What is it that humans do that AI can't?

A casual breakfast in Hong Kong drifted towards this thought.

I met David last month, a dear old friend and LSE classmate, now a professor at HKU Business School. When you meet someone after 14 years, time collapses a little. You are reminded of who you were before the weight of the world felt heavy. And when those early years were spent doing organisational psychology at LSE, the present has a way of pulling you back into the bigger questions: where is the world going, and what does it mean for us?

Between catching up on work and kids, we drifted towards AI. David said something that stayed with me - the essay questions professors are setting today are fundamentally different. Not because the subject has changed, but because they are asking for something AI cannot produce: a genuinely personal point of view, shaped by a life actually lived.

The paths our children will walk will look very different from ours. Their world is shaped by AI, the way ours was shaped by the personal computers and the internet while we were growing up.

So I found myself sitting with a question that David brought forth - what remains irreducibly human?

A decade ago, my list would have looked different. Today, AI can coach, counsel, predict emotion from voice and expression, write a persuasive opinion, and, somewhat ironically, even hallucinate. So what is left that is still ours?

Three things, I think.

Lived experience as judgement. AI can generate options. What it cannot do is weigh them against the grain of a specific context, relationship, or moment in time. I remember a negotiation program at IIM where, in a case study, I flagged that the real deadlock between two parties wasn't just about terms but also about how each felt they were being treated by the other. Resolving that dissonance was the unlock. That kind of read doesn't come from data. It comes from years of being in rooms and reading people, knowing what's actually at stake.

The warmth of human connection. AI can process emotional signals. But there is a difference between reading an emotion and meeting someone in it. Human emotions are layered, contradictory, and deeply contextual. I think AI will absorb the first level of support people need, it can even handle most of the available-at-2am crisis. But what lies beyond that still requires human cognition, presence, touch and the particular kind of care that comes from shared vulnerability.

The genuinely unpredicted thought. AI is trained on what has already happened. Its intelligence is, in a meaningful sense, retrospective. The human capacity for jugaad, for lateral leaps, for the question that comes from nowhere is harder to replicate. Though I'll admit: the meaning of "out of the box" is probably itself shifting. And I am very curious to see how that shapes.

In my own practice, AI now handles work that would once have gone to a junior. But the thing I keep returning to is what Kahneman would call intuition, pattern recognition built slowly through experience, encoded not in data but in judgement. That, for now, remains ours.

The question isn't whether AI will change work. It already has. But the capability that will matter most in the decade ahead isn't the one AI is building; it is the one we risk neglecting precisely because it can't be automated. Judgement, connect, the unpredicted thought. These don't develop by default. They require intention.