The prevailing intuition says: AI advances, therefore more technical skills are needed. The empirical evidence says the opposite. According to the OECD Employment Outlook devoted to AI and the labour market, employers report that AI has increased the importance of specialist skills — but has increased the importance of human skills even more (OECD, 2023). The thesis of this article is that the human capital of the coming decade will be hybrid: technical enough to converse with the systems, human enough to do what the systems do not.
What the evidence actually says
A second OECD study, on the demand for skills in AI-exposed occupations, adds the decisive piece: most workers exposed to AI will not need specialist AI skills. They will need different skills — and at the top of the list of those most demanded in high-exposure occupations sit management and business capabilities: project management, coordination, administration (OECD, 2024).
The logic is economic, not sentimental, and was already visible in every previous technological transition: the spreadsheet did not make fast bookkeepers valuable, it made valuable those who knew what to calculate. When a capability is automated, its price collapses; value migrates to the complementary capabilities. If the system produces the analysis, the person who can pose the right question and recognise the wrong analysis is worth more. If the system writes the draft, the person who can judge what should not be written at all is worth more.
It is not a race to learn what AI can do. It is a race to excel at what AI makes scarce: judgement, relationships, accountability.
The four families of hybrid work
Judgement under uncertainty. Deciding with incomplete information, weighing incommensurable risks, owning the error. It is the heart of what we elsewhere call decision capital, and no model replaces it: models estimate, they do not answer for the outcome.
Competent interrogation. The value of the output depends on the quality of the question and the capacity to verify it. Competent interrogation requires understanding the domain: a finance director questioning a model about cash flows sees errors a prompt specialist never will.
Relationships and trust. Negotiating, motivating, settling conflicts, representing the organisation. These are activities in which the counterparty is human and demands a human interlocutor — not from nostalgia, but because trust is granted to someone who can answer for it.
Cross-domain integration. Connecting what the systems treat separately: the market signal with the production constraint, the legal datum with the commercial choice. It is the characteristic competence of middle management — the layer many organisations are cutting precisely when they need it most, as we argue in our analysis of middle management.
The T-shaped profile turns over
For years the model of talent was the T-shaped profile: deep vertical specialisation plus a horizontal base of transferable capabilities. AI is reversing the proportions. Vertical depth is exactly what the systems replicate best: codified knowledge, procedures, the recurring patterns of a domain. The horizontal bar — connecting domains, translating between technical languages, reconciling points of view — is what remains scarce.
This does not mean specialisation dies: it means it stops being enough. The specialist who knows only his own vertical now competes with a system that knows it almost as well, at zero marginal cost. The specialist who can also interrogate, verify and connect uses that system as a lever — and his vertical, instead of depreciating, multiplies.
For organisations the consequence is concrete: skills taxonomies built for pure verticals — and the career paths that reward them — select for the profile that is depreciating. Revising them is not an HR project: it is maintenance of competitive advantage.
How corporate training changes
If the skills that matter are judgement, interrogation, relationships and integration, the training that develops them cannot come from a catalogue. Three shifts define the training that works in the post-automation era.
From content to cases: judgement is not transferred by slides, it is trained on real decisions — ambiguous, incomplete, with something at stake. Classroom hours produce vocabulary; cases produce capability.
From the individual to the decision chain: training one person at a time leaves untouched the processes in which people decide together. The most advanced organisations train whole chains — those who propose, those who prepare, those who approve — on the same case, because that is where collective judgement is actually exercised.
From the certificate to the test: a programme nobody can fail selects for nothing. Serious verification — recognising a flawed analysis, arguing a dissent, stopping a project — is uncomfortable and therefore rare. It is also the only kind that distinguishes training from compliance, as with the literacy the European regulation requires.
The risk of the technical monocle
Many organisations are responding to AI with wholly technical training plans: tool courses, platform certifications. Useful, but built on a fragile assumption — that competitive advantage lies in the use of the tool. It cannot lie there: the tool is the same for everyone and is bought at list price.
The signal also comes from the labour market. The organisations that hire best have stopped asking for lists of tools in their requirements and started testing situations: an automated output with three hidden errors to find, a decision to argue with incomplete information, a technical disagreement to settle. The interview that simulates real work selects the hybrid profile; the list of platforms selects whoever updates a CV well.
The advantage lies in the combination: people who understand the technology well enough to use it without being used by it, and the business well enough to know where the tool lies. The OECD also notes the reverse side: two employers in five cite the lack of adequate skills as a barrier to adoption. The answer is not to train a hundred specialists; it is to raise the minimum threshold of understanding across the whole management line — the same principle the European regulation turned into an obligation with AI literacy.
What this means for those who lead people
First: rewrite job descriptions starting from decisions, not tasks. Tasks get automated; decisions get guarded; a position is justified by the judgement it exercises.
Second: in selection, weigh the capacity to verify as heavily as the capacity to produce. The candidate who spots a wrong analysis is worth more than the one who produces ten in an hour.
Third: protect the paths where judgement is formed. Judgement grows from guided experience — hard cases faced beside someone who has seen them before. If AI absorbs all the entry-level work, the organisation stops producing tomorrow’s seniors: an invisible debt that accrues compound interest.
And there is a responsibility for those reading these lines from the top: hybrid skills spread by imitation before they spread by training. An executive committee that verifies outputs instead of accepting them, that rewards the awkward question over quick consensus, teaches more than any course catalogue. The first classroom of human capital is the room where decisions are made.
Human capital is not what remains once technology has taken its share. It is what determines how much the technology is worth. The organisations that understand this first will buy the same tools as everyone else — and obtain results the others will not see.