GUIDE · AI ACT

Fundamental rights impact assessment (FRIA): when and how to carry it out

After this guide you will be able to tell whether your case requires a fundamental rights impact assessment, prepare its structure in concrete steps, identify the people involved and the risks to describe, and link the assessment to data protection and to the supervision of the system. The scope should be confirmed with a consultant.

Updated october 2026
7 min read
Altovar Technical Editorial Team

1. What it is and why it matters

The fundamental rights impact assessment, often referred to by the acronym FRIA from the English Fundamental Rights Impact Assessment, is a written analysis that some deployers of AI systems must prepare before putting them into use. It helps establish how the system may affect rights such as non-discrimination, confidentiality, the protection of workers and access to services. The document should also be readable by people who are not technical specialists.

It is not a technical evaluation of the model. It asks you to reason about the context in which the system operates: who is affected, at what moments, with which alternatives and with what possibilities of redress. For this reason it should be prepared together with people who know the work process, not only with those who know the software, because the consequences show up in the process and in people's lives.

2. When it is required: it depends on the entity and the use

The obligation does not apply to every deployer. It depends on the type of entity, for example public bodies or private entities providing public services, and on the use, for example certain creditworthiness assessments or certain risk assessments for life and health insurance. Check both parameters before starting, because a single one of them can change the picture. If the case is uncertain, document the scoping assessment, so that the reason for the choice remains traceable.

Even where the obligation does not apply, a similar assessment can be good practice for systems that affect people. The work involved takes time, but it is not wasted: it forces you to describe the system's purpose precisely, which is exactly what the consultant needs to confirm the risk classification. Even in this case, the assessment becomes a useful exercise for the organisation.

3. The steps of the assessment

Organise the work in steps, each with an owner and a supporting document. The format can be the one your organisation already uses for risk analyses, provided it covers the points below, in the order in which it is best to address them. The first step is always the longest, because it requires you to put in writing things everyone takes for granted. From the second version onwards the work gets shorter.

  1. Describe the process in which the system is used, its purpose, frequency of use and the people involved.
  2. Identify the categories of people who may be affected, including those who are not directly served by the system.
  3. Identify the rights that may be affected and how, describing concrete cases rather than generic formulas.
  4. Assess the likelihood and severity of each risk, with a written rationale for each assessment.
  5. Describe the planned human oversight and indicate who can suspend the system.
  6. List the measures to reduce or manage the risks, with internal owners and deadlines.
  7. Define how the assessment is updated and what happens if the system or its use changes.

4. Data and the relationship with the GDPR assessment

If you already have a data protection impact assessment under the GDPR for the same processing, do not start from scratch. You can use it as a basis and complete it with the points specific to the AI system. A separate assessment that is inconsistent with the first creates more problems than it solves, because two documents saying different things about the same processing are hard to defend.

Pay attention to training data and outputs: they may contain personal data even when you did not input it yourself. Document where the data comes from, who can consult the outputs and how long they are kept, coordinating the choice with the data protection officer. Also describe who has access to the outputs and for how long they remain available.

5. Common mistakes

The first mistake is writing a generic assessment that could apply to any system. A useful document cites the process, the department and the type of decision taken. The second is describing risks as if they were purely technical, forgetting those affecting people, such as the difficulty of understanding and contesting a decision assisted by the system. This applies even if the assessment was carried out by another department.

The third is writing the assessment once and never updating it. A change of supplier, a new function or the extension of the system to another department alters the picture and must be reflected in the document. The fourth is failing to say who can stop the system: without a name and a role, the measures remain on paper. Specify also which event triggers a review.

6. How to check that the document holds up

A simple test: read the assessment as if you were a person subject to a decision assisted by the system. Can you understand what happened, why, and how to challenge it? If the answer is no, the document is still too technical or too abstract and needs to be rewritten in the part concerning people. Give the reader concrete examples of scenarios and ask whether the explanation is understandable.

Have the text reviewed by someone who did not write it, preferably a person from the department that uses the tool and a consultant. Keep the versions and the comments received. If an authority asked for the document, you should be able to show how it was built, who approved it and on what date. The final document must record the observations received and how they were handled.

7. How to start

This week, check with your consultant whether the system you are planning falls within the entities and uses for which the obligation applies. If the answer is uncertain, prepare it as though it were mandatory: the work is useful anyway, because it makes the risks visible before launch. If the consultant has already indicated some steps, add them to the draft straight away and do not postpone them. Even a partial answer, provided it is in writing, is a useful starting point.

Then prepare a first draft with the process, the purpose and the people involved, and a list of risks to discuss with the department. You do not need a final document straight away. The draft tells you which information is missing from the supplier and which internal decisions you still have to take. Once the draft is ready, schedule the first meeting with the department and the consultant.

Frequently Asked Questions

Does the assessment have to be carried out for every AI system?

No, only for the cases the regulation provides for, that is, for certain deployers and certain uses. For other systems a lighter assessment may still be useful to understand the risks. In any case, the decision not to carry one out must be documented with its reasoning, so that it can be verified.

Who should sign the assessment?

Responsibility usually lies with senior management or a designated manager, with input from the department that uses the system, from whoever manages data protection and from the consultant. The signature should not be a formality: the signatory must have read the document and understand the measures planned.

Do I have to report the result to the authority?

In some cases the regulation requires the results to be communicated to the competent authorities. The method and timing should be checked against the up-to-date text and with your consultant. Do not send anything without verification, but do not treat the matter as irrelevant.

Can I use a generic document template?

You can use it as a structure, but not as content. A generic template does not describe your process or your people, and that is exactly what authorities and reviewers ask to see. Adapt each section to the specific system and remove any parts you do not apply, explaining why.

How long does it take?

It depends on the complexity of the system, the availability of information from the supplier and the number of departments involved. A simple system with one department is much quicker than one that affects different processes. Plan adequate time for the review, and do not treat it as a task to be closed in an afternoon.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →