
The AI Act's FRIA Template Is Out: Who Fills It In and What It Asks
On 16 November the European Commission published the draft template for the fundamental rights impact assessment (FRIA) required by Article 27 of the AI Act, together with a draft explanatory notice. The Commission's announcement sets two dates: feedback is open until 15 January 2027, and deployers who will have to carry out a FRIA can volunteer by 4 December 2026 to test the draft on one of their own systems. The obligation itself applies from 2 December 2027.
Until now, Article 27 was a paragraph that listed six things an assessment must contain. The draft document turns that into a questionnaire of about a dozen pages, with tick boxes, risk scales and a signature line. Once the template is final, covered deployers will have to use it and attach it when they notify the results to their market surveillance authority. The Commission also plans to offer it as an automated tool on the AI Act Single Information Platform.
The draft carries a clear warning on its first page: it has not been adopted or endorsed by the Commission and the questions can still change. Even so, it is the first official picture of what the AI Office will expect, and for a bank, a lender or an insurer it is concrete enough to start a dry run.
Who actually has to do one
Most companies using high-risk AI will not need a FRIA. The draft notice repeats the narrow list from Article 27(1) of the AI Act:
- bodies governed by public law, and private entities providing public services, when they deploy a high-risk system in one of the Annex III areas (critical infrastructure is excluded);
- deployers of AI systems that evaluate creditworthiness or set a credit score, except systems meant for detecting financial fraud;
- deployers of AI systems used for risk assessment and pricing in life and health insurance.
"Private entities providing public services" is the line that will cause debate. The notice points to recital 96 of the Act, which gives education, healthcare, social services, housing and the administration of justice as examples, and says the list is not closed. A private hospital or a private school using a high-risk system should assume it is in scope until someone with legal authority says otherwise.
For Bulgaria, the most obvious group is lending. Any bank or non-bank lender that uses an AI model to decide who gets a consumer loan, or on what terms, is a deployer of a high-risk system under Annex III, and Article 27 applies to it directly, whether it built the model or bought it from a vendor.
What the form asks
The template follows Article 27(1) closely. It starts with general information about the organisation, the system and the responsible person. Then it asks how the system fits into your processes, for how long and how often it will be used, and which groups of people it affects, with explicit boxes for children, older people, persons with disabilities and people in a vulnerable social or economic situation.
The core is section 4. For each concrete risk of harm, the deployer describes the risk, ticks the Charter rights it touches, names the affected groups and rates likelihood and severity as low, medium or high. Preventive measures are not required only where both likelihood and severity are low. Then comes a second rating: the level of risk after those measures.
Section 5 covers human oversight, section 6 what happens if a risk materialises, and section 7 asks for a short overall conclusion. Public authorities that must register the system in the EU database also fill in a summary that will be published there.
The DPIA does part of the work
A lender scoring loan applicants has almost certainly already done a data protection impact assessment under GDPR. The draft recognises that. The FRIA "complements" the DPIA, and deployers may cross-refer to the relevant sections or copy them in, especially for the right to the protection of personal data. The template even asks you to restate the DPIA's likelihood and severity ratings on its own low-medium-high scale.
Deployers may also rely on an earlier FRIA or on an impact assessment done by the provider, where the system and the risks are similar. The catch, which the notice states plainly, is that the provider is not obliged to share its assessment. Asking your vendor for it now, while contracts for 2027 are being negotiated, costs nothing.
Where the hard questions are
The rights analysis is the part people expect to struggle with. Reading the draft, I think the questions that will take longer are operational.
Question 5.1 asks who carries out human oversight, by role and position, what authority they have to override, disregard or reverse the system's output, and what tools support them. If you deviated from the provider's instructions for use, you must explain why. Question 5.2 asks how those people were given the competence, training and authority, and it mentions AI literacy under Article 4 by name. A sentence such as "a credit officer reviews the results" will not answer either question.
Question 6.2 asks how complaints from affected people are received and processed, and how complainants learn what was done. For a lender that already has a complaints procedure for its loans, this is mostly about connecting that process to the AI system. For a public body that never thought of its scoring tool as something people could complain about, it may be new work.
Then there is the update rule. The notice says the FRIA must be updated when relevant factors change, and its examples include the processes concerned, the persons affected and "the version of the system or its underlying model". For anyone running AI in production, that line matters. Model versions change several times a year, sometimes at the vendor's initiative. That is an interpretation of an example in a draft, not settled law, but a FRIA that nobody revisits after go-live is unlikely to satisfy it.
A hypothetical: a consumer lender in Sofia
Take a hypothetical online lender in Sofia that uses a vendor's scoring model to approve or decline small consumer loans. It has a DPIA from two years ago and a complaints procedure for its credit products.
Going through the draft template, it could fill sections 1 to 3 in an afternoon. Section 4 would lean on the DPIA, but it would also have to think about harms the DPIA may not have covered, such as a model that systematically scores pensioners or people with irregular income lower. Section 5 would force it to write down whether the credit officer can actually overturn a decline, how often that happens and what training the officer received. And when the vendor moves to a new model version next spring, the lender would need to know about it in time to update the assessment.
None of this requires waiting for 2027. A dry run on one system shows where the gaps are while there is still time to close them.
Who receives it in Bulgaria
Article 27(3) says the results go to the market surveillance authority of the Member State. The draft leaves the submission link "to be added later". In Bulgaria there is a further gap: the parliamentary committee on innovation rejected the AI bill at first reading on 4 November, so the law that would name the national authorities is not in place. That does not change the obligation or its date. It does mean Bulgarian deployers cannot yet ask a local authority for guidance, which makes the EU-level testing call more useful.
Private employers and CV screening
One group is often surprised to find itself outside Article 27: private employers using AI to screen job applications. Recruitment systems are high-risk under Annex III, but a private company that is not providing a public service has no FRIA obligation. It does still have GDPR. On 11 November the European Data Protection Board chose automated decision-making in recruitment, including AI-driven candidate screening, as the topic of its coordinated enforcement action for 2027, focusing on the right not to be subject to solely automated decisions and the related right of access.
Before 4 December and 15 January
If your organisation is in scope, two things are worth doing this year. Read the draft against one real system and note which questions you cannot answer today. And decide whether to put that system forward for testing by 4 December through the expression-of-interest form; the AI Office says it will support participants with an introductory session, answers along the way and a closing session. Comments on the wording itself can be sent until 15 January 2027.
If you want a structured view of where your systems sit under the AI Act, our AI Act readiness test is a starting point, and our AI governance work covers assessments like this one.


