AI Agents: Autonomy Without Framework, a Risky Bet?

AI Agents: Autonomy Without Framework, A Risky Bet?

We are entering a decade where we will need to learn how to manage a new type of resource: agents and the unprecedented collaboration dynamics they establish between your teams and your AI agents.

Until now, the main question has been: "What AI use cases should we launch?" Useful, but already outdated. The real question becomes: "Are we capable of measuring, comparing, and improving the performance of our human teams and our agents with the same rigor?" As long as this question remains unanswered, agentic AI will neither scale nor become a driver of profitable growth.

Before going further, let’s clarify the terminology.

Analytics is not just "making graphs." It is a discipline that closely resembles economics: a way of observing reality to understand what creates value, what destroys value, and how to arbitrate between several imperfect choices. While economics focuses on markets, analytics focuses on operations. It is the art – and science – of transforming what is actually happening in the company into informed decisions.

For years, we have practiced a form of static analytics: dashboards, indicators, monthly reports.

With AI agents, we are shifting towards a form of dynamic, concrete analytics. Analytics no longer merely describes performance retrospectively; it becomes the very way to orchestrate work in real-time.

The First Disruption: The Notion of Role

AI agents are not employees, but they will reveal whether your company knows – or does not know – how to manage performance. An agent that responds to customers or prepares quotes is no longer just an executor: it must also be capable of making certain decisions to be autonomous. This autonomy requires defining a mandate and clear, precise, and measurable success indicators:

• What it is allowed to do alone
• What it must submit for validation
• What it must absolutely pass on to a human

As long as this mandate remains vague, the results will be inconsistent. The agent is blamed for decisions it should never have made, and praised for tasks that a script could have executed. Analytics then merely measures the confusion.

The Second Disruption: Context

An agent without context is like an analyst locked in a windowless room. It sees numbers, but it does not know what they represent. Initial feedback on agentic analytics shows that true value emerges when the agent understands the business vocabulary, the structure of customer relationships, and the company’s priorities, not just the data patterns.

This context is based on a simple yet demanding idea: methodical trust, built on governed data, shared definitions, and identified sources. Without this governance, data alone cannot provide the agent with the context necessary to perform its role. It is this trust based on context (“context-driven trust”) that allows for accepting that an agent can make proposals and execute actions without going through systematic human validation.

The Third Disruption: Supervision

An AI agent does not need an annual review; it needs a specific performance indicator. A continuous flow where we measure not only the volume of tasks completed but also:

• The success rate without human intervention
• The quality perceived by customers
• The cost per task
• The AI model used
• The frequency of escalations
• The types of errors made

At this stage, analytics ceases to be a rearview mirror. It becomes a nervous system: it picks up weak signals, alerts on deviations, and highlights blind spots. When an agent escalates multiple times on the same process, it is not just the agent that needs questioning; it is the process itself. Agents then become seismographs of our dysfunctions.

Ultimately, the issue is no longer whether you will use AI agents. You will, sooner or later. The challenge is whether your analytics is ready to welcome and manage them.

If analytics remains confined to dashboards that recount the past, AI agents will only add a layer of opacity to an already unclear system. If, on the contrary, you accept to make it a true science of hybrid work – with its hypotheses, models, experiments, and corrections – then agents will become what they were always meant to be: decision partners, not black boxes.

On Monday morning, instead of asking your teams: "What new AI use case should we pursue?", another question deserves to open your meetings: "How can we ensure that our employees and our AI agents have the necessary means to make the right decisions for the company, and clearly explain their effects?"