Public bodies, private firms providing public services, and deployers that use high-risk AI to assess creditworthiness or price life and health insurance must complete a fundamental rights impact assessment before first use. Article 27 of the EU AI Act sets its contents: how the AI agent is used, who it affects, what could harm them, and who oversees it.
What does Article 27 require?
A fundamental rights impact assessment, usually shortened to FRIA, is a documented assessment of how a high-risk AI system could affect people's rights in the deployer's own use. It is done before the system is first put into use. In substance, it must describe:
- the deployer's processes in which the system will be used, in line with its intended purpose;
- the period of time over which, and how often, it will be used;
- the categories of people and groups likely to be affected in this context;
- the specific risks of harm likely to affect those people, taking account of the provider's information;
- the human oversight measures, as set out in the instructions for use;
- what the deployer will do if those risks materialise, including internal governance and complaint routes.
Once done, the deployer notifies the market surveillance authority of the results, using a standard questionnaire to be made available at EU level. If something changes that makes the assessment out of date, it must be updated. Where a data protection impact assessment already covers some of these points, the FRIA complements it rather than repeating it.
Our explainer on what a fundamental rights impact assessment is covers the concept; this piece is about the duty.
Who must carry one out?
Three groups of deployers of high-risk systems:
- Bodies governed by public law. Ministries, agencies, local authorities and similar.
- Private entities providing public services. Firms delivering services of a public nature, such as some education, health or housing providers.
- Deployers of two specific Annex III uses. Systems that assess the creditworthiness of individuals or establish their credit score, and systems used for risk assessment and pricing of life and health insurance for individuals.
For financial services the third group is the one that matters. A lender running a credit agent on individual applicants, or an insurer using an agent to price life cover, needs a FRIA. A bank running a service agent or a payment agent does not, unless that agent takes part in one of those listed uses. The question whether AI agents need a fundamental rights impact assessment usually turns on that capability-level detail.
What changes when the high-risk system is an agent?
A FRIA for a scoring model describes one decision point. An agent in a credit process may gather documents, read bank statements, compute affordability, draft a recommendation and write to the applicant. Each step can affect rights differently.
Three points are specific to agents:
Scope by capability, not by product. Describe which of the agent's capabilities touch the listed use. An agent that also answers general questions has capabilities outside the FRIA; say so and show they are separated.
Risks from actions, not only from outputs. An agent that writes to a customer can mislead them. One that requests documents can burden some applicants more than others, for example people without digital bank statements. One that summarises income can drop irregular earnings that matter. These are rights risks, and a model-centred FRIA will miss them.
Oversight described at the point of action. The FRIA must name the oversight measures. For an agent, state which actions a person must approve, who that person is, what they see when they approve, and how often they disagree with the agent.
How do you meet Article 27 for an agent?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Describe the process | Process map showing every step where the agent acts, and at what depth | Map attached to the FRIA, versioned |
| Period and frequency | Volume and schedule of use stated per capability | Usage figures from the action record |
| Affected people and groups | Groups listed, including those more exposed, such as thin-file or vulnerable applicants | Group list with reasoning |
| Specific risks of harm | Risks per capability, informed by the provider's instructions and your own tests | Risk table with likelihood and severity |
| Human oversight | Named approvers for decisions affecting the applicant; hold points before any adverse step | Approval records and override rates |
| Measures if risks materialise | Complaint route, review procedure, suspension trigger | Procedures and logged uses of each |
| Notify the authority | Submission using the authority's template | Submission receipt |
| Keep it current | Review on any change of capability, source or population | Revision history |
A practical template for these fields is in our fundamental rights impact assessment template for AI agents, and the step-by-step method is in how to run a FRIA for an AI agent.
What will a supervisor or auditor ask?
- Why does this agent fall within the FRIA duty, and which capabilities are in scope?
- Which groups did you consider, and how did you identify the ones more exposed?
- What risks did you find that the provider did not mention?
- Show us the oversight in practice: approvals, overrides, waiting times.
- What triggered the last update?
- When did you notify the authority?
What does Article 27 not cover?
It is not a conformity assessment and does not certify the system; the provider's own assessment is separate, as explained in EU AI Act conformity assessment and AI agents. It is not a DPIA, though the two should be run together. It does not apply to private firms using high-risk systems outside the listed credit and insurance uses unless they provide public services. And it does not prohibit anything: a FRIA that finds high risk obliges you to mitigate and govern, not to stop, though stopping may be the right answer. For the full set of duties, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can a FRIA be reused across similar agents?
Parts can, such as the description of affected groups. But the Act ties the assessment to the deployer's own process and use, so each deployment needs its own risks, oversight and frequency stated.
Does a FRIA have to be published?
The Act requires notification of the results to the market surveillance authority. It does not require publication to the public, though public bodies may face transparency duties under other law.
Who should own the FRIA inside a lender?
The business owner of the credit process, with compliance, data protection and model risk contributing. The owner should be someone who can change the process if the assessment says it should change.