While AI assistants respond to instructions, AI agents act proactively to achieve a given objective. This evolution marks a major shift that we are currently observing among our clients: we are moving from a logic of automation to one of supervised autonomy. Agents operate within a dynamic environment, guided by objectives and boundaries defined by humans.
The question is therefore no longer whether organizations will adopt agentic AI, but how they will organize its governance and supervision.
What mechanisms should be put in place to effectively supervise an AI agent?
Bertrand Merlier: We observe that supervision is the primary challenge of agentic systems. The value of the expert no longer lies solely in producing an outcome, but in designing a framework that ensures the quality and compliance of the decisions made by agents.
Three key mechanisms are essential:
- Clearly defined performance indicators: acceptance criteria, expected quality levels, and quantified targets for time, cost, or resource consumption.
- Full traceability of actions performed, enabling organizations to understand decisions made and facilitate audits.
- A level of supervision tailored to the criticality of tasks, ranging from occasional reviews to systematic human validation.
How can trust be built between teams and AI agents?
Bertrand Merlier: The arrival of AI agents is fundamentally reshaping the distribution of work. The question is no longer which tasks should be automated, but how to organize effective collaboration between humans and autonomous systems.
The challenge is to establish a clear framework in which each party remains within its role: employees retain control, while AI agents operate strictly within their designated scope of execution. AI brings speed and processing power, but judgment, arbitration, and accountability for decisions remain exclusively in human hands.
How can organizations balance AI agent performance, cost control, and digital sustainability?
Bertrand Merlier: The effectiveness of an agentic system should not be measured solely by its business performance. It must also take into account operational costs, resource consumption, and environmental impact.
An agent may perform dozens of model calls to complete a single task. Every token consumed, every model invocation, and every execution loop carries a cost.
Optimizing an agentic system requires a constant trade-off between performance and the efficient use of resources. This balance is essential to ensure system effectiveness while limiting its environmental footprint.
Toward governed AI that creates value
Agentic AI represents a new stage in digital transformation. To fully unlock its potential, its adoption must be both thoughtful and responsible.
At Sopra Steria, this conviction is embodied through the AI Backbone, a shared infrastructure that enables employees to develop and use AI agents within a secure, governed, and economically controlled framework.
This approach demonstrates our ability to intelligently orchestrate specialized AI agents throughout the entire solution lifecycle. By enabling them to collaborate at scale, we are reinventing systems engineering and delivering tangible, measurable value to our clients.
Tomorrow, organizational performance will not depend solely on the ability to use AI, but above all on the ability to govern it effectively.
Author:
Bertrand Merlier
Solution Architect and AI Expert, Sopra Steria Switzerland