With the rise of AI agents, the role of the Digital Service Infrastructure (DSI) is evolving from a traditional manager of digital tools to an architect of digital action. Previously, the DSI focused on providing applications, securing data, integrating systems, and ensuring their availability. Now, it must define what AI agents can do, control their actions, and measure the value they create within business functions. These agents can consult data, use applications, and trigger actions, adding a new layer of complexity to how companies manage their digital operations. AI agents can analyze customer requests, search for information across multiple applications, prepare decisions, create tickets, or even launch operations. This shift means the DSI must rethink how work is delegated, controlled, and measured. The question is no longer just whether a technology works, but what an agent is authorized to do, under what conditions, with what level of supervision, and with what responsibility. As these agents become more prevalent, managing their use becomes increasingly challenging. Who owns each agent? What data does it access? Which systems can it modify? How can its actions be verified? The DSI must ensure that each agent has a defined objective, a business owner, a technical identity, rights that match its mission, and conditions for being deactivated or modified. These are familiar concepts in application and access management, but they become more complex when an agent can interpret a situation and make decisions. For example, in a production incident, an agent might gather alerts, technical logs, and recent changes, then propose a cause and open an incident. Should it be allowed to restart a service, cancel a deployment, or modify a configuration? The answer depends on the system's criticality, the reversibility of the action, and the trust built over time. The DSI must formalize these levels of autonomy, ensuring that each action is appropriately authorized and monitored. Deploying AI agents is only the beginning. Their behavior must be observable and evaluated over time, including the data they access, the tools they use, the decisions they propose, and whether a human has corrected them. Technical metrics, such as the number of requests or model costs, are not enough to determine if a process is working better. The DSI must work with business functions to measure processing time, result quality, errors, human interventions, and the complete cost per operation. This shift also changes how operations are monitored. It is no longer just about ensuring an application is available, but verifying that agents act within defined limits and produce acceptable results. The DSI cannot alone define the value or rules of each process; business functions must remain responsible for objectives, decisions, and expected quality. The DSI provides the necessary architecture, integration, security, and operational support, while control functions help set appropriate risk limits. This collaboration must start with the process itself, ensuring that agents take on tasks that enhance efficiency without automating unnecessary complexity.