Autonomy Advancing Without an Owner
According to Gravitee's State of AI Agent Security Report 2026, cited by Agenda Digitale, in 2026 83% of companies began working with autonomous AI agents—systems that do not merely respond to a prompt but plan, coordinate, and execute chained actions across multiple enterprise systems. Despite this uptake, only 7.2% of organizations have identified a person with formal responsibility for the behavior of their agents. More than six organizations in ten describe that responsibility as unclear or shared without a true owner.
The risk identified by the source is not a supposed rebellion of the systems, but the difficulty of reconstructing with certainty what an agent did, when it did it, and what data it relied on. It is a problem of attribution even before it is a problem of technology.
Back Office Under Pressure and the Costs of Incidents
From observation of companies' integration needs, it emerges that 80% of integrated AI agents are intended for internal back-office processes, while only 20% involve contact with external customers. The areas in which companies seek greater autonomy are the elimination of manual data entry and automated document reading (28%), data analysis and querying internal archives and procedures (28%), and the automatic drafting of document drafts or contracts (24%).
This marked internal autonomy makes it harder to isolate the individual step that generates a problem. The source reports that 47% of European CISOs consider autonomous agents the main threat to corporate security, ahead of ransomware and phishing, with average losses estimated at €4.2 million per incident at companies without specific governance. These costs include operational disruptions, non-compliance penalties, legal expenses, forensic consulting, and reputational impact.
Logging, Supervision and AI Act Obligations
Without a continuous log of the agent's reasoning, reconstructing an incident requires working backward through outputs and system requests, often with tools that can be trusted only in part. The source identifies three indispensable conditions for those who entrust operational tasks to agents, valid regardless of sector and size: recording of steps, data provenance, and the ability to reconstruct quickly.
The AI Act points in the same direction: for high-risk systems it requires automatic logging of events throughout the lifecycle (Art. 12), human oversight (Art. 14), and an obligation for users to retain logs (Art. 26). The source stresses that traceability is not a constraint on the speed of agents, but the condition for benefiting from their autonomy with confidence.
The methodological details of the Gravitee report, which are not described in the material, and their representativeness with respect to the Italian business landscape remain to be verified.
