In many AI projects, technical discussions are rarely the problem. Model choice, integration, data quality, guardrails; these are issues with a solution, however annoying they may sometimes be. Where things truly stall is with the people who actually have to work with the output. And that is a different kind of resistance than we were used to. 

With previous automation, it was relatively easy to explain. With RPA, we told the employee: the 200 invoices that you are currently retyping manually? From now on, the robot will take care of it. Usually, that was good news. The work that disappeared was the work that nobody liked. The discussion was about the process, which exceptions, which fallback, who manages the bot, but rarely about who someone’s role and identity at work. The role remained intact, with the boring part taken out. 

AI does something else. AI does not affect the actions, but the judgement. And that is precisely the part from which people derive their craftsmanship and identity. 

From executing to assessing 

Take a claims handler who has been assessing files for fifteen years. That person has developed a feeling for when a claim is not right. That feeling is difficult to make explicit, but it is what he or she is asked for and appreciated for. Put a model next to that, which proposes a decision with supporting arguments in two seconds, and the role shifts: from assessing to assessing an assessment. 

That sounds like a small step. In practice, it is a fundamentally different skill. You must be able to estimate when a plausible-looking output is nevertheless incorrect. You need to know where the model is weak. You must dare to deviate from something that looks very convincing, and then be able to justify that. Thinking critically about someone else's work is different from delivering good work yourself. And not everyone who can do the second can automatically do the first. 

Why the resistance feels different 

The classic reflex is to frame this as fear of job loss. That plays a part, but it is not the whole story and, in my opinion, not the most important part either. I believe it is more the feeling of loss of status and autonomy. "I am the expert here, not the system." Or more subtly: people who consistently rate the output of the model too low, because accepting it feels like admitting that their experience is less unique than thought. That is not a crazy behaviour. That is someone who has invested twenty years in a certain way of working and is now being told that things are going to change. 

Organisations underestimate this from the start. The change trajectory is budgeted as a communication trajectory, a kick-off, training, some instructional videos, whilst the real task is that people have to redefine their professional identity. 

What managers can and cannot do here 

This is where leadership makes the difference, and where it often goes wrong. Not because of bad intentions, but because of avoidance. It is uncomfortable to tell a team that their work will be set up differently in two years' time. So the soft message is chosen: "AI takes over the boring work, you will have more time left for the fun things." Everyone in the room knows that that is not the whole story, and with that, you lose exactly the trust that you need. 

The truth is that honesty probably works better, even if it chafes. Explicitly state that the role is changing. Explicitly state what will disappear. And then be very concrete about what replaces it and how people grow towards that, including time, including room to make mistakes, including adjusted performance standards. Without the latter, it is a nice story without backing. 

"With Copilot, I can do this myself too" 

There is something else that directly stems from this. As personal AI tools improve, employees solve their own problems. Someone builds a prompt that does exactly what they need, shares it with two colleagues, and done. Great, there is real value in that, and you should not want to put the brakes on that. 

But it is not a solution for structurally embedding AI in processes. The prompt is in someone's personal environment. There is no version control, no logging, no ownership, no handover when that person leaves. In an audit, there is nothing to show. And you get thirty slightly different ways of working for the same process, which is exactly the fragmentation we wanted to get rid of ten years ago. 

The tricky nuance: this is not a choice between two flavours. You need both. Personal tools for speed and ownership. Generic agents and process integration for continuity, compliance, and securing knowledge.  

In conclusion 

The technology gets better, cheaper, and easier to implement every quarter. The human part does not. The ability of an organisation to let people accept a new role, and even more to embrace it, moves at a very different pace. 

If your AI programme stalls, look at the last part first. Not at the AI model or the tool.