The most important AI skill of the next few years has surprisingly little to do with prompting.
Almost every discussion about AI today revolves around models, agents, automation, and productivity.
What keeps me busy is a different question:
How do I prevent an AI from misleading me with an exceptionally well-written answer?
At first glance, that sounds paradoxical. After all, better answers are usually considered progress.
Yet over the past few years, I've noticed something:
The biggest risk of AI is often not the bad answer. It's the plausible one.
A bad answer is easy to spot.
A plausible answer is easy to approve.
And that's where the interesting problems begin.
These days, I use AI more often to challenge than to generate.
That tends to surprise people.
Most users start with the question:
"What can AI create for me?"
More and more often, my question is:
"What can AI dismantle in my approach?"
I start by developing a hypothesis, a structure, a concept, or a decision proposal myself.
Then I ask AI to look for weaknesses.
- Where is the evidence missing?
- Which assumptions haven't been tested?
- Which conclusions do not actually follow from the data?
- Which counterarguments have I overlooked?
I like this division of labor far better.
The human does the thinking.
The AI does the nitpicking.
[And before anyone worries: I'm not advocating for even more nitpicking in the world. I'm advocating for skipping the boring parts and getting straight to the interesting ones ;-) ]
Surprisingly enough, AI is often very good at exactly that.
The most valuable prompt may not be a generation prompt.
It may be a review assignment.
Not:
"Write a strategy for me."
But:
"This is my strategy. Find its weaknesses."
Not:
"Create a prioritization."
But:
"Which priorities would look completely different if my core assumption turned out to be wrong?"
That changes the dynamic entirely.
The AI is no longer producing the next plausible piece of text.
It becomes a counterpart.
A critical sparring partner.
And that is where I currently see the greatest value.
Many organizations underestimate the real challenge.
It is not called Agentic AI.
It is called data.
Again and again, I hear the expectation:
"We'll connect our data to AI and discover entirely new insights."
Maybe.
Maybe not.
When you combine a powerful model with an unstructured graveyard of data, you often end up with something quite dangerous:
Highly qualified uncertainty.
The output sounds convincing.
But it may be based on incomplete, contradictory, outdated, or barely discoverable information.
That has led me to an observation that occasionally runs against the current AI narrative:
"Unstructured data is often better suited for challenging a claim than for deriving a new truth."
"Find evidence against my assumption."
That works surprisingly well.
"Analyze all available data and tell me our strategic priorities for the next three years."
At that point, I become much more cautious.
In my view, Agentic AI is often explained the wrong way.
People frequently talk about agents as if their greatest advantage were automation.
- Writing emails.
- Processing tickets.
- Operating systems.
- Executing workflows.
Interesting?
Absolutely.
But that is not the core idea.
For me, the core idea is something else:
Agents can stay focused on a goal.
A traditional language model answers a question.
An agent pursues an intention.
It explores sub-problems.
It records intermediate findings.
It adjusts its search path.
Step by step, it works its way closer to the actual objective.
Which means the more interesting question is not:
"How autonomous will the agent become?"
But:
"Does it understand my real objective better than through a single, concise instruction?"
That is a subtle but important distinction.
Human in the Loop is often not enough.
The phrase appears everywhere.
A human reviews the result.
Problem solved.
Unfortunately, reality is not quite that simple.
A person can be formally involved while exercising almost no meaningful control.
- They may be overloaded.
- Under time pressure.
- Unable to see the intermediate steps.
- Or simply accustomed to approving everything automatically.
For that reason, I find a different term far more compelling:
Human in Control.
Control is not a switch.
Control is a slider.
The real question is:
What intervention capabilities does the human actually have?
Not:
Does a human appear somewhere in the process?
The real problem behind many AI-generated results is trust.
Most people have experienced this moment by now.
You receive a text.
Three paragraphs in, you think:
"This looks suspiciously AI-generated."
Instantly, perception changes.
Credibility drops.
Regardless of whether the content is correct or incorrect.
Some people call this phenomenon AI slop.
I think the term captures something important.
Not every automation-optimized answer creates trust.
Which is why I believe the question should always be:
Does this output improve my credibility?
Or does it diminish it?
It sounds almost trivial.
Yet surprisingly few people use it as their guiding principle.
My prediction for the coming years
I don't believe the world will split into people with AI and people without AI.
That phase is already behind us.
The more interesting divide will emerge elsewhere.
Between people who consume AI.
And people who shape it.
One group uses the tools they are given.
The other understands why those tools work, where their limits are, and when they should be used differently.
You don't have to be a developer to belong to the second group.
But you should learn to distinguish between generating, challenging, structuring, searching, evaluating, and acting.
Because increasingly, those decisions determine the quality of the outcome.
My conclusion
The most productive people of the next few years will probably not be those who trust AI blindly.
Nor those who reject it outright.
They will be the people who learn where it belongs.
Personally, my focus keeps shifting:
- Away from generating.
- Toward reviewing.
- Away from quick answers.
- Toward robust insights.
Because good AI does not replace thinking.
Good AI increases the resistance that thinking must withstand.
(Based on an AI-generated transcript of a 1:1 human interview. AI helped with the words. The conclusions are still my own.)