Human capability in the age of AI

Human capability in the age of AI is the study of whether an organisation still holds the judgement, practice and accountability it depends on, as machines absorb the work those things were built from. It is a different question from whether the technology works, whether people are using it, or whether anyone has been trained. Those three are well covered. This one is barely measured at all.

The category, defined

An organisation's capability is the stock of judgement, skill and tacit knowledge held by the people in it, together with the work design that keeps replenishing that stock. Both halves matter. A firm can employ people who are individually excellent and still lose capability, if the work that made them excellent has been removed from the jobs beneath them.

The category asks one question in six forms: can this organisation still do the thinking it is accountable for? Not today, when the tool is working and the experienced people are still in post. In four years, when the people who learned the job the slow way have moved on and their replacements learned it a different way.

What this is not, and the four things it keeps being mistaken for

Category definition is mostly boundary work, so the useful part of this page is the part that says what belongs elsewhere.

• AI adoption asks whether people are using the tools. It is measured in seats, licences, weekly actives and completed onboarding. All of those can rise while capability falls, and the estate calls that pattern usage theatre. Adoption is a question about deployment. This is a question about consequence.
• AI literacy asks whether people understand the tools well enough to use them sensibly. Necessary, and separate. Someone can be fluent with a model and still have stopped forming a view before they prompt it.
• AI governance asks whether the system is controlled, documented and lawful. It has the strongest institutional backing of the four and the largest blind spot: most governance frameworks treat human oversight as the control that makes the rest safe. The evidence that oversight works as designed is weaker than the frameworks assume, which is the argument at human in the loop is not a safeguard.
• Skills training asks whether people have been taught something. It is measured in courses delivered and modules completed. Capability is built by doing consequential work and being answerable for it, and a course is a poor substitute for the repetitions that were removed to make room for it.

Each of those four has established owners, established budgets and established metrics. The gap sits underneath all of them: an organisation can score well on every one and still be spending capability it has not noticed it holds.

The problem, and why it stays invisible

AI rarely removes a whole job. It removes the first draft, the initial analysis, the routine review. Those were also the tasks through which people built the judgement that made them senior. The work looks the same from the outside and the output often improves, so nothing triggers an alarm.

The most direct evidence is clinical. Budzyń and colleagues found that adenoma detection in unassisted colonoscopy fell from 28.4 per cent before AI exposure to 22.4 per cent afterwards. The clinicians were the same clinicians. What changed was what they had been practising. The study is observational rather than randomised, covers one procedure in one country, and cannot fully exclude other changes over the period, so it establishes a pattern rather than a law.

The education equivalent runs the same shape. Bastani and colleagues found grades rose 48 per cent while an unrestricted AI tutor was available, and that when it was taken away those students scored 17 per cent lower than students who had never had it. Performance and capability moved in opposite directions, and only one of them was being measured.

The cost of not measuring it

The bill arrives late and in a form that does not obviously connect to the cause.

• Nobody senior enough to check. Verification requires the ability to do the work being verified. An organisation that automates the junior tier is choosing who will be able to supervise in a decade, at the missing rungs.
• Oversight that only looks like oversight. Across 106 experiments and 370 effect sizes, Vaccaro, Almaatouq and Malone found human and AI combinations performed worse on average than the better of human alone or AI alone. The losses concentrated in decision-making and there were gains in content creation, so the finding is conditional on the task rather than general.
• Fragility that only shows under stress. Capability is invisible while conditions are normal, and the test of it arrives when the system is wrong, unavailable or being used outside the range it was validated for.
• An expensive repair. Rebuilding judgement takes years of consequential work, and the years cannot be bought back. This accumulating gap is what the estate describes as capability debt.

The six dimensions

A category needs something measurable. These six are the dimensions this research will measure, each drawn from work already published rather than invented for this page.

• Judgement. Can people form a view, challenge the machine's, and own the decision? The diagnostic is whether anyone forms a position before prompting, at human at the start.
• Practice. Are people still getting the repetitions that build the capability the organisation is paying for? See missed reps.
• Verification. Is checking real or ceremonial? Someone who could not have produced the work cannot meaningfully check it, at the verifier's discount.
• Accountability. Is there a named human owner for an AI-assisted decision, and could that decision be reconstructed later? See auditing an AI-assisted decision.
• Origination. Can people still frame the problem and generate the question, or has the range of proposals narrowed? See does AI make everyone think alike.
• Resilience. Can the work be done when the system is unavailable or wrong? Dependence is measured by whether someone could still work unaided, at am I becoming dependent on AI.

Six rather than seven, deliberately. The seven SuperSkills are a different list doing a different job, and giving both the same count would invite the assumption that they map one to one.

Five stages, offered as a framework and not as a finding

The stages below are a way of organising a conversation. They have not been validated against outcomes, no organisation has been scored on them, and they should be read as a hypothesis about what deliberate practice looks like.

1. Unaware. AI adoption is measured through usage. Nobody has asked what is happening to capability, because nobody has framed it as a question.
2. Adopting. Tools are introduced with little redesign of the work around them. Individual productivity is the metric and the junior tier absorbs the change first.
3. Managing. Governance, policy and training appear. Risk is being handled. Capability formation is still assumed rather than designed.
4. Designing. The organisation makes explicit decisions about where judgement stays human, which work is kept unaided on purpose, and who owns which decision. This is the point at which drift becomes design, at drift versus design.
5. Capability-building. The six dimensions are measured, tracked over time, and acted on. Capability becomes something the organisation manages rather than something it hopes it still has.

Most organisations this research has worked with sit between the second and third stage. That is an impression from advisory work rather than a measurement, so treat it as one.

What SuperSkills is for inside this

The category is the territory. SuperSkills is one contribution to it, and separating the two keeps both honest.

The seven human skills describe what an individual practises. The six dimensions describe what an organisation is measured on. A person builds curiosity; an organisation either does or does not preserve origination. Those are related and they are not the same variable, and forcing them into one list would make the commercial framework and the measurement framework identical without evidence that they are.

The evidence base holds the graded sources. The questions map holds what is still open, including a good deal of this. The thesis makes the argument at length.

What this page does not claim

• That the category is original. Deskilling, automation bias and the ironies of automation have been studied for forty years. Bainbridge described this problem in 1983. The contribution here is translation into something an organisation can be measured on, not discovery.
• That the six dimensions are validated. They are a proposed structure. No instrument has been fielded against them and no scores exist.
• That the stages predict anything. They organise a conversation. Any claim that a higher stage produces better outcomes would need evidence that has not been gathered.
• That the terms here are coined. Capability debt, usage theatre and the verifier's discount are used without a claim of first use. No dated first publication is documented, and independent prior use by others is possible.

Key research and primary sources

• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful. Nature Human Behaviour, 8.
• Budzyń, K., Romańczyk, M., Kitala, D. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology and Hepatology.
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, O. and Mariman, R. (2025). Generative AI can harm learning.
• Autor, D. and Thompson, N. (2025). Expertise. NBER Working Paper 33941.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at work.

Related SuperSkills research

On the argument, the SuperSkills thesis and drift versus design. On the mechanism, capability debt, the missing rungs and synthetic seniority. On the measurement problem, measuring adoption properly and usage theatre. For boards, what a board should ask. For HR, the CHRO guide. For the state of the evidence, what we actually know.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page defines a category and proposes a structure for measuring it. By the evidence hierarchy used across this estate, the framework here is not graded evidence: the cited studies carry their own grades and limits at the evidence base, and the dimensions and stages are the author's proposed structure, unvalidated.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.