What is a DPIA? A Data Protection Impact Assessment (DPIA) is a structured process for systematically mapping, assessing and reducing the privacy risks of processing data. A DPIA matters all the more when you use AI systems. Under the GDPR, the requirement is set out in Article 35.
Why is a DPIA important?
For professional offices, a DPIA is not a luxury but an obligation with real financial consequences. A DPIA serves three purposes at once. It helps you spot risks early, before they turn into problems. It gives you a paper trail that demonstrates accountability: if a regulator checks later, you can show you took this seriously. And it builds trust with your clients, because you are transparent about how you handle their data.
What many offices do not realise is that the risk of having no DPIA is considerable. Failing to carry out a required DPIA (Article 35 GDPR) falls under the fine category of Article 83(4) GDPR: up to 10 million euros or, for a business, 2% of worldwide annual turnover if that is higher. The higher category of 20 million euros or 4% applies to other infringements, such as processing without a legal basis. Worse still is the reputational damage if clients find out you never assessed the risks to their data.
When do you need a DPIA?
A DPIA is required as soon as your data processing is likely to result in a high risk to the rights and freedoms of individuals. In practice, offices quickly hit one of the triggers.
The first trigger is automated decision-making with legal effects. Imagine an accountant using an AI tool to sort clients’ bookkeeping automatically by risk profile. That is a DPIA situation, because the system directly determines how the client is treated. The same applies to profiling based on personal data: if you map client behaviour to sort clients into categories, you are in DPIA territory.
The second trigger is large-scale processing. Large scale is one of the criteria in the DPIA list of the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) and the European guidelines, and on its own it does not create a DPIA obligation; a DPIA is usually needed as soon as processing meets two or more criteria. According to Recital 91 GDPR, processing by an individual professional does not count as large scale, so hundreds of clients do not automatically make an office large scale. Novel technology is the third trigger. This means technology that is new by the state of the art, not a tool that is only new to your office, and this too is one criterion alongside the others.
A practical example for notaries: you start using an AI tool to draft wills more quickly. The tool processes names, addresses, family relationships and financial details. If you go through the criteria in the guidelines, two probably apply. The first is sensitive data or data of a highly personal nature. Besides the special categories, the guidelines include data linked to private life or whose misuse would seriously affect someone, such as financial data. The second is innovative use of new technology, which a generative AI model can fall under. According to Recital 91, large scale does not count for an individual professional. Nor does automated decision-making with legal effects, as long as you decide which clauses and warnings go into the deed. Such a tool does not appear as a separate category on the AP’s list. A DPIA is then probably required; record your assessment.
DPIA and the GDPR: the legal framework
The GDPR (General Data Protection Regulation) is the European data protection regulation; in the Netherlands it is known as the AVG. Article 35 GDPR sets out when you must carry out a DPIA. Paragraph 3 names three situations in which a DPIA is always required; for all other processing you weigh the criteria from the previous section.
First: when you systematically and extensively evaluate personal aspects of people based on automated processing, including profiling, and base decisions on it that have legal effects or a similarly significant impact. Storing client data centrally and using AI tools does not automatically fall under this.
Second: when you process special categories of personal data (Article 9 GDPR, such as health, religion or ethnicity) or criminal offence data (Article 10 GDPR) on a large scale. Financial data and household circumstances are not a special category, however sensitive they are for a notary.
Third: when you systematically monitor publicly accessible areas on a large scale. This is not relevant for most offices, unless your office has cameras pointed at public roads.
As the controller, you assess for yourself whether processing is likely to result in a high risk. In the Netherlands, the AP publishes a list of processing operations that require a DPIA under Article 35(4), but it does not decide case by case for each organisation. For AI systems, this often runs in parallel with broader obligations under the AI Act for businesses, which sets its own requirements alongside the GDPR.
DPIA for AI systems: what do you need to know?
AI systems make DPIAs more complex, because they introduce risks that traditional data processing does not have. A spreadsheet of client data is static; an AI model is trained and changes over time.
Automated decision-making is a major risk. AI makes decisions without human intervention. If you use AI to assess creditworthiness or sort clients into risk classes, you need to examine whether those decisions are fair and do not discriminate.
Bias and discrimination are a second concern. AI models are trained on historical data. If your training data contains discrimination (for example, men get more promotions, women more rejections), the model picks up that discrimination. For accountants this rarely looks dramatic, but subtle bias (for example, certain types of business being charged higher rates) can occur.
Model drift is a third risk. A model that is accurate today can perform worse tomorrow. Why? The training material ages, reality changes, the input data changes. This affects offices that rely on AI without regularly checking whether the model still works properly.
Explainability, the ability to explain what AI does, is another important consideration. If an AI says “no”, can you explain why? For many offices this also matters legally. A notary must be able to explain why she offers specific clauses; an accountant must be able to justify why she recommends a particular risk class. If you use AI and cannot explain what the AI did, you have a problem.
For AI systems with a higher risk level, a DPIA is sometimes not enough. An AIIA may then also be relevant. A legally required FRIA applies from 2 December 2027 only to bodies governed by public law and private parties providing public services, plus deployers of high-risk systems for creditworthiness or life and health insurance. Both instruments look more broadly than privacy alone.
How do you carry out a DPIA?
A DPIA consists of four steps. You will not finish them in a day, but they follow a logical order.
- Data mappingDescribe how data flows through your organisation and which AI tools see it.
- Identify risksFor each processing step: leaks, misuse, discrimination, unauthorised decisions.
- MitigateTechnical, organisational and contractual measures, plus regular bias testing.
- DocumentRecord the data, the risks and the measures as evidence of your work.
The real value lies in the conversation you have while filling it in, not in the template.
Start with data mapping. Describe step by step how data flows through your organisation. Who collects the data? Where does it go? Who has access? Which AI tools see this data? It sounds simple, but many offices do not know the exact answers. A notary, for example, does not always know that support staff run bulk exports of client data to a cloud tool. This is often the first eye-opener.
Follow data mapping with risk identification. For each processing step: which risks can arise? Can data leak? Can the system be misused? Does it discriminate? Can someone make decisions without authorisation? You do not do this alone, but together with the team that actually uses the data.
In the third step you mitigate: what can you do about these risks? Technical measures: encryption, access control, logging. Organisational: who may see what, who checks whom? Contractual: what is in your data processing agreement with suppliers? For AI specifically: can you test the model for bias regularly?
Finally, you document. Write down which data, which risks and which measures. This document is your evidence that you worked seriously, rather than just ticking off a checklist.
Templates help, but every office is different. The AP offers guidance; some AI tool suppliers also provide templates. The real value is not in the template, but in the conversation you have while filling it in.
Where can you get help with an AI risk analysis?
You have several options. Carrying out a DPIA yourself is possible, especially if your office is small and the data processing is not complicated. But as soon as AI comes into play, the complexity grows.
Doing it yourself with a template: free, but you miss out on insight into risks specific to your office. An external consultant: billed by the hour, which can add up, but gives you an independent perspective. Tools and software: many AI tool suppliers include DPIA templates.
A practical starting point is your own inventory. It is not a full DPIA and does not replace one, but it is a risk map: which AI tools you have, which risks they bring and where the blind spots are. That gives you the information you need for a targeted, much quicker DPIA. For the people who will work with those AI tools every day, the AI Literacy Course gives a shared basis in using AI responsibly at work.
Frequently asked questions about the DPIA
When you carry out a DPIA, questions often come up about who should be involved and what happens once it is finished.
Who should you involve in a DPIA?
In a DPIA you always involve the data protection officer, if your organisation has designated one: Article 35(2) GDPR requires you to seek their advice. The controller remains responsible for the assessment itself. Article 35(9) adds that, where appropriate, you seek the views of data subjects or their representatives on the intended processing.
What do you do if a DPIA shows a high residual risk?
If a DPIA shows a high residual risk that you cannot reduce with measures, you consult the supervisory authority before processing; in the Netherlands, that is the AP. Article 36 GDPR calls this prior consultation. If the AP considers that the processing would infringe the GDPR, it will in principle give written advice within eight weeks and can use its powers. So plan that consultation before you start using an AI tool.
How often should you review a DPIA?
You review a DPIA at least whenever the risk of the processing changes, as Article 35(11) GDPR states. The GDPR does not set a fixed period. With AI tools, the risk can change, for example, when you switch to a different model or when new types of data go into the tool. Record each review in the DPIA document, so you can show which assessment applied at which time.
Is there a free DPIA template?
One free DPIA template is the Dutch government’s Model DPIA Rijksdienst (in Dutch), available from the Dutch government’s knowledge centre for policy and regulation (Kenniscentrum voor beleid en regelgeving). The model was made for central government, but its structure is also usable outside government: first the facts of the processing, then its lawfulness, then the risks to data subjects and finally the planned measures. The GDPR does not prescribe a specific tool, but Article 35(7) does set out the minimum content of a DPIA.
