From tokenmaxxing to token discipline: Orange considers how to measure the efficacy of agentic AI in network operations
Orange Fellow and VP of Cloud and Software Engineering Philippe Ensarguet recently discussed the path to agentic network operations; he characterized the journey as a progression that moves from assisting humans, to recommending actions, to acting within defined boundaries, and eventually to orchestrating operations across domains. The critical requirement is that operators earn trust and prove business value at each step rather than treating autonomy itself as the objective.
Ensarguet said success is not about the number of agents running or tokens consumed. Rather, it is operational: incidents resolved faster, fewer customer disruptions and less manual coordination. The path there, he said, moves through “assist, recommend, act, and orchestrate.” Assistive AI makes humans faster and better informed; supervised agents take on more cognitive load; bounded autonomy expands what systems can do; and only then does cross-domain orchestration become realistic.
“Assist first, earn trust, expand the scope,” Ensarguet said. Operators that try to jump directly to autonomy may produce a “very, very impressive demo,” he warned, while leaving operations teams deeply uncomfortable.
Orange is already applying that progression to concrete network domains. At Orange OpenTech 2025, the operator demonstrated LiveCentriX, an agentic, multi-agent approach to 5G security monitoring. Ensarguet also pointed to telco cloud root-cause analysis, RAN energy optimization, RAN lifecycle management and incident management. Orange’s work in the Linux Foundation Europe-hosted Sylva project provides a broader cloud-native substrate for this direction; an April 2026 Sylva white paper explicitly connects standardized telco cloud infrastructure with closed-loop automation, distributed intelligence and autonomous operations.
But Ensarguet’s key filter is not technical feasibility. Orange uses what he described as a “business value first selection,” including a methodology called “high value scenario” to identify problems that can create value across Orange affiliates. As he put it, “The use cases that scale for me are the ones chosen for business value, not for technical implementation.” And, “We need to start with the problem and not with the technology.”
That discipline also applies to AI economics. As token consumption becomes a real operating cost, Ensarguet argued that aggregate token counts say little on their own. “I would treat tokens as unit cost, but measure them at a level of real outcome, token cost per assisted decision, per resolved case, per completed workflow, not total token conception.”
The measurement chain then runs from AI use, to a changed decision, to an operational KPI, to measurable business value. Each link needs evidence and, where possible, A/B testing or phased rollouts with control groups. Otherwise, Ensarguet argued, operators risk mistaking correlation for causation.
That may be Orange’s most useful contribution to the agentic network discussion. Autonomy is not a leap. It is an evidence-based progression. Instrument the causal chain early, prove value at each step and let earned trust determine how much agency comes next.
