EU AI Act Article 27: Fundamental Rights Assessment, Implemented in Practice
EU AI Act Article 27 requires public authorities, providers of public services, banks and insurers to carry out a fundamental rights impact assessment before they first use a high-risk system, and at SIFAMO we keep it as a versioned record inside the registry entry, one that references the data protection impact assessment instead of copying it and without which the gateway opens no route.
Article 27 is the only article in the high-risk chapter that addresses the deployer alone and gives it an assessment of its own to carry out before use. It addresses public authorities, private providers of public services and deployers of systems for creditworthiness assessment or for risk assessment in life and health insurance. In other words, exactly the organisations we work with. I am an engineer, not a lawyer. What interests me is how the assessment becomes an artefact that counts at runtime and that an auditor can still read later. The basics on deadlines and roles: my EU AI Act guide.
For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 27 belongs to the Annex III obligations. The Omnibus moved deadlines; it removed no obligation.
What Article 27 requires
Article 27 does not ask for a new risk analysis of the system, the provider does that under Article 9. It asks for an assessment of the specific use and its effect on the rights of the people affected:
Art. 27(1), who: Before first using a high-risk system under Art. 6(2) (Annex III), bodies governed by public law, private entities providing public services, and all deployers of systems under Annex III point 5(b) (creditworthiness) and 5(c) (life and health insurance) must assess the impact on fundamental rights. Critical infrastructure under Annex III point 2 is excluded.
Art. 27(1)(a) to (f), what: (a) the deployer’s processes in which the system will be used in line with its intended purpose; (b) the period and frequency of use; (c) the categories of persons and groups likely to be affected; (d) the specific risks of harm to those groups, taking into account the provider’s information under Art. 13; (e) how human oversight will be implemented according to the instructions for use; (f) the measures to be taken if the risks materialise, including internal governance and complaint mechanisms.
Art. 27(2): The obligation applies to the first use. In similar cases the deployer may rely on earlier assessments, including those carried out by the provider. If any element changes or is no longer up to date, it must update it.
Art. 27(3): Once the assessment is done, the deployer notifies the market surveillance authority of the results and submits the completed template under paragraph 5. In the case of Art. 46(1) (derogation from the conformity assessment procedure for exceptional reasons), the notification duty may be waived.
Art. 27(4): Where a data protection impact assessment under Art. 35 GDPR or Art. 27 of Directive (EU) 2016/680 already meets obligations of this article, the fundamental rights impact assessment complements it. Under the Digital Omnibus wording, the deployer may cross-reference the DPIA or include its relevant parts.
Art. 27(5): The Commission’s AI Office develops a questionnaire template, including through an automated tool, with cross-references to the DPIA. The template is being prepared; until then we work with a structure of our own along the six points.
The technical example
Two systems, one governance layer. First: a public authority introduces a system for a social benefit that prioritises incoming applications as urgent, standard or review case. It does not decide on the benefit; it reads the pseudonymised application and its metadata and proposes a queue. The system comes from a provider, the authority is the deployer. Annex III point 5(a) covers systems used by public authorities to evaluate eligibility for public assistance benefits. A prioritisation does not formally evaluate eligibility, but whoever ends up at the back of the queue waits for money they need to live on. We take the stricter reading rather than the exception in Art. 6(3). As a body governed by public law, the authority falls under Art. 27(1) in any case.
Second: a bank runs a pre-check in retail lending that routes applications to a fast track, a standard review or a specialist review, Annex III point 5(b). The bank built the model itself, so it is provider and deployer at the same time. Article 27 hits it in its deployer role.
The data flow is the same in both cases: the business application calls the gateway with a use-case ID, the gateway checks the registry entry and the data zone, pseudonymises with IDs, calls the model, writes the evidence log and places cases in the review queue according to their oversight tier. Article 27 does not bite at the model. It bites at the process around it. Which groups are standing in this queue? What happens to those the model sorts wrongly? Whom does an affected person turn to? Only the authority or the bank can answer that, not the maker of the model.
How we implement it
At SIFAMO we do not keep the fundamental rights impact assessment as a document next to the system, but as its own versioned block inside the registry entry, which the governance layer reads at runtime. Each field corresponds to one point of Art. 27(1), plus the reference to the DPIA, the notification and the triggers for an update. Simplified, for the authority:
{
"model_id": "DE-P-2026-0062",
"purpose": "Prioritisation of applications for a social benefit, no decision on entitlement",
"risk_class": "high (Annex III point 5(a)), tier 3 (internal)",
"fria": {
"version": "benefit-priority-fria-v2", "status": "approved",
"a_process": "Benefit intake, step 2 (pre-sorting); the decision stays with the caseworker",
"b_period_frequency": "open-ended, continuous, all incoming applications",
"c_affected_groups": ["single parents", "people without German language skills", "people in acute hardship"],
"d_harms": [
{"id": "FR-01", "right": "equal treatment", "harm": "downgrading via proxy features (household size, place of residence)",
"source": "provider's instructions for use", "mitigation": "gateway rule G-14, bias run per version"},
{"id": "FR-02", "right": "access to benefits", "harm": "delayed payment when wrongly sorted",
"mitigation": "maximum waiting time per queue, monitoring M-07, notice in the decision letter, complaint channel B-02"}
],
"e_oversight": "review case: full review; standard: 5 % sample with blind re-check",
"f_measures": {"kill_switch": "head of office", "complaints": "B-02, reply within 10 working days", "escalation": "supervisory unit, data protection officer"},
"dpia_ref": {"id": "DPIA-2026-011", "reused_sections": ["description of processing", "data categories", "technical and organisational measures"]},
"notification": {"authority": "competent market surveillance authority", "date": "2026-10-06", "format": "FRIA export v1, until the Commission template is available"},
"signed": ["head of office (business owner)", "data protection officer", "IT operations"],
"update_on": ["process", "applicant group", "model or prompt version", "oversight tier", "complaint pattern", "expiry 2026-12-15"]
}
}
Six points as fields, not as chapters
Points (a) and (b) come from the registry: purpose, scope, process step, plus period and frequency from the business owner, because “three-month pilot” and “open-ended for all applications” are two different assessments. For point (c) we name groups concretely. “Applicants” is not a group, it is the population, and the same groups end up as dedicated cases in the evaluation set. For point (d), every risk of harm gets an ID, the right affected, the source (usually the instructions for use under Art. 13) and a measure that exists as a gateway rule, a monitoring rule or an oversight tier. Point (e) is not “a human checks” but the oversight tier per route with sampling and blind re-checks. Point (f) means for us: a kill switch with a name, a complaint channel with a deadline, an escalation with a role. Without a complaint channel, (f) is not met, even if the gateway works perfectly.
Reference the DPIA, do not copy it
Art. 27(4) is the invitation not to write the data protection impact assessment a second time. We accept it. Description of processing, data categories and technical and organisational measures live in the DPIA and are referenced by ID and section. The FRIA adds what the DPIA does not ask: equal treatment, access to benefits, effective remedy. The data protection officer co-signs. Two documents about the same processing drift apart after the first change, a reference does not.
No assessment, no notification, no route
For systems within the scope of Article 27, the registry entry carries a flag. Without an approved FRIA version and without a notification date, the gateway opens no route for the productive use case. That is our decision, not the wording: paragraph 1 requires the assessment before use, paragraph 3 the notification afterwards. We turn that into a fixed order: assessment, notification, first use. Until the Commission template is available, we export the FRIA block along the six points with the cross-reference to the DPIA and file the acknowledgement in the entry. For the bank, Art. 74(6) makes the financial supervisor the market surveillance authority.
Reuse with a delta
When the authority rolls the system out for a second benefit, it does not start a new FRIA under Art. 27(2) but a version with a delta: same provider, same oversight, different groups, different process. A provider’s assessment is the basis for point (d), never a substitute for (a), (b), (c) and (f). Only the deployer knows those.
Update when something changes
The FRIA version is tied to the model, prompt and policy versions in the registry entry. If one of them changes, the FRIA comes up for review. Then there are triggers from operations: sampling rate lowered (e), new applicant group (c), complaint patterns or reviewer actions in the evidence log that match no known risk of harm (d), pilot becomes regular operation (b). The expiry date of the registration forces a review even without a change. We run the FRIA today, ahead of December 2027, because the questions (a) to (f) have to be answered anyway before a model sorts people’s applications.
What the auditor sees: the versioned FRIA block with groups, risks of harm, measure IDs and signatures, the DPIA reference, the notification date with acknowledgement and the version history with triggers. For the authority, the same entry feeds the registration in the EU database under Art. 49(3). More under AI Governance.
Typical mistakes
The FRIA is a renamed DPIA. It says which data is processed, but not who is disadvantaged and where they can turn.
The provider wrote it. Its assessment knows the system, but not the deployer’s process, groups and oversight. Art. 27(2) allows it as a basis, not as a replacement.
Affected group: “customers”. Without concrete groups there are no specific risks of harm and no cases in the evaluation set.
Point (f) without a complaint channel, notification after go-live or not at all. And never updated, although the sampling rate and the model version changed long ago.
EU AI Act Article 27 checklist
Every high-risk system within the scope of Art. 27(1) has a versioned FRIA block in its registry entry, linked to the model, prompt and policy versions.
Process, period and frequency are described by the business owner and match the provider’s intended purpose.
Affected groups are named concretely and represented as dedicated cases in the evaluation set.
Every risk of harm has an ID, the right affected, its source in the Art. 13 information and an enforced measure; the oversight tier per route is enforced by the gateway.
Kill switch, complaint channel with deadline and escalation are recorded with names and roles; the DPIA is referenced by ID and section.
The notification to the market surveillance authority was made before first use and is filed with acknowledgement; the triggers for an update are defined.
Open questions
Which authority receives the notification? For banks, Art. 74(6) points to the financial supervisor. For public authorities and private providers of public services the national allocation of competences decides, and the Commission template is still missing.
What is a “similar case”? Art. 27(2) does not say whether a second social benefit with a different applicant population is still similar. We justify the comparison in the entry.
How do you measure a risk of harm to a fundamental right? There are metrics for discrimination, none for effective remedy. When a private entity “provides public services” is not conclusively settled either.
Frequently asked questions
Does this apply to us as deployers?
Article 27 addresses deployers only, but only the groups it names. An industrial company with a high-risk system for screening job applicants has the deployer obligations under Art. 26, but no fundamental rights impact assessment. A bank, a life insurer and a public authority always have one.
Is our data protection impact assessment enough?
No. Art. 27(4) allows you to reference it, but the FRIA complements it. What is usually missing: groups, risks of harm beyond data protection, concrete oversight, a complaint channel.
Do we have to publish the FRIA?
No, Article 27 requires the notification to the market surveillance authority with the template under paragraph 5. Public authorities additionally register the use in the EU database under Art. 49(3). What affected persons learn is governed by Art. 26(11) and Art. 86, in our case the notice in the decision letter.
This post is part of the series on the EU AI Act, article by article; see all articles at a glance.
EU AI Act Article 27: Fundamental Rights Assessment, Implemented in Practice
EU AI Act Article 27 requires public authorities, providers of public services, banks and insurers to carry out a fundamental rights impact assessment before they first use a high-risk system, and at SIFAMO we keep it as a versioned record inside the registry entry, one that references the data protection impact assessment instead of copying it and without which the gateway opens no route.
Article 27 is the only article in the high-risk chapter that addresses the deployer alone and gives it an assessment of its own to carry out before use. It addresses public authorities, private providers of public services and deployers of systems for creditworthiness assessment or for risk assessment in life and health insurance. In other words, exactly the organisations we work with. I am an engineer, not a lawyer. What interests me is how the assessment becomes an artefact that counts at runtime and that an auditor can still read later. The basics on deadlines and roles: my EU AI Act guide.
For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 27 belongs to the Annex III obligations. The Omnibus moved deadlines; it removed no obligation.
What Article 27 requires
Article 27 does not ask for a new risk analysis of the system, the provider does that under Article 9. It asks for an assessment of the specific use and its effect on the rights of the people affected:
The technical example
Two systems, one governance layer. First: a public authority introduces a system for a social benefit that prioritises incoming applications as urgent, standard or review case. It does not decide on the benefit; it reads the pseudonymised application and its metadata and proposes a queue. The system comes from a provider, the authority is the deployer. Annex III point 5(a) covers systems used by public authorities to evaluate eligibility for public assistance benefits. A prioritisation does not formally evaluate eligibility, but whoever ends up at the back of the queue waits for money they need to live on. We take the stricter reading rather than the exception in Art. 6(3). As a body governed by public law, the authority falls under Art. 27(1) in any case.
Second: a bank runs a pre-check in retail lending that routes applications to a fast track, a standard review or a specialist review, Annex III point 5(b). The bank built the model itself, so it is provider and deployer at the same time. Article 27 hits it in its deployer role.
The data flow is the same in both cases: the business application calls the gateway with a use-case ID, the gateway checks the registry entry and the data zone, pseudonymises with IDs, calls the model, writes the evidence log and places cases in the review queue according to their oversight tier. Article 27 does not bite at the model. It bites at the process around it. Which groups are standing in this queue? What happens to those the model sorts wrongly? Whom does an affected person turn to? Only the authority or the bank can answer that, not the maker of the model.
How we implement it
At SIFAMO we do not keep the fundamental rights impact assessment as a document next to the system, but as its own versioned block inside the registry entry, which the governance layer reads at runtime. Each field corresponds to one point of Art. 27(1), plus the reference to the DPIA, the notification and the triggers for an update. Simplified, for the authority:
Six points as fields, not as chapters
Points (a) and (b) come from the registry: purpose, scope, process step, plus period and frequency from the business owner, because “three-month pilot” and “open-ended for all applications” are two different assessments. For point (c) we name groups concretely. “Applicants” is not a group, it is the population, and the same groups end up as dedicated cases in the evaluation set. For point (d), every risk of harm gets an ID, the right affected, the source (usually the instructions for use under Art. 13) and a measure that exists as a gateway rule, a monitoring rule or an oversight tier. Point (e) is not “a human checks” but the oversight tier per route with sampling and blind re-checks. Point (f) means for us: a kill switch with a name, a complaint channel with a deadline, an escalation with a role. Without a complaint channel, (f) is not met, even if the gateway works perfectly.
Reference the DPIA, do not copy it
Art. 27(4) is the invitation not to write the data protection impact assessment a second time. We accept it. Description of processing, data categories and technical and organisational measures live in the DPIA and are referenced by ID and section. The FRIA adds what the DPIA does not ask: equal treatment, access to benefits, effective remedy. The data protection officer co-signs. Two documents about the same processing drift apart after the first change, a reference does not.
No assessment, no notification, no route
For systems within the scope of Article 27, the registry entry carries a flag. Without an approved FRIA version and without a notification date, the gateway opens no route for the productive use case. That is our decision, not the wording: paragraph 1 requires the assessment before use, paragraph 3 the notification afterwards. We turn that into a fixed order: assessment, notification, first use. Until the Commission template is available, we export the FRIA block along the six points with the cross-reference to the DPIA and file the acknowledgement in the entry. For the bank, Art. 74(6) makes the financial supervisor the market surveillance authority.
Reuse with a delta
When the authority rolls the system out for a second benefit, it does not start a new FRIA under Art. 27(2) but a version with a delta: same provider, same oversight, different groups, different process. A provider’s assessment is the basis for point (d), never a substitute for (a), (b), (c) and (f). Only the deployer knows those.
Update when something changes
The FRIA version is tied to the model, prompt and policy versions in the registry entry. If one of them changes, the FRIA comes up for review. Then there are triggers from operations: sampling rate lowered (e), new applicant group (c), complaint patterns or reviewer actions in the evidence log that match no known risk of harm (d), pilot becomes regular operation (b). The expiry date of the registration forces a review even without a change. We run the FRIA today, ahead of December 2027, because the questions (a) to (f) have to be answered anyway before a model sorts people’s applications.
What the auditor sees: the versioned FRIA block with groups, risks of harm, measure IDs and signatures, the DPIA reference, the notification date with acknowledgement and the version history with triggers. For the authority, the same entry feeds the registration in the EU database under Art. 49(3). More under AI Governance.
Typical mistakes
EU AI Act Article 27 checklist
Open questions
Frequently asked questions
Does this apply to us as deployers?
Article 27 addresses deployers only, but only the groups it names. An industrial company with a high-risk system for screening job applicants has the deployer obligations under Art. 26, but no fundamental rights impact assessment. A bank, a life insurer and a public authority always have one.
Is our data protection impact assessment enough?
No. Art. 27(4) allows you to reference it, but the FRIA complements it. What is usually missing: groups, risks of harm beyond data protection, concrete oversight, a complaint channel.
Do we have to publish the FRIA?
No, Article 27 requires the notification to the market surveillance authority with the template under paragraph 5. Public authorities additionally register the use in the EU database under Art. 49(3). What affected persons learn is governed by Art. 26(11) and Art. 86, in our case the notice in the decision letter.
This post is part of the series on the EU AI Act, article by article; see all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments