Nominations now open — Submit before 15 December 2026 · Free to enterSubmit Now

AI Act Rights Impact Assessments Become Legal Evidence

★
Global Awards Admin
Editorial Team
DATENov 21, 2026
AI Act Rights Impact Assessments Become Legal Evidence

The EU Artificial Intelligence Act's fundamental rights impact assessment is usually filed under policy: a document prepared before a system goes live, signed off by a committee and stored. For the deployers obliged to carry it out, that framing understates the exercise. From 2 August 2026, when the Act's high-risk obligations for Annex III systems begin to apply, the assessment is better understood as an evidentiary record — a contemporaneous account of who may be affected by a system, what risks were identified, which mitigations were adopted and who is accountable for oversight.

The distinction matters because the obligation arrives alongside a broader shift in how legal teams are judged. Regulators, customers and courts rarely ask whether a document exists. They ask whether it is specific, whether it was produced before deployment rather than after a problem surfaced, and whether it is supported by records that fit together. An assessment that cannot be reconciled with procurement files, monitoring logs and board minutes is difficult to defend, however carefully it is drafted.

Who must assess, and why August 2026 matters

The Act sets a staged timetable. Prohibited practices applied from February 2025, obligations for general-purpose models followed, and, under the Act as adopted, the high-risk regime for systems listed in Annex III applies from 2 August 2026. The fundamental rights impact assessment sits within that high-risk phase.

Article 27 does not apply to every organisation using a high-risk system. It applies to deployers that are bodies governed by public law or private entities providing public services, with a limited exception for certain critical infrastructure systems, and to deployers of Annex III systems used to evaluate the creditworthiness of natural persons or to assess risk and pricing in life and health insurance. A bank scoring loan applicants, a municipality allocating social assistance, an insurer pricing cover: the common thread is not sector but the effect of the system on access to a service or a benefit.

Two features of the duty are easily missed. First, it attaches to the deployer, not the provider. A supplier's conformity assessment, technical documentation and market access formalities do not discharge it, because the assessment turns on the deployer's own context: its processes, its population, its oversight arrangements. Second, the results must be notified to the market surveillance authority, which converts an internal exercise into a regulatory communication. The assessment is also expected to complement, rather than duplicate, a data protection impact assessment under the GDPR where one is required — the two examine overlapping facts from different angles.

What the assessment must contain

The statutory content is deliberately concrete, and each element is a factual assertion that can later be checked against other records.

  • The deployer's processes. How the system will be used, for how long and how frequently.
  • Affected persons. The categories of individuals and groups likely to be affected in the specific context — not an abstract list, but the actual population the system will touch.
  • Specific risks of harm. The particular ways those groups could be disadvantaged, including through inaccurate outputs, inaccessible interfaces or over-reliance on automated recommendations.
  • Human oversight. Who reviews outputs, with what authority, and how that oversight is resourced in practice.
  • Mitigation and governance. What happens if a risk materialises, who handles complaints, and which internal functions are accountable.

Read together, these are the components of an accountability narrative. The difficulty is rarely knowing what to write. It is producing an account that survives comparison with the vendor's instructions for use, the deployment decision, the logs and the minutes.

From document to evidentiary chain

The assessment becomes useful when it is treated as one link in a chain rather than a standalone file.

Vendor information. The assessment depends on material the deployer does not generate: intended purpose, known limitations, performance metrics, instructions for use and incident notifications. Procurement should secure those items as contractual deliverables, together with change notification and audit rights, so that the assessment can be updated on the basis of information rather than assumption.

Deployment decisions. Someone approved the use of the system. The record should show who, when, on what evidence and what alternatives were considered. An approval that exists only in an email thread is weaker than one reflected in a minute.

Monitoring and logs. Deployers must monitor the system in operation and keep the logs automatically generated by it, to the extent those logs are under their control, for at least six months. Monitoring findings are the evidence that the risks identified in the assessment were actually watched.

Governance. Board or committee minutes that record the assessment, the residual risk accepted and the follow-up actions convert a compliance document into a management decision. Contract terms, meanwhile, should allocate responsibility for supplying the information on which each later version depends.

Keeping the record alive

An assessment written once and never revisited is the easiest kind to challenge. The Act expects it to be updated when the elements on which it rests change, and legal teams should define those triggers in advance rather than wait for an annual review cycle.

Practical triggers include a new model version or material update, a new use case or a new population, expansion into another jurisdiction, a change of provider, a serious incident, evidence of drift or degraded performance, a shift in monitoring findings and relevant supervisory guidance. Each should route to a named owner, produce a dated version and preserve the version it replaces. Retention periods should be set with regulatory requests and litigation in mind, not only with data minimisation.

How the assessment will be tested

Market surveillance authorities are the most immediate audience, and their questions tend to be granular: which groups were considered, what evidence supported the risk rating, who exercises oversight in practice. A generic list of rights-related risks copied across systems will not answer them.

Customers and counterparties are the second audience. Public sector buyers and regulated institutions may ask suppliers and partners to demonstrate how high-risk systems were assessed, and contractual warranties can turn those answers into enforceable statements.

Courts are the third, and the least predictable. Where harm is alleged, the assessment is likely to be read as the organisation's own account of what it knew before deployment. A document that identified a risk and committed to a mitigation, with no evidence the mitigation was implemented, is more damaging than one that recorded a narrower set of risks and showed how they were managed.

Three signals tend to distinguish a substantive assessment from a paper exercise: it names affected groups and explains why they are exposed; it records decisions and owners rather than intentions; and it can be reconciled with the logs, contracts and minutes that surround it.

The fundamental rights impact assessment will not be won on drafting. It will be won on whether the organisation can reconstruct, months or years later, how it decided to deploy a system, what it understood the risks to be, and what it did about them. Legal teams that design the assessment as a maintained evidentiary record — connected to procurement, contracts, monitoring and board governance — will be able to produce that account. Those that treat it as a pre-deployment formality will have a document, but not a defence.