Back to the blog

The AP's AI risk report of March 2026: four checks for AI chat at work

· By Bee Flow · 4 min read

On 5 March 2026 the Dutch data protection authority published its sixth report on AI and algorithm risks. What it says in its own words, and four things a small company that uses AI chat can check calmly.

A wall panel with nine round gauges, four of them turned up and lit in amber, above a desk with a four-item checklist and a pen

On 5 March 2026 the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, published the sixth Report AI & Algorithms Netherlands, known as the RAN. The AP publishes it twice a year in its role as coordinating supervisor for algorithms and AI. The press release leads with a count: four of the nine indicators in its AI Impact Barometer are now red, up from two in the previous edition.

Most of the message is addressed to the new Dutch government. The AP wants it to adopt the national implementing act, designate supervisors and put the funding of supervision on a structural footing. Its chairman, Aleid Wolfsen, says in the press release: "Anyone who wants to prevent a new scandal must act now."

For a small company that uses ChatGPT, Copilot or a similar AI chat, the report adds no new obligation. It does contain observations you can use to look over your own use calmly. This article is not legal advice.

What the AP writes

The passages that matter most for office use, with page numbers from the English edition.

On the quality of answers: AI chatbots "gave answers with errors more often than a year ago, and used sources incorrectly" (p. 10). On hallucination, the AP finds that "New models do not consistently hallucinate less. Some new models hallucinate less, others more" (p. 10).

On AI agents, systems that carry out actions themselves: "In the design of these systems, it is therefore important to consider what the system gets access to, and what actions it can perform" (p. 15). The AP calls limiting an agent to one specific task an important good practice. On 12 February it had already warned about the security risks of AI agents such as OpenClaw.

On smaller organisations: the frontrunners on transparency are often in the public sector, "while smaller organisations tend to lag behind" (p. 4). The AP links this to limited capacity for organising control of new technology.

On incidents, the indicator reads "insufficient progress". Incidents, the report says, are not yet used sufficiently as learning opportunities and often stay out of sight (p. 7).

On recruitment: AI systems used for it are high-risk under the AI Act. The press release says the strict requirements apply from August 2026. The report itself is more careful and says "2026 or 2027", depending on European negotiations, and notes that the European Commission has proposed to partially postpone application by about one to one and a half years (p. 4 and p. 44). That is a proposal. Nothing has been decided.

Four things to check

1. Who checks the answer?

The AI literacy duty has applied since 2 February 2025, the AP reminds readers (p. 44): employees should "continue to critically assess the output of AI". So agree which answers never leave the building unchecked, such as a quote, an answer to a legal question or the figures in a report. And prefer answers that say where they came from: one with a source takes a minute to check.

2. What can the AI reach, and what can it do?

More and more chat tools connect to mail, calendars and shared folders. Make a short list per connection. What data can the tool read? Can it send, change or delete anything on its own? That is exactly the question the AP asks about agents. A reasonable starting point is that reading is allowed and acting waits for a person to click. Put on the same list which personal data reaches an AI vendor through that connection, and whether a data processing agreement covers it.

3. Where do you write down what goes wrong?

When the AP says incidents stay out of sight, it mostly means large systems. The same pattern shows up on a small scale: a wrong answer that went to a client, a client's e-mail pasted into a free chat service. A shared document with the date, what happened and what you changed is enough. It also helps you notice in time when something is a data breach you must report under the GDPR. A one-page AI policy is a good home for that log.

4. Do you use AI when hiring?

A small employer that has a chat tool summarise or rank CVs is already close to what the AP describes. According to the AP, organisations can be expected to meet the requirements on the right to explanation, transparency and meaningful human intervention already now, because these follow from the GDPR (p. 44). From 2 August 2026, the report adds, you must also make it clear when an applicant interacts with AI. So let a person make every rejection, and tell applicants beforehand that AI is involved.

How we read it

We do not read the report as an alarm for small companies. The red indicators concern the set-up of supervision, standards, the registration of government algorithms and visibility of incidents. Those are tasks for The Hague and Brussels.

For a small company one plain question remains. Do you know what your AI tools are allowed to do, and who checks their answers? On recruitment we would wait for the outcome of the postponement proposal before changing any plans. The GDPR points apply today, postponement or not.

What Bee Flow does about it

Bee Flow is an AI workspace that can run on your own server. Answers from a knowledge base point to the document they came from, so a colleague can open the source in one click. Privacy Shield checks messages for personal data before they reach a language model. Every agent works with a fixed set of tools that you choose, and you decide which models your organisation uses, including a European provider. How this fits together is on the privacy and security page. If you want to hold it against your own checklist, get in touch.

#compliance #privacy