Agentic AI: what makes the promise hold

A promise seen everywhere

The word "agentic" has taken hold within a few months. Vendors, the press and suppliers describe agents capable of pursuing a goal, planning and acting. The promise is appealing, because it speaks to entire processes, not simple answers.

The movement runs deep. According to Gartner, a third of enterprise software could integrate this kind of capability by 2028 (forecast reported by L'Usine Digitale). So there's a real trend behind the noise.

But a market promise isn't a result in production. Before committing a budget to it, it's worth understanding what "an agent that acts on its own" really means.

"Autonomous", a false friend

The word "autonomous" either excites or worries, depending on the mood. In a business context, it deserves to be defused. The agents deployed today don't act unchecked.

They operate within a defined scope, with precise rules and controls. Their objectives are set upfront by the company. Sensitive actions remain supervised. This is governed delegation, not absolute autonomy.

This nuance isn't a detail. It's what separates an impressive demo from a system that holds up in production. AI proposes and executes. You arbitrate, you decide, you take responsibility.

The three keys to an agent that holds up

When an agent creates value, it's rarely by chance. Three conditions come up again and again, always the same ones. When they're missing, effectiveness drops fast.

• A defined scope. A clear objective, explicit action limits, a bounded use case rather than a vague, cross-cutting ambition.
• Reliable, structured data. An agent acts based on your data. If it's fragmented or unreliable, its decisions will be too.
• Continuous human supervision. Someone keeps a hand on the wheel, controls sensitive actions and corrects drift. This isn't a crutch, it's a condition of reliability.

One consequence follows. An agent that acts can't be generic. It must understand your business, your rules, your constraints. That's why offerings specialize by sector and by function, instead of promising a universal agent.

A concrete example

Take an agent assigned to customer service or operations. Where a simple chatbot merely answers, the agent executes a sequence within a bounded scope.

It checks the real status of a case in your systems. It diagnoses an issue by cross-referencing several sources. It triggers a corrective action, then updates the case and keeps the teams informed.

The gains are real when the framework holds. A telecom operator cut the processing time for certain tickets by 75% using this kind of agent (Orange ClariFibre case reported by L'Usine Digitale). The condition remains the same: a clear scope and usable data.

What decides the move to production

Here's the point the prevailing narrative often glosses over. Many agent projects never make it to production. This isn't a failure of the technology, it's a lack of framework.

Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027. The causes cited are prosaic: costs spiraling out of control, poorly defined business value, insufficient safeguards (Gartner press release, June 2025).

These causes aren't inevitable. They're matters of preparation. A vague success criterion, overly limited data access, weak governance: all points to address upfront, before deploying a single agent.

The controllable path

The good news is that the path is well known. It consists of bringing together the three keys before delegating, on a deliberately narrow scope.

• Choose a bounded use case with clear value, rather than an overly ambitious cross-cutting rollout.
• Check and prepare the data the agent will need.
• Name someone responsible, define what stays supervised and what can be automatic.
• Measure the results, then only widen the scope once value has been proven.

That's exactly what a tight first step allows: proving value on a pilot scope before industrializing. And that value is real: companies that target precise operational use cases report an average return on investment of 1.7 times (Capgemini Research Institute, via L'Usine Digitale). The return follows the framework, not the promise.

One last point deserves your attention. An agent that acts within your processes touches your data and your decisions. Keeping control of where that data lives, and of your ability to take back the reins, is part of the framework, not an optional extra.