Experiment Microsoft Copilot 365

Municipality of Zaanstad

Explore the use of Microsoft Copilot 365, integrated into Office applications (including Word, Excel, Outlook, Teams) and the municipal work context, to investigate whether employees can work more efficiently and reduce administrative burdens.

Last change on 13th of November 2025, at 12:04 (CET) | Publication Standard 1.0

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

General information

Theme

Organisation and business operations

Begin date

2025

End date

2026

Contact information

https://www.zaanstad.nl/contact/

Responsible use

Goal and impact

The experiment investigates whether Copilot 365 increases staff productivity by reducing administrative burdens (such as documentation, reports and emails), thus freeing up more time for substantive tasks and better services to residents.

Considerations

When deploying Copilot 365, there are clear ground rules for responsible use and careful handling of data. Employees always remain responsible for the final choices and content, and assess whether the output generated by Copilot is correct and usable. Privacy and information security are safeguarded through the DPIA, agreements with the supplier and our ground rules.

Human intervention

Always: employee decides whether Copilot output is used, modified or rejected.

Risk management

In this experiment, we looked specifically at the requirements of the European AI Regulation. In doing so, we fit in with our existing policies on privacy, information security and procurement, as well as topics such as awareness, code of conduct and use of cloud services. To provide employees with guidance, rules have been drawn up for the responsible use of generative AI: what is and is not possible. Furthermore, we work on awareness, communication and training so that colleagues become familiar with AI and know how to use it responsibly.

Elaboration on impact assessments

The risks were assessed by the Data Protection Officer (FG) for relevance and applicability when using AI. Based on this assessment, and in the light of the publications of VNG and IBD, it was indicated that the AI project can be continued responsibly, provided the emphasis remains on careful and explainable use.

Impact assessment

Data Protection Impact Assessment (DPIA): https://www.surf.nl/nieuws/privacyrisicos-microsoft-365-copilot-naar-oranje

Operations

Data

Copilot uses all the data within the Microsoft 365 graph to which an employee himself has access (such as emails, documents and Teams messages). This data is converted by Microsoft into a semantic index, allowing Copilot to make connections. Copilot sees data only when an employee actively asks for it in a prompt. The data does not leave the Microsoft tenant and is not used to train AI models.

According to the municipality's Generative AI game rules, when an employee enters a prompt, Copilot may only be used with unstructured work data.

Not allowed: personal data and other classified or sensitive information (such as payroll, healthcare or criminal records).

We intend to exclude certain SharePoint channels from indexing to avoid risks of AI misuse and non-compliance with the rules of the game.

Technical design

Copilot uses Large Language Models (LLMs) integrated into Microsoft 365. When an employee enters a prompt, Copilot combines this question with relevant context from the Microsoft Graph and the semantic index. Only information to which the employee himself has access can be used in the answer.

The processing of prompts and context takes place within the secure Microsoft tenant. Copilot then generates a response based on the prompt, the context and the underlying AI model. The temporary prompt data is not stored and not used for training the model.

Although Copilot provides support, human control remains necessary: Copilot may make mistakes or misinterpret information.

External provider

Microsoft (Copilot 365)