Purpose and context
The intended purpose and expected benefit, the process the system will be used in, and the period and frequency of use (Article 27(1)(a) and (b)).
What an AI impact assessment documents, how to run one, and how it differs from the EU’s fundamental rights impact assessment and a GDPR DPIA.
An AI impact assessment is a documented evaluation of how an AI system and its foreseeable uses could affect individuals, groups and society, for better and for worse.
Where a risk assessment scores what could go wrong, an impact assessment looks outward, at the people on the receiving end of the system’s outputs.
elements a fundamental rights impact assessment must contain: Article 27(1), points (a) to (f)
Source: EU AI Act, Article 27
items in the AI impact assessment US federal agencies complete for high-impact AI
Source: OMB Memorandum M‑25‑21
when EU AI Act rules for Annex III high-risk systems, Article 27 included, start to apply
Formats vary. These four elements draw on EU AI Act Article 27 and OMB M‑25‑21, the two texts that spell the contents out.
The intended purpose and expected benefit, the process the system will be used in, and the period and frequency of use (Article 27(1)(a) and (b)).
The categories of people and groups likely to be affected in that specific context, not users in general (Article 27(1)(c)).
The specific risks of harm to those groups (Article 27(1)(d)). M‑25‑21 names privacy, civil rights and civil liberties. NIST MAP 5.1 asks for the likelihood and magnitude of each impact, beneficial or harmful.
How human oversight is implemented and what happens if a risk materializes (Article 27(1)(e) and (f)), plus a schedule for reassessment, which M‑25‑21 adds.
For how these differ from a risk assessment, see AI risk assessment.
| AI impact assessment | Fundamental rights impact assessment | Data protection impact assessment | |
|---|---|---|---|
| Source | ISO/IEC 42005:2025; ISO/IEC 42001 clause 6.1.4 | EU AI Act, Article 27 | GDPR, Article 35 |
| Status | Voluntary guidance | Legal duty | Legal duty |
| Who does it | Any organization developing, providing or using AI | Public bodies, private providers of public services, and deployers of credit scoring or life and health insurance pricing AI | The data controller, where processing is likely to result in a high risk |
| Looks at | Impacts on individuals and society | Fundamental rights of the people affected by one deployment | Rights and freedoms affected by personal data processing |
| Overlap | Can be integrated into AI risk management | May cross-reference a DPIA that already meets an obligation (Article 27(4)) | Can be cross-referenced by a FRIA instead of repeated |
Screen every proposed use at intake. Uses that make or shape decisions about people go on to a full assessment.
Record purpose, users, data, deployment setting and the decisions the output feeds.
Name them, and gather input from operators, users and affected communities, as the NIST Playbook suggests under GOVERN 4.2.
Record each impact with its likelihood and magnitude, the mitigation and the residual effect.
Use the result in the go or no-go decision, and update it when a documented element changes, as Article 27(2) requires for a FRIA.
Straight answers, kept current. Last updated .
Ask us anything →It depends on who you are. ISO/IEC 42005 is voluntary guidance. In the EU, Article 27 requires a fundamental rights impact assessment from certain deployers of high-risk AI before first use, under rules the Commission’s timeline applies from December 2, 2027. In the United States, OMB M‑25‑21 requires one of federal agencies for high-impact AI. It does not bind private companies.
Deployers that are bodies governed by public law, private entities providing public services, and deployers of the Annex III systems in point 5(b) and (c): AI that evaluates creditworthiness or sets a credit score, and AI for risk assessment and pricing in life and health insurance. They notify the results to the market surveillance authority (Article 27(3)).
A risk assessment scores what could go wrong with a system: likelihood, severity and treatment. An impact assessment documents who is affected and how, including benefits. ISO/IEC 42001 lists them as separate clauses, 6.1.2 and 6.1.4, and the results of one feed the other. Running both in one review avoids asking a team the same questions twice.
Primary sources for the figures and frameworks cited on this page. Last verified .
See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.