data is necessary for the fulfilment of a task of general interest (execution Participation Act and IOAW) that has been assigned to the controller. The interview is intended to support residents and do what is necessary within their capabilities and the law to promote self-reliance. The approach of the interview is mainly service-oriented but can also be of a legitimacy and/or efficiency nature.

Links to legal bases

• Article 9 and 17 PW: https://wetten.overheid.nl/BWBR0015703/2025-07-01
• Section 34 IOAW: https://wetten.overheid.nl/BWBR0004044/2025-07-01
• Article 6, first paragraph under c AVG: https://eur-lex.europa.eu/legal-content/NL/TXT/?uri=legissum:310401_2

Elaboration on impact assessments

The AI Assistant is not intended to be an AI system for assessing (the degree of) access to essential government benefits and services. In particular, its purpose is to provide better and faster assistance to residents and reduce the administrative burden. As a result, the application falls under one of the exceptions in the AI Act (Section 6(3) AI Act) where AI systems that fall under one of the application areas are nevertheless not considered high-risk AI (Essential private and public services and benefits (Section 6(2) in conjunction with Annex III AI Act)) and conducting an IAMA test is not mandatory. An internal ethical values dialogue on public values was conducted using the Code for Good Digital Public Administration (CODIO). In addition, the municipality received an opinion from the external sounding board group on digitisation and ethics in April 2025 (residents).

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

The algorithm was developed with a description about the conversation protocol. Simulated conversations between staff members were used to refine the prompting model. The algorithm processes only data obtained from an audio recording of the conversation between the resident and the staff member. Personal data that could potentially be processed are Name, address, place of residence, BSN, information on work and income, housing, health, social relations, day care, safety, self-reliance, meaning. Directly traceable personal data is deleted during the conversion of every 30 to 60 seconds of audio to text. Data such as family situation, health data and other special personal data that are mentioned in the conversation and may be relevant to provide the necessary support to the resident will be processed on the supplier's platform and stored for a maximum of 28 days.

Technical design

The AI assistant is a tool to convert the spoken word during a 'know your customer' conversation into written record. The software converts audio to a written-out, de-indented transcript. This transcript is presented to several locally hosted copies of LLMs (Large language Models) for writing the conversation report. To ensure that these reports are written in the desired form, the software contains specific rules about the prompts offered to the LLMs. The software forms, as it were, an intermediate layer between the interview transcript and the LLMs that ensures that the ouput of the LLMs better matches the district teams' method of reporting than if one were to offer a transcript directly to an LLM.

The software uses web services and locally hosted copies of LLMs. The following models are used: speech modelling services 'Azure Language' and speech-to-text services (Microsoft), GPT4o and GPT4o mini from OpenAI, Claude Sonnet from Anthropic and Llama (open source). LLM providers (e.g. OpenAI) do not connect to the servers of the vendor platform. LLM provider may provide the models, but hosting is done independently on supplier servers within the EEA. As a result, the data is not accessible or viewable by LLM providers.

External provider

Verticai B.V.