AI agents: How do we manage these brilliant yet unstable new colleagues?

This op-ed written by Ghislain de Pierrefeu, AI expert at Wavestone, was first published in La Tribune, a French media, in May.

Whether you’re a startup looking to automate your sales or a large corporation aiming to optimize customer relations, logistics, or recruitment, AI agents — algorithms capable of autonomously handling tasks traditionally performed by humans — offer a compelling promise.

Riding the wave of Generative AI

This enthusiasm stems from recent advances in generative AI and language models, which now offer near-perfect understanding and generation of natural language, as well as streamlined navigation through dense documentation.

Agent-based systems operate on a simple principle: break down complex processes into smaller tasks, each handled by a specialized AI, often working in concert.

Take responding to a request for proposal (RFP) as an example: one AI extracts key information, a generative chatbot interacts with the documents, a predictive model suggests pricing, and another AI drafts the response. It’s fast and efficient, but still far from flawless.

A powerful promise, but a complex craft

Success stories abound from early adopters who are already reaping the benefits of AI agents in areas like market monitoring, communications, and marketing campaign management. These are ideal candidates for automation: they carry limited risk and often rely on large volumes of data, which help offset the occasional unreliability of AI.

However, for more complex scenarios, especially in large organizations where excellence is non-negotiable, expectations are higher. Human involvement remains essential, particularly for handling emotional, situational, and sometimes irrational factors that are still difficult to encode into algorithms.

Brilliant and overtrained…but unstable

Who among us would want to manage employees with an IQ of 150, but can’t explain their decisions, admit mistakes, or maintain moral consistency? That’s essentially the profile of today’s AI agents. Even their creators like Dario Amodei of Anthropic admit they don’t fully understand how these models work, as they learn autonomously by adapting to input data.

So, can we really rely on these agents to improve critical business processes? The answer is yes, for three key reasons:

• Agile competitors are already doing it. Companies like Qonto (in finance) and Alan (in health insurance) are using AI to reshape their industries. Large enterprises can’t afford to fall behind.
• AI agents are becoming embedded in mainstream solutions. Understanding how they work is essential to maintaining control over your processes.
• Despite their flaws, their capabilities often surpass ours in reading speed, translation, anomaly detection, and forecasting complex phenomena.

Know your AI agents like you know your team

Just like human employees, AI agents need to be selected, trained, and supervised. This means breaking down business processes into discrete tasks, identifying which ones can tolerate imperfect automation, and keeping humans in the loop where judgment is critical.

This requires clear governance: data quality, technology choices (vendors, platforms, open-source tools), continuous oversight, and most importantly, active involvement from business teams.

Ultimately, the real transformation won’t come from algorithms alone, but from our ability to rethink organizational structures and workflows around them. Governing AI will likely be the next major challenge for large enterprises.