The Skills AI Cannot Touch (According to the Research)

The question everyone is asking is: which jobs will AI replace? It is the wrong question. Not because the answer is unimportant, but because it leads you toward the wrong unit of analysis. Jobs are bundles of activities. What matters is not whether your job title survives, but which of the specific activities within it will still require a human in five, ten, or fifteen years.

McKinsey Global Institute has been working to answer exactly this. Their Skill Change Index, published in the Agents, Robots, and Us research series (2025-2026), measures the automation exposure of over two thousand individual professional skills across the US and ten European economies, including the Netherlands.

The results are more nuanced than the headlines suggest. And more useful.

The skills with highest automation exposure

At the top of the Skill Change Index, the skills most exposed to automation over the next five years, are ones that are highly structured, rule-based, and digital in nature.

• Data processing and entry
• Quality assurance and routine checking
• Invoicing and financial record-keeping
• SQL programming and database querying
• Inventory management and tracking
• Routine writing and templated research

If a significant portion of your current work lives in this list, that is worth taking seriously. Not as a reason for panic, but as a reason for intentional movement.

The skills with lowest automation exposure

At the other end of the index, the skills with the lowest automation exposure are a different kind of list entirely.

• Coaching and developing others
• Leadership and team direction
• Negotiation and conflict resolution
• Caring and assisting roles requiring physical presence
• Ethical reasoning in novel, ambiguous situations
• Building and sustaining trust in complex relationships

Notice what these have in common. They all require something AI cannot structurally possess: embodied presence, genuine accountability, and the kind of judgment that comes from having lived through things. An AI can describe what good coaching looks like. It has never coached anyone. That distinction is not trivial.

The finding that changes the frame

Here is the part of the research that almost nobody reports. McKinsey found that 70 to 75 percent of the skills sought by employers today are used in both automatable and non-automatable work. The same skill that helps you process data also helps you synthesize insight from a messy conversation. The same communication skill that goes into a templated report goes into navigating a difficult client relationship.

Skills do not disappear. They shift. The question is whether you are building toward the shift or away from it.

This means the frame of "AI replaces skills" is not quite right. The more accurate frame is: AI takes over the low-judgment expression of a skill, while the high-judgment expression becomes more valuable. Writing as data entry becomes automated. Writing as strategic communication becomes irreplaceable.

What the three horizons look like through this lens

In the next one to five years, the urgent work is building AI fluency, not tool proficiency, but the judgment to use AI well. Demand for this capability grew seven times in two years, from roughly one million to seven million workers globally (MGI, 2025-2026). This is the fastest-growing skill in job postings worldwide.

In the five-to-fifteen year window, the structural advantage belongs to people who have developed the skills at the bottom of the automation index: coaching, leadership, ethical reasoning, relational depth. MGI finds that one-third of all non-physical work relies on social and emotional skills that remain beyond AI's structural reach. These are not supplementary. They are the core.

Beyond fifteen years, the research stops being useful and philosophy becomes the better guide. But it is worth noting that every capacity on the lowest-exposure list is also, arguably, what makes a life worth living: the ability to care for someone, to lead with integrity, to navigate genuine moral complexity.

What to do with this

Map your own work against this framework. For each major thing you do, ask: is this closer to structured and rule-based, or to judgment-dependent and relational? The first category needs your attention now, not to protect it, but to evolve it. The second category needs investment, because its value is increasing precisely because AI cannot touch it.

A caveat worth stating plainly

It would be dishonest to present the low-exposure list as a safe harbor with no further thought required. Coaching, leadership, and negotiation are low-exposure as categories, but a specific job with that title is not automatically protected if most of what fills the actual hours is templated and low-judgment. A manager who spends most of a week on status reporting and calendar coordination is doing high-exposure work under a low-exposure job title. The index measures skills, not titles, precisely because titles are a poor proxy for what someone actually spends their time doing, and the gap between the two is exactly where people get caught off guard.

A worked example of moving across the index

An operations coordinator whose role was mostly data entry and status tracking, squarely in the high-exposure category, did not wait for the role to disappear. Over eighteen months, she used AI to absorb the data entry and reporting almost entirely, and deliberately used the recovered time to take on the parts of the job nobody had gotten around to: mediating vendor disputes, training new hires, and building the informal trust that made cross-team coordination actually work. None of that was in her original job description. All of it was low-exposure, judgment-dependent work that had always been adjacent to her role but never central to it. Eighteen months later, the data entry part of her job barely existed. The rest of it had grown to fill the space, and it was the part that mattered.

The honest limitation of this framework

Not every role has an adjacent judgment-dependent version waiting to be grown into, and it would be misleading to suggest otherwise. Some roles genuinely were built almost entirely around structured, repeatable tasks with very little room for the kind of pivot described above, and for people in those roles the honest advice is different: build transferable skills deliberately, outside the role if necessary, rather than assuming the current job has room to evolve. The framework is useful for most people in most roles. It is not a guarantee, and pretending it is would be a disservice to the people for whom it genuinely does not apply.

• Audit your actual hours, not your job title, against the two ends of the automation index
• For high-exposure tasks, use AI to absorb them rather than compete with it on speed
• Identify the judgment-dependent work already adjacent to your role, even if it is informal or unrecognized
• Deliberately grow the time spent on that adjacent work as the structured tasks shrink
• If your role has no adjacent judgment-dependent work to grow into, build transferable skills outside it rather than waiting

The uncomfortable middle ground

Most people reading a piece like this want a clean answer to which category they fall into, safe or exposed, and the honest answer for most professionals is neither. Most real jobs are a blend, some hours spent on genuinely high-exposure tasks and some spent on genuinely low-exposure ones, in a ratio that shifts gradually rather than all at once. The useful question is not which category am I in, it is which direction is my ratio moving, and is that movement something I am actively shaping or something happening to me while I am focused on other things. Waiting for a clean signal that the shift is complete before responding to it means responding after the window to shape it deliberately has already mostly closed.

The McKinsey data is useful precisely because it turns a vague anxiety into a specific, actionable map. Anxiety about AI and jobs tends to stay diffuse and unproductive because it has no concrete shape, just a general sense of unease. A skill-level map converts that unease into a checklist: which of my specific skills sit where, and what is the deliberate next move for each one. That conversion, from anxiety to inventory, is most of what makes the framework worth using, independent of exactly how the automation timelines eventually play out.

It is also worth being honest about the emotional weight of this kind of decision, not just the practical logistics of it. Building deliberately toward the low-exposure end of a career sometimes means walking away from work that is comfortable, familiar, and currently well compensated, in favor of work that feels less certain and, in the near term, may pay less. That tradeoff is real, and dismissing the difficulty of it does a disservice to anyone actually facing it. The argument here is not that the tradeoff is easy. It is that making it deliberately, with a clear view of where the automation exposure actually sits, tends to produce better outcomes than avoiding the decision until it gets made for you.

The goal is not to make yourself AI-proof. Nothing is AI-proof. The goal is to build in the direction of what will always require a human, and to use AI to handle everything else so efficiently that you have more time to do exactly that.