Which Assessments Does an AI System Need? DPIA, FRIA, EQIA and ATRS: Practitioner Q&A
DPIA, FRIA, EQIA, ATRS and AI impact assessment compared: which apply to an AI system, who each is aimed at and how to reuse evidence across them.
An AI system can fall under several assessments at once: a data protection impact assessment (DPIA), a fundamental rights impact assessment (FRIA), an equality impact assessment (EQIA), an Algorithmic Transparency Recording Standard (ATRS) record and an AI impact assessment. They overlap, but each answers a different question: personal data, fundamental rights, equality, public transparency or the AI system as a whole. The practical gain is gathering evidence once and reusing it across every assessment that applies. This page is general information and is not legal advice; check each source against your own circumstances.
What is the difference between a DPIA and a FRIA?
A DPIA asks what risk the processing of personal data poses to individuals. A FRIA asks how deploying a high-risk AI system could affect people’s fundamental rights. They rest on different laws. The DPIA sits in UK data protection law (the UK GDPR and the Data Protection Act 2018). Article 35 aims it at controllers whose processing is likely to result in a high risk. The FRIA comes from Article 27 of the EU AI Act and is aimed at certain deployers of high-risk systems. Article 27(4), as amended by Regulation (EU) 2026/1744, lets a deployer cross-reference or include relevant parts of a DPIA in a FRIA. The FRIA remains a separate obligation.
Do we need a DPIA for every AI system that uses personal data?
Not automatically, though ICO guidance points to one in most AI cases. UK GDPR Article 35 requires a DPIA where processing, in particular using new technologies, is likely to result in a high risk to individuals’ rights and freedoms. The ICO’s guidance on AI and data protection says that in the vast majority of cases AI use will involve processing likely to be high risk, which triggers the requirement. Where screening concludes that no DPIA is needed, the ICO says to document the decision and reasons, including the DPO’s advice. The guidance carries a notice that it is under review because of the Data (Use and Access) Act 2025 and may change. Automated decision-making has separate rules in UK GDPR Articles 22A to 22D. See also our FAQ on using AI on customer data.
Who has to carry out a fundamental rights impact assessment under the EU AI Act?
Article 27 of the EU AI Act places the duty on deployers of high-risk AI systems and leaves providers outside it. It names three groups: bodies governed by public law, private entities providing public services and deployers of the Annex III systems used for creditworthiness or credit scoring and for risk assessment and pricing in life and health insurance. The assessment comes before deploying an Annex III system, with critical infrastructure excluded. Following Regulation (EU) 2026/1744, these obligations apply from 2 December 2027. Under Article 2, the Act reaches organisations outside the EU where they place systems on the EU market or the system’s output is used in the EU. A UK NHS trust or local authority with no EU link is unlikely to carry the duty and may use the FRIA as a benchmark only.
Do private companies need an equality impact assessment for AI?
No statute requires a private company to produce an EQIA. The public sector equality duty in section 149 of the Equality Act 2010 is aimed at public authorities, which include NHS trusts and local authorities. It reaches other bodies only in respect of public functions they exercise, for example under some outsourced contracts. Even for public bodies, the Act requires due regard to its three equality aims and does not prescribe a document called an equality impact assessment; an EQIA is the common way to evidence due regard if a decision is challenged. The specific duties in Wales and Scotland require listed authorities to assess the impact of proposed policies and practices. The Act’s general anti-discrimination provisions apply to every organisation, so a private company still has reason to check an AI system for bias.
Which organisations have to publish an ATRS record?
The mandatory ATRS requirement currently applies to central government. The ATRS Mandatory Scope and Exemptions Policy covers all ministerial and non-ministerial government departments, plus arm’s-length bodies (executive agencies and non-departmental public bodies) that provide public or frontline services or routinely interact with the public. Tools in scope are those that significantly influence public decision-making or interact directly with the public. Under that policy, NHS trusts and local authorities fall outside mandatory scope. The Data Standards Authority recommends the ATRS across the wider public sector without requiring it. Some bodies publish voluntarily: by May 2025, 10 records had come from devolved administrations and local government. An ATRS record is published and does not assess risk, so it sits alongside the other assessments.
Do we need to update our DPIA when we add AI to an existing system?
Adding AI is a prompt to review the DPIA. ICO guidance does not say that every AI addition automatically requires a new one, but it describes a DPIA as a living process. Its guidance on how to do a DPIA says to keep the DPIA under review and that it may need repeating after a substantial change to the nature, scope, context or purposes of processing. After significant changes to how or why personal data is processed, or to the amount collected, the controller needs to show that the DPIA covers the new risks. Adding AI fits these triggers because it changes the technology and usually the scope. Roles need review too. Where an AI vendor processes data on the controller’s behalf, UK GDPR Article 28 governs the Data Processing Agreement. ICO guidance also says to assess whether joint controllership under UK GDPR Article 26 exists.
Can a DPIA, FRIA and EQIA be combined into one assessment?
A DPIA, FRIA and EQIA can share one evidence base and one process, provided each still meets its own test. The law supports reuse: Article 27(4) of the EU AI Act, as amended, lets a deployer cross-reference or include relevant parts of a DPIA in a FRIA. That provision permits cross-referencing only, and the deployer must still carry out the FRIA. A single merged document risks satisfying none of the tests, because each assessment has a different threshold, owner and output. A DPIA can lead to prior consultation with the ICO, the deployer notifies a market surveillance authority of the FRIA results and an EQIA evidences due regard. A practical approach is to collect the facts once and then write a separate conclusion for each obligation that applies, so each output stands alone.
How does an ISO 42001 AI impact assessment relate to a DPIA?
An ISO/IEC 42001 AI impact assessment and a DPIA can share evidence, but they differ in legal basis and scope. A DPIA is a statutory duty limited to the processing of personal data, while the AI system impact assessment in ISO/IEC 42001 is voluntary unless an organisation operates a 42001 management system or a contract requires it. Clause 6.1.4 requires a process for assessing an AI system’s potential consequences for individuals, groups and societies and clause 8.4 requires the assessment to be performed and its results kept. ISO/IEC 42005:2025 gives guidance on carrying it out and cannot be certified alone. The ISO assessment covers fairness, safety and societal effects, including in systems that process no personal data. BSI describes ISO/IEC 42005 as aligning with privacy and human rights impact assessments. Neither assessment replaces the other, and both can inform a wider AI risk assessment.
What information do we need before starting an AI impact assessment?
Gather the facts that every assessment draws on before choosing a template. Eight inputs recur across the sources:
- An inventory record of the AI system and its intended purpose
- Personal data flows and the controller and processor roles (UK GDPR Articles 26 and 28)
- The provider’s instructions for use and transparency information (EU AI Act Article 13)
- The categories of people affected, including groups with protected characteristics
- The period and frequency of use
- The human oversight design
- Complaint and redress routes
- Whether the organisation falls within ATRS mandatory scope
Which obligations apply depends on the organisation and the system, so check each against its source. Our FAQ on UK AI regulation by sector gives the wider regulatory context.
AI Security Gap Analysis
Our AI Security Gap Analysis identifies where AI is being used, maps the risks against regulatory and industry expectations and produces a prioritised action plan.