DPIA for generative AI: when it is required and how

A data protection impact assessment (DPIA) is required when at least two of the nine criteria on the list of IMY, the Swedish Authority for Privacy Protection, are met. A generative AI tool in legal work meets the new technology criterion, and in employment matters a second one too. The assessment should be done before first use, because the answer decides which matters can be entered. This guide covers the criteria, what the assessment must contain at a minimum, and how a department of one to ten lawyers can do it without a long document.

Updated by Nils-Erik Jansson

When it is required

Article 35 GDPR requires an impact assessment before a type of processing that is likely to result in a high risk to the rights and freedoms of natural persons, in particular when new technology is used [2]. IMY's list under Article 35(4) has nine criteria, and the rule of thumb is that an assessment is required when at least two are met, and in individual cases when only one is [1]. The criteria are: evaluation or scoring of people, automated decisions with legal effects, systematic monitoring, sensitive personal data, processing on a large scale, combining data from several processing operations, people in a weak or dependent position, new technology or new organisational solutions, and processing that prevents access to a service or a contract [1].

Put a generative AI tool used in legal work next to the list. New technology: yes, almost always. Sensitive data: yes in employment matters, in disputes about health or trade union membership, and in many law firm matters. Dependent position: yes when the data concerns employees. Large scale: yes if the archive is connected to the tool. Combining processing operations: yes when the contract archive, email and matter management system are connected to the same tool. Two criteria are easily met. The conclusion is that an impact assessment is required for the matter types where two criteria are met, and that it is simpler to do one for the tool, with limits per matter type, than to try to avoid it.

What it must contain

The minimum under Article 35(7), as IMY sets it out: a systematic description of the processing and its purposes, an assessment of necessity and proportionality, an assessment of the risks to the rights and freedoms of data subjects, and the measures planned to address those risks. The data protection officer must be consulted, the views of data subjects sought where appropriate, and the assessment reviewed [2].

For an AI tool, the four parts look like this.

The description. Which tool, which plan, which connections, which matter types, which categories of data subjects (employees, customers, counterparties, clients) and data, where the processing takes place and at which sub-processors. Here you reuse step one of the transfer assessment; see the guide on transfer impact assessments.

Necessity and proportionality. Why the tool is needed for the purpose, whether the same thing could be done with less data, and which legal basis the processing relies on for each matter type. This is where the department has to answer whether every matter type really needs to go in, and where the list of matters kept out gets its justification.

The risks. To the data subjects, not to the company: unauthorised access, reuse for training, incorrect output leading to incorrect decisions about a person, disclosure following an authority request, and retention longer than necessary. Likelihood and severity for each risk, briefly.

The measures. The agreement and its addenda, the retention periods, the permissions, the list of matters kept out, the review rule in the policy, and the training. And the residual risk after the measures, described honestly.

How to keep it short

One assessment per tool, with a table per matter type, is the format that works. The table has four columns: matter type, data involved, criteria met, and conclusion (entered, entered with conditions, kept out). The description and the measures are written once and apply to every row. The table is the part of the document used in day-to-day work. This is what the rows look like when filled in from the criteria above.

Matter type Data involved Criteria met Conclusion
Employment matters Data about employees, sensitive data New technology, sensitive data, dependent position Decided per tool
Disputes about health or trade union membership Sensitive data New technology, sensitive data Decided per tool
Many law firm matters Sensitive data New technology, sensitive data Decided per tool
Archive connected to the tool Data across the whole archive New technology, large scale Decided per tool
Contract archive, email and matter system connected to the same tool Data from several processing operations New technology, combined processing Decided per tool

In the conclusion column, you choose between entered, entered with conditions and kept out.

The assessment can be reused. The next tool gets the same structure, and half the content is the same. The next matter type is one more row. The GDPR guidance of the Swedish Bar Association (Sveriges advokatsamfund) treats the impact assessment as part of a firm's ordinary data protection work, not as a one-off project [3].

The data protection officer and IMY

The data protection officer must be consulted, and the document must show what the officer said. If the company has no data protection officer, record who held that role in the assessment.

Prior consultation with IMY is requested only if the assessment shows a high risk despite the measures [2]. If a legal department ends up there, it should consider moving a matter type to the kept out list, since the need for consultation then disappears.

HOVR Legal carries out the impact assessment where one is needed, as one of five parts of the data protection package, in your template and with a review with your data protection officer. For Claude, it is part of the setup, with the description of the processing already written from Anthropic's documentation. What else needs to be in place before the first matter is entered is in the checklist.

Frequently asked questions

Is a DPIA required for every AI tool?
No, only for processing that is likely to result in a high risk. IMY's rule of thumb is that an assessment is required when at least two of the nine criteria on its list are met, and in individual cases when only one is [1]. A generative tool almost always meets the new technology criterion. If it is also used for sensitive data, for data about employees or on a large scale, two criteria are met. The answer is yes for the matter types where two criteria are met.
Can we use the vendor's own DPIA?
As input, not as your assessment. The vendor can describe the processing on its side, but only you can assess necessity, proportionality and the risks to your data subjects, because only you know what data you enter and why. Article 35 places the responsibility on the controller.
Do we have to consult IMY?
Only if the assessment shows a high risk despite the measures you plan [1]. With the right contract, retention periods and a list of matters kept out of the tool, a legal department's use should not end up there. If it does, that is a sign that the use should be narrowed, not that prior consultation should be requested.

Sources

  1. [1]När ska en konsekvensbedömning genomföras, IMY, in Swedish; IMY's list under Article 35(4) with nine criteria
  2. [2]Konsekvensbedömningar och förhandssamråd, IMY, in Swedish; IMY on impact assessments and prior consultation
  3. [3]GDPR-vägledningen, Sveriges advokatsamfund, 10 september 2026, in Swedish; Swedish Bar Association GDPR guidance

Back to AI for legal

More under AI for legal