Source: Bird & Bird (Two Birds insights)

An agentic AI system is defined as a goal-driven loop that plans sub-tasks, calls tools and APIs, checks outputs against objectives and re-plans autonomously; it differs from reactive generative tools in autonomy, real-time API perception, action (sending emails, booking, signing), proactivity, planning and persistent memory. Those traits create three heightened GDPR risks: scale and speed across broad datasets, unanticipated sensitive inferences, and real-world automated actions that can occur before human review. Examples include lending agents declining applications, HR agents sending rejections, diagnostic agents booking follow-ups and dynamic pricing agents adjusting offers in real time.

Regulators have already issued targeted material: the EDPS framed human oversight as a new privacy obligation (Sept 2025), the UK ICO mapped eight novel agentic risks (Jan 2026), the Spanish AEPD published a 71-page technical guide (Feb 2026), and the Dutch AP opened a public consultation on the right to explanation (Apr 2026). Controllership can shift mid-process when agents call third-party APIs or sub-agents; a worked RecruitBot example shows one task producing four sub-calls that assume different GDPR roles (independent controller, joint controller, controller-to-controller, and a true processor). Recommended mitigations include mapping every agent data flow, classifying each actor per step, refreshing vendor DPAs to cover dynamic sub-processors and agent memory, updating ROPA and privacy notices, applying least-privilege architectures, whitelisting approved tools, tagging lawful purpose per pipeline, and adding agent-specific transparency measures. Purpose limitation and lawful basis are at risk when agents infer additional purposes (e.g., leadership, age, culture fit) beyond the original collection purpose; architecturally enforced restrictions and pipeline-level purpose tags are advised. Article 22 exposure is a central compliance pinch-point: the two-part test—solely automated decision and legal or similarly significant effect—must both be met. Likely in-scope scenarios include underwriting decisions, CV screening that filters candidates before human review, and account suspensions without human appeal; rubber-stamp reviews do not remove Article 22 scope. The CJEU has broadened the concept of automated decisioning (SCHUFA C-634/21) and set explanation standards in Dun & Bradstreet (C-203/22, 27 Feb 2025): organisations in scope must provide exceptions or safeguards including meaningful human intervention, a channel to express views, contestation mechanisms, and both general and person-specific explanations.