The real AI debate: Intelligence vs power

We are moving from AI that follows instructions to AI that pursues goals. And that changes everything.

The debate about AI is no longer primarily about whether it works well. Today, the real debate is about power — who decides, where authority begins, and where it must end. These are not new questions. They are the core questions of political philosophy. It is precisely there that we can look for inspiration on how to structure our relationship with AI.

The autonomy–accountability tension

In 1762, Jean-Jacques Rousseau’s Du contrat social proposed legitimate power exists only when those subjected to it are also its authors.

For the first time in history, we are surrounded by systems that influence us, decide for us, and act on our behalf without us having authored the rules by which they operate.

Artificial intelligence is no longer just a tool. It is becoming an actor. And that changes more than technology itself. It challenges how we design work, allocate responsibility, and preserve human agency. The real question is no longer whether AI performs well. It is what kind of relationship we choose to build with it, and what role humans retain within that relationship.

This tension sits at the center of today’s AI transformation. More policies, more compliance reviews, more governance boards will not resolve it. They sit on top of the system, not inside it. Incremental adjustments are no longer enough. A new social contract is the need of the hour

How AI evolved from assistant to actor

AI is the first non-human actor that is exercising delegated autonomy inside human institutions For years, AI was reactive. It answered questions, followed rules, and automated predefined steps. Agentic AI changes this model fundamentally. These systems understand intent, reason across multiple options, decide how to act, and adapt dynamically. Humans are no longer embedded everywhere in the process - they are involved selectively and deliberately. This marks a transition from deterministic workflows where every step is predefined to probabilistic ones where the system weighs options, judges confidence and risk, and decides when to act on its own.

This shift quietly reshapes the digital workplace. It is no longer a place where humans execute and machines assist, but where decisions are shared between humans and machines.

Power requires structure

As AI gains the ability to act, it also gains a form of power. And power, by definition, requires control. Three centuries ago, Montesquieu framed this principle clearly, laying the foundation for the separation of powers on which modern democratic systems are built.

"To prevent the abuse of power, it is necessary that by the very disposition of things, power should be a check to power."
- Montesquieu

The same principle now applies to AI. We no longer design every step. We design the objective and the guardrails. This is where governance becomes architectural, not procedural. Control must sit inside the system, not on top of it. Above this structure stands the human, not as an operator controlling every step, but as the sovereign: not operationally, but normatively. The one defining purpose, limits and responsibility.

Separating powers in agentic AI

This principle takes a concrete form in agentic systems:

The Doer: The agents that propose an action
The Critic: The agent that challenges it by testing risk, evidence and impact
The Judge: The agent that decides by calibrating the level of autonomy
This is not a linear flow. It is a system of checks and balances. Every check exists to protect human judgment, not to replace it.

AI may govern processes, but humans govern meaning, intent and accountability.

Not every decision needs a human

After the level of autonomy is calibrated (based on risk, confidence and impact), five modes emerge across autonomous systems:

Autonomous (Human out-of-the-loop) - AI acts independently.
Audit (Human-after-the-loop) - Actions are reviewed retrospectively.
Monitor (Human-on-the-loop) - Decisions are monitored in real time.
Approve (Human-in-the-loop) - Approval is required before execution.
Escalate or stop (Veto-driven) - Certain actions are blocked entirely.

This is not about handing over decisions to AI but about augmenting decision-making by extending judgment across the system.

It is important to note that not everything should be automated or escalated. Human judgment is applied exactly where it matters most.

Human-centric, not human-operated

This is what human centricity really means: not humans in every step, but human values, limits and intent embedded into how systems act on their own. However, presence is not the same as agency. A human who approves every step may have less real influence than one who defines the rules by which the system decides.

The impact of a human-centric system can be measured by whether their attention is spent on what actually matters like the judgment calls, the exceptions and the moments where being human makes the difference.

Where this becomes real: Contact centers

Now, this is not theory. At Atos, within Digital Workplace services, we are already applying these principles in environments such as customer support. Every day, systems receive incomplete signals:

“I can’t log in.”

“My system is not working.”

“I don’t know what’s wrong.”

AI must interpret, decide and act. And it does not wait for humans. It continues operating while routing attention to where it matters most. Cases are handled differently depending on the situation.

Critical issues are escalated immediately. Uncertain situations surface for human approval. And stable patterns are handled autonomously.

AI does not replace people. It orchestrates their attention.

The new social contract for the AI age

Work is not separate from social life. It is one of its most structured, consequential and visible forms. In workplaces, autonomous systems do not remain abstract; they allocate tasks, prioritize actions, escalate or suppress risk and influence outcomes that affect people directly.

This is why we need a new social contract for the AI age. And it will be born at work because that is where autonomy becomes real. This is where autonomy is exercised daily. This is where responsibility has clear, tangible consequences.

When a system is wrong, someone is accountable. When a recommendation is followed, someone accepts its impact. When autonomy is granted, someone defines its limits.

Work is where society first learns how to live with autonomous actors. Not in theory, but in practice. Not in isolation, but in a relationship.

From automation to accountable autonomy

The central question about AI is no longer technical performance. It is political legitimacy.

This is why the question of autonomy is not only technical, but fundamentally human.

We don’t give AI freedom. We give it responsibility, backed by human oversight. In Immanuel Kant’s philosophy, autonomy is the foundation of dignity. But autonomy is not the freedom to do whatever one wants. It is the ability to act according to principles one can recognize as universally valid.

AI cannot do this.

It can act and follow rules, but it does not understand moral obligation and cannot be held accountable for its actions. This is why every AI decision must remain auditable, explainable and ultimately human-owned. The shift, therefore, is from automation to accountable autonomy. Because ultimately, this is not only about the autonomy of our systems but about our own autonomy. We cannot build systems we do not understand, and we cannot delegate decisions to systems whose rules we have not agreed upon.

>> The real challenge is no longer technical - it is about design. If your organization is looking to balance AI autonomy with accountability, connect with me and let’s discuss.