Back to the blog

Introducing AI with your works council: when you need consent and what to bring

· By Bee Flow · 6 min read

On 14 September 2026 the Dutch union FNV decided to bargain for a say over AI for the first time. For a company of 30 to 200 people the Works Councils Act already covers most of it: when article 27 requires consent, what to bring to the first conversation, and how to turn it into a short agreement.

A long meeting table with chairs on both sides, a notebook at each side and one sheet of paper with an amber fountain pen in the middle

On 14 September 2026 the federation council of the Dutch trade union FNV adopted its wage demand and its bargaining agenda for 2027. Next to pay, the union says that this year, for the first time, it will bargain at the collective agreement (cao) tables for a say for employees over AI. The agenda itself spells out the aim: when new technology that touches the future of a sector is introduced, the unions want cao clauses saying that employees, works councils and unions have a say in it.

How that lands in agreements is open. If you run HR or the company itself, with 30 to 200 staff, and want to introduce AI now, you do not have to wait. The Works Councils Act (Wet op de ondernemingsraden, WOR) already covers most of it. This article is not legal advice: put your own situation to an employment lawyer or your data protection officer.

Who you talk to depends on your size

If you normally have 50 or more people working for you, article 2 of the WOR obliges you to set up a works council (ondernemingsraad, OR).

Between 10 and 50 people it works differently. Without a works council, article 35b says you must give staff the chance to meet with you at least twice a year. You may set up a staff representation (personeelsvertegenwoordiging), and must if a majority of staff asks (article 35c). That body has a say over a shorter list: working and rest hours, health and safety and sick leave, and the whistleblower procedure. The two items on personal data and monitoring, where AI usually falls, are not on it.

Legally, you then have more room. The same people will still use the tool, and taking a staff meeting through the questions below costs an hour.

When article 27 requires consent

Article 27(1) of the WOR lists the arrangements for which the employer needs the works council's prior consent. Two almost always apply to AI (our translation):

  • under k: an arrangement on the processing and protection of the personal data of the people working in the company
  • under l: an arrangement on facilities aimed at, or suitable for, observing or checking the attendance, behaviour or performance of the people working in the company

The word that matters is "suitable". The Dutch data protection authority (Autoriteit Persoonsgegevens, AP) explains it in its booklet on works councils and privacy: if a system can be used to follow staff, even if that does not happen in practice (yet), it is still a staff monitoring system. Its examples include software that records keystrokes, e-mail or internet use, and logging of who opens which files. An AI chat with personal accounts and a usage overview per person fits that description easily.

If you use AI in performance reviews or in hiring, items g (staff appraisal) and e (appointment, dismissal and promotion policy) come into play too.

Skip the consent and, under paragraph 5, the decision is void if the works council invokes that in writing within a month. If the council refuses, you can ask the cantonal court for permission (paragraph 4), which it gives only for an unreasonable refusal or weighty business reasons.

Paragraph 3 shows why the FNV agenda matters. Consent is not required to the extent that a subject is already settled in substance in a cao. If AI clauses end up in your cao, that part moves to the bargaining table. What the cao leaves open stays with you.

When advice or information is enough

If you plan to introduce or change "an important technological facility", the works council has a right to advise under article 25(1)(k). You ask for that advice at a point where it can still have a substantial influence on the decision, so before the licence is signed. On request you give the council, in time, the information it reasonably needs (article 31).

Advice or information alone is enough when a tool processes no personal data of staff and is not suitable for following them. With an AI chat that has user accounts, we rarely see that. Our view: for such a tool, assume consent is needed and simply ask for it. The conversation beforehand is shorter than a dispute afterwards.

What to bring to the first conversation

A works council consents to an arrangement, not a product name. Bring answers to the questions it will ask anyway:

  • Which tools, what they may be used for, and who gets access.
  • Which data may go in: customer data, data about colleagues, confidential documents. Where the model runs, with which provider, and whether there is a data processing agreement.
  • What is recorded: conversations, usage per person, alerts from a filter, sign-ins. And how long it is kept.
  • Who can see it: which admin roles exist, whether a manager can see individual use, and in which cases someone may look.
  • What it will not be used for, such as judging performance.
  • The DPIA, if you need one. The same AP booklet says that large-scale or systematic processing of personal data to monitor employees' activities is on its list of processing that requires a DPIA. Share it with the council.

Also show the screen where an administrator looks at usage. It says more than a paragraph of policy.

A laptop on a meeting table facing a row of empty chairs, its screen showing a bar chart with one amber bar, next to a small stack of papers

From conversation to a short agreement

An AI arrangement does not need to be long. A made-up example, for Example Installations Ltd with 80 staff:

  1. Scope: the company's AI chat, for drafting, summaries and searching the company's own manuals.
  2. Data: no citizen service numbers (BSN), medical data or salary data.
  3. Records: an appendix lists what the tool logs per person and for how long.
  4. Access: administrators look at usage per team. Individual use only on suspected misuse or a security incident, after telling the council chair, with a note of who looked and why.
  5. No appraisal: AI use does not come up in performance or appraisal reviews.
  6. Review after six months. A new tool or a new kind of recording goes back to the council.

This differs from the policy that tells staff what they may do with AI. We wrote about that earlier in a one-page AI policy. The arrangement sets limits for the employer, the policy for the user. You need both.

What you can show in Bee Flow

The admin area of Bee Flow has a usage overview that shows per person which model was used, how much, and from which part of the product. That is the overview you put on the table under article 27(1)(l). If you also make meeting notes from recordings, name them in the arrangement too, because a recording holds what everyone at the table said. Privacy Shield checks a question for personal data before it goes to a model, with detection running on the Bee Flow server, and records which kind of data it found and what it did, without the text of the message. You choose where the model runs: a cloud provider, an EU provider or your own hardware. The Privacy & Security page explains how that works. If you want to walk through the admin screens with your works council before you decide, get in touch for a demo.

#ai-at-work #compliance