Agentic AI as a means of processing
An approach under which an agentic AI system is not a processing operation in itself, but one of the means by which the controller carries out processing of personal data. The controller is whoever determines the purposes and means of the processing (Article 4, point 7, of Regulation (EU) 2016/679), and the AEPD guidelines on agentic AI apply this to agents whatever form those means take, whether agentic AI systems or others (section V.A, p. 33).
Data protection obligations attach not to the agent but to each processing operation in which it acts.
Which obligations it carries
It is an interpretative criterion proposed by the AEPD guidelines, not a new obligation. The obligations are those of Regulation (EU) 2016/679 for each processing operation, applicable from 25 May 2018: among them, the principles of Article 5 with accountability, the record of processing activities under Article 30 and, where a high risk to the rights and freedoms of natural persons is likely, the impact assessment under Article 35.
What it is not
It is not a register of agents: the Article 30 record is a record of processing activities, and an agent appears in it through the processing operations in which it takes part. Nor does it allocate responsibilities by itself: each third-party service the agent accesses has the role that its relationship with the controller gives it — none if it receives no personal data linked to a person, processor if it processes them on the controller's behalf, or controller where they are disclosed to it for its own purposes — and the guidelines analyse this case by case (pp. 33-34); and where a processor infringes the Regulation by determining the purposes and means of processing, it is considered a controller in respect of that processing (Article 28(10)). And it does not leave out what the agent itself generates: the guidelines call for taking into account the additional processing of data that the components of the system may give rise to (p. 33).
The nuance almost nobody captures
Being a means does not make the agent a neutral one. The guidelines warn that including agentic AI as a means of processing introduces new uncertainties (section V.G, p. 40) and note that the controller has to manage the new risks its use may create (section V.A, p. 33). That is why the analysis is done processing operation by processing operation: the same agent may be harmless in one and decisive in another.