Agentic AI in HR: "You don't use an agent, you collaborate with"

Arago: Agentic AI is becoming a boardroom priority, yet many decision-makers struggle to grasp its real implications. Where do you start to understand what you are actually dealing with?

Karam El Wardi : "There is a quote I love: 'The greatest shortcoming of the human race is our inability to understand the exponential function.' It comes from Professor Albert Allen Bartlett, and it says it all.

When I speak to leadership teams, I often ask them this: if you take 30 steps in a linear fashion, you travel 30 or 40 metres. Predictable. But if each step doubles the previous one, exponentially, how many metres have you covered by step 30? The answer: 1,073,741,824 metres. Enough to go to the Moon, come back, and go again.

That is exactly where we are with AI. Its evolution is not linear. It doubles, every month, in performance and capability. And that is why when clients ask me for a 5 or 10-year ROI, I tell them honestly: that ROI will be wrong. Not because I don't want to help you build it, we can, and we will if you need it. But it will be based on what we know today, not on what AI will be in three years.

For an ERP like SAP SuccessFactors, projecting a 3-year ROI is legitimate: the technology evolves linearly, two or three times a year, predictably. For agentic AI, it is fiction. Capabilities that existed six months ago are already obsolete."

So how do you move forward if you cannot plan ahead?

"You go back to basics. Faced with the unknown, you move step by step, on what you understand. And to understand agentic AI, you first need to understand what an agent is.

That word is everywhere, it is overused, and that is a problem, because you cannot adopt what you do not understand. And above all, you cannot 'use' an agent. You 'collaborate' with one. That is not a semantic detail: it is a complete paradigm shift."

So what, concretely, is an AI agent?

"An agent is four components working in a loop.

First, a brain: what we call a Large Language Model (LLM). That is what reasons, plans, and understands intent. Without it, you have a script, not an agent. Then, tools. An agent can read, write, and act on real systems. It is not stuck behind a pane of glass looking at your data without being able to touch it. Without tools, you have a chatbot. There is also memory: the agent remembers context and interaction history. Without memory, it is a simple question-and-answer exchange. And finally, autonomy: it decides when to act, when to ask for help, when to stop. That is what fundamentally sets it apart from a command or a script.

Chatbot, script, or agent, they are not the same thing. A chatbot answers. A script executes. An agent decides, acts, measures, adjusts, and can replace an entire workflow, not just a single task.

What matters to understand is that an agent resembles a human. Not by accident, but because humans created it. And since the dawn of time, when humans create something, it resembles them."

And this resemblance to humans, what does it change about how we evaluate them?

"Everything. When a candidate comes to interview with you, you do not ask them about their memory capacity or how fast they type. You assess their skills, their expertise, what they have already done, their references. An AI agent is recruited in exactly the same way.

The problem today is that agent vendors will highlight power, speed, and the incredible memory of their solution. It is like a candidate telling you in an interview: 'I have an excellent memory and I type very fast.' That is not what counts.

What you need to ask is: does this agent know my sector? Does it understand collective agreements? Does it know the logic of my HRIS? Can it audit my SAP SuccessFactors configuration? Does it have experience in my type of organisation?

A generic agent that knows everything yet nothing in depth is like a brilliant generalist colleague with no specialism. It will not be truly useful to you."

There is much talk of AI pilots that never make it to production. Do you observe this too?

"Yes, and it is systematic. Take any high-performing LLM, set it against an HR business function: the first 10 minutes are impressive. Then five gaps appear, always the same ones.

The agent is not expert: it reasons in a generalist way, ignoring your policies, local rules, and business specificities. It is not connected: it does not read your systems in real time. How many employees are on SAP SuccessFactors? How many open tickets on ServiceNow? It does not know. And an agent that does not know cannot augment anyone. It is not reliable: ask the same question five times and you get five slightly different answers. That is not acceptable in a business, and even less so under the European AI Act, which requires traceability for AI-assisted decisions. It is not understood: it is a black box. Neither IT nor the DPO can explain why it gave that answer rather than another. And it is not secure: the agent's access rights are not aligned with those of your HRIS. You create a back door in your security architecture.

A single one of these five gaps is enough to kill a pilot. Not five, just one."

So what makes a good AI agent?

"A good agent combines five qualities simultaneously, because none of them are optional.

It is expert: it knows your business, with tested and documented competencies grounded in your actual policies. Not 'I know HR in general', but 'I know HR in France, collective agreements, and your SAP SuccessFactors instance.' It is connected: it reads your systems in real time (SAP, ServiceNow, Microsoft 365, Cornerstone, Jira...). Not yesterday's copy, live data. It is reliable: every response is sourced, traceable, and auditable. Same question, same answer. This is what we call determinism, non-negotiable for compliance. It is understood: IT and the DPO can inspect the reasoning loop and understand why the agent said what it said. And it is secure: it inherits the rights and permissions of your HRIS. No second door. A profile without access to payroll in SAP SuccessFactors does not access it through the agent either.

These five qualities combined are what takes an agent from gadget to genuine digital collaborator."

How does an agent's skill development work once it has been deployed?

"Exactly like with a new colleague. When someone joins an organisation, you do not ask them to draw up a 5-year career plan on their first day. You evaluate them, give feedback, and they progress.

With an AI agent, it is the same. You ask questions, you give feedback: 'that answer was perfect', 'that one does not reflect our company culture', 'next time, present it differently.' Every piece of feedback is encoded. The agent trains on it, refines, improves. Until you reach the sweet spot, the response that matches exactly what you expect, in your context, with your company culture.

At that point, you no longer talk about 'using' an agent. You collaborate with one. And that is the real transformation."

What role is left for humans in this model?

"A central and irreplaceable one. An agent makes mistakes. All current agents make mistakes, because the moment you allow it to reason, there is a margin for error. It is inherent. It is therefore impossible to have a fully autonomous agent on sensitive decisions. A human must oversee, validate, and correct.

I often use the image of tennis. What makes a match beautiful is not the player who hits hardest. It is the rally. The value lies in the quality of the exchange between the two. Having the world's most powerful agent is useless if the governance between it and the human is poor. The value is not in the intelligence of one or the other, it is in the balance you find between them."