An AI agent can check a customer record and prepare a reply. Can it also modify the record? Send the message? Trigger a refund? These are not the same decisions, and the fact that the agent can do something does not mean you should let it do it alone.
Agents that act within a frame
The trend is real. According to a Gartner forecast reported by L'Usine Digitale, a third of enterprise software could include agentic capabilities by 2028. But the agents companies deploy today do not run free: they operate within a defined scope, with goals set upfront and sensitive actions under supervision.
Before giving it access to tools, you need to write down what it can read, what it can modify and which actions require approval. You can perfectly well automate the information search and leave the sending or the refund to a person. Autonomy is set operation by operation.
A customer service agent, action by action
Take an agent assigned to customer service. On its own, it can check the actual status of a case in your systems and diagnose an anomaly by cross-checking several sources. It can prepare the reply and the case update. Sending the message to the customer, granting a goodwill gesture or triggering a refund can remain subject to an advisor's approval, at least at first.
The delegation limit then shifts with experience: an action approved a hundred times without correction can move to automatic, an action whose error would be costly stays under approval. Gains exist when that frame holds. L'Usine Digitale reports, for example, that Orange cut the processing time of certain tickets by 75% with an agent of this type (ClariFibre); that figure applies to that case, not to every agent.
What makes a deployment fail
In June 2025, Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027. That is a forecast, not a finding, but the causes cited are telling: escalating costs, unclear business value, inadequate risk controls.
They can be addressed before deploying anything:
- A bounded scope, with clear value, rather than a cross-functional deployment.
- Reliable data: an agent acts on your data, and its decisions are no better than that data.
- Tight access rights: reading a record does not imply being able to delete it, or access the whole database.
- A named owner, who decides what stays under approval and corrects drift.
- Measurement of results before widening the scope.
An agent that acts within your processes touches your data and your decisions. Knowing where that data lives and being able to take back control is part of the frame.
For a first deployment
For a first deployment, choose a limited process and decide in advance which actions remain subject to approval. You will then see what it is reasonable to automate further.
