AI agents are rapidly gaining a place within organizations: digital assistants that independently perform tasks, prepare decisions, and manage processes. Convenient — but also exciting and sometimes unsettling. How do you deploy such agents safely without losing sight of the human element?
That’s what we discuss with Wouter van der Harg, AI expert at CGI. He designs new AI solutions, often using AI agents. He also guides and advises organizations through this transition. Wouter shows how agents do not replace people, but instead create space for more meaningful work — and for the next step in your career.
“It’s crucial that you automate the right things without employees losing their autonomy and sense of control,” he says. But how do you determine which tasks to entrust to an AI agent — and which ones you prefer to keep in human hands?
This article previously appeared on Begeleidingsethiek.nl (Dutch)
Control first, then AI
According to Wouter, working safely with AI starts with a solid foundation. “At CGI, we only start using AI once the guardrails are watertight,” he explains. “If we’re not in control or not certain about what happens to data, we simply don’t use it. Period.” This approach enabled CGI to be among the first to sign the European AI Act. Some of the preconditions: their ChatGPT servers are located in Europe — no exceptions — and client data may never be used to train models.
According to Wouter, the idea that an AI agent constantly needs to “learn” is often overestimated. Like humans, agents draw on previous experiences. Actual model training rarely happens. “Most of the time, it’s about adjusting the instructions given to the agent, not retraining the underlying language model itself. We only train the language model at a very specific, predefined moment. That requires significant computing power and supervision.”
Agentic AI, therefore, is less about “magically learning systems” and more about clearly formulated tasks within well-defined boundaries. Humans remain firmly in control. From that solid foundation, Wouter looks beyond risks and asks: what can AI agents mean for people’s work?
Less administration, more valuable interaction
In this transition, Wouter always starts from a simple principle: people thrive through contact with other people. “I strongly believe that the more people’s work consists of interaction and connection, the better and more fulfilled employees feel.”
From that perspective, he examines which tasks are not about human interaction — and are therefore most suitable for automation. He also sees an important ethical dimension here. “The question is: how do you ensure you automate the right things? What I sometimes encounter is that you could automate tasks that employees themselves see as their core responsibility.”
He gives the example of someone who manages databases and transfers data from one system to another. “AI agents can now take over that entire process. But if you take that away from someone without consideration, that person may start to feel useless.”
A sharper look at ‘Human in the loop’
These concerns about human dignity and autonomy were part of a session on guidance ethics about Agentic AI with procurement professionals. Guidance ethics is a structured dialogue method in which all stakeholders explore the impact, values, and concrete actions surrounding new technology. These dialogues help guide technological developments responsibly in society, in line with our shared values.
During this session, one of Wouter’s own assumptions was nuanced: the well-known concept of “human in the loop.” This principle ensures that a human remains actively involved in assessing, steering, and approving decisions made by an AI agent. “Especially when sensitive data or high-impact decisions are involved, humans must of course remain decision-makers,” Wouter explains.
However, during the session — thanks to conversations with procurement professionals — he discovered that this principle can be applied more precisely. For certain well-defined parts of the procurement process, constant human oversight is not necessary, as long as the boundaries are clearly defined in advance.
“If someone knows a process so well that they can help design it and validate beforehand whether it stays within agreed parameters, you can fully automate that part,” he says. “The human is no longer involved in every single step, but instead at the design stage. That was a real eye-opener for me: not removing humans everywhere, but looking more carefully at where human intervention truly adds value.”
The human as manager of AI
Because AI agents can take over much of the operational work, more space is created to think about how something should work. “Right now, due to all the paperwork and red tape, procurement professionals often have little time for conversations in which they can truly explore what the client needs. We see that burden across many fields,” Wouter says.
Automation creates more time for genuine conversations with end users: how do you want it to work? What is practical? What feels intuitive? “These are crucial questions where human interaction is indispensable — and where the end product truly improves as a result,” he says. “The agent executes, but people determine the direction and the quality.”
In this next phase, he sees employees as “the managers of AI.” The employee directs the AI agent within their field of expertise, precisely because they understand what the agent should be doing and can assess whether the process remains on track.
A More Level Playing Field
According to Wouter, AI not only increases efficiency and makes work more meaningful, but can also make processes fairer. “In practice, people are not completely independent,” he says. “A procurement professional may unintentionally have a preference for a particular supplier.” That makes it difficult for new parties to enter the market. “Yet especially in government, you want new market players to have a fair chance.”
That is why Logius is exploring the possibilities of an AI-based procurement advisor. This concept was also examined in a guidance ethics session. “Such an AI agent can look at a tender more independently,” Wouter explains. “When there is a request for proposal, the system can validate whether a supplier meets all criteria. And if not, that doesn’t immediately mean disqualification. The agent can also provide targeted feedback.”
This reduces the gap between large, experienced parties and a starting SME. According to Wouter, this demonstrates that AI does not necessarily have to reinforce inequality through bias in data — a common concern. “By consciously investing in these kinds of applications, AI can actually create a more neutral playing field and more equal opportunities.”
Be Critical of Your Own AI Assumptions
Getting started with AI agents also means remaining critical of your own perceptions. “There are many ideas about AI that are not always based on facts,” he explains. “Things you may have heard in a blog, vlog, or video. But innovation moves so fast that knowledge from six months ago may already be outdated. So check your assumptions and test your knowledge against the most recent information.”
Finally, Wouter has a clear call to action for organizations: “Have the conversation with each other. Share concerns and fears openly and genuinely pay attention to one another. But also explore the opportunities together, and bring people along patiently.”
His conclusion? If we use AI to remove administrative burdens and free up time for real connection, technology can actually make work more human — and improve outcomes. “The more room there is for genuine conversations, the better you can uncover what users truly need.”
