A DPIA for AI chat in a small company
· By Bee Flow · 4 min read
When a data protection impact assessment is required for an AI chat, what goes in it, and how a small company keeps it proportionate.

An installation company with forty staff wants to bring in an AI chat for quotes, service mails and summarising work orders. The managing director asks the office manager, who also handles privacy, whether they need a DPIA first. She has never written one and finds examples online that run to sixty pages.
The short answer: a data protection impact assessment (DPIA) is required when processing is likely to result in a high risk to the people whose data it is. That is Article 35 of the GDPR (the AVG in Dutch). An AI chat for ordinary writing tasks without sensitive data need not fall under it. Once health data, data about vulnerable people or decisions about individuals are involved, it quickly becomes required. If you decide against a DPIA, you have to be able to explain why.
This article is not legal advice, and your own situation may differ. We checked the guidance below against the sources themselves on 14 November 2026.
When it is required
The Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, uses nine criteria that the European data protection authorities drew up together in the WP248 guidelines, later endorsed by the European Data Protection Board:
- evaluation or scoring
- automated decision-making with legal or similarly significant effect
- systematic monitoring
- sensitive data or data of a highly personal nature
- processing on a large scale
- matching or combining datasets
- data about vulnerable people
- innovative use or new technology
- processing that prevents people from exercising a right or getting a service or contract
The AP's rule of thumb: if your processing meets two or more criteria, you have to do a DPIA. An AI chat can count as new technology. If yours does, and staff also paste client files with medical information into it, you are at two.
The AP also has a list of seventeen kinds of processing that always need a DPIA. Monitoring employees is one of them.
A trial counts too. In its guidance on AI and algorithms the AP says the DPIA obligation also applies to pilots and tests. And the DPIA is finished before the processing starts.
What goes in it
Article 35(7) names four parts. For an AI chat it could look like this.
A description of the processing. Who uses the chat, for which tasks, what data ends up in it, which language model answers and where it runs.
An assessment of necessity and proportionality. Can it be done with less data? Does the chat need to see the customer's name to write a polite mail? How long are conversations kept, and why that long?
The risks to the people concerned. Data that ends up with a provider you have no agreement with. Conversations used for training. Answers that state something wrong about a person. Colleagues seeing conversations that were not meant for them.
The measures. A business account with a data processing agreement, written rules on what may and may not go in, a retention period, a filter that holds back personal data, and an explanation for the team.
If you have a data protection officer, ask for their advice. If a high risk remains after the measures, you have to consult the AP beforehand (Article 36), and according to the AP you may only start after that. And when the risk changes, because you pick another model or add new tasks, you review the DPIA.
How to keep it proportionate
One assessment may cover several similar processing operations with similar risks. If three departments use the same chat in the same way, you write one.
Use what others have done. SURF, the IT cooperative of Dutch education and research, published a DPIA on Microsoft 365 Copilot and has updated it twice. In the second update two medium risks remain, one of them the processing of possibly inaccurate personal data. That analysis is about the education version, but the kinds of risk are familiar. Is Microsoft 365 Copilot GDPR compliant? sums up what SURF and the Dutch regulator concluded. The Dutch government has a model DPIA (in Dutch) with seventeen parts. And the European Data Protection Board (EDPB) had an external expert write a report on the privacy risks of language models, which you can use as a list of risks and measures. That report does not replace a DPIA.
Start small. A trial with one team, without the most sensitive files, with a start date, an end date and a point where you evaluate, is easier to assess than a rollout across the whole company.
DPIA and register in the same system
The Compliance Center in Bee Flow has a DPIA section per assistant. Bee Flow flags assistants that point to a high risk, for instance because they work with certain kinds of personal data or use an external model. Per assistant you fill in a short questionnaire, or confirm that the assessment has been done, and you export the result as a PDF. When a DPIA expires within thirty days, the administrators get a reminder. The record of processing activities sits next to it.
If you run the chat on your own server with the filter that holds back personal data, those are measures you can write down in the DPIA. The Compliance Center is not in every licence; see the plans page.
More about the chat is on private AI chat, and about security on privacy and security, with the GDPR side in plain words on security, privacy and GDPR. Whether ChatGPT is safe at work in the first place is covered in is ChatGPT safe to use at work. Questions can go to contact.