When Technology Outruns Competence
We are used to discussing artificial intelligence primarily as a threat to jobs.
But another problem may become at least as important:
technology may develop faster than the human and organizational capabilities required to understand, implement, verify, and maintain it.
If that happens, the main constraint on economic growth will not necessarily be computing power, software, or access to AI.
It may increasingly become complementary human and organizational capital.
The central risk is not that technology becomes useless.
It is that technological capability grows faster than society’s ability to convert that capability into reliable economic output.
The bottleneck may be shifting
AI is becoming cheaper, more capable, and more widely available.
According to the Stanford AI Index 2025, the share of surveyed organizations reporting AI use increased from 55% in 2023 to 78% in 2024. The share using generative AI in at least one business function rose from 33% to 71%.
Global corporate AI investment reached $252.3 billion, while private AI investment in the United States alone amounted to $109.1 billion.
Source: Stanford AI Index 2025.
https://hai.stanford.edu/ai-index/2025-ai-index-report/economy?utm_source=chatgpt.comAn enormous accumulation of technological capital is taking place.
Technology itself is not disappearing as a constraint. AI systems still face problems involving reliability, hallucinations, energy consumption, data quality, cybersecurity, semiconductor capacity, and integration.
But in many sectors, the bottleneck may gradually shift from:
access to technology toward:
the ability to deploy technology effectively.
Technology by itself does not solve a problem.
AI can write a program, perform a calculation, produce a technical diagram, analyze data, identify possible faults, or propose an engineering solution.
Someone must still determine:
what problem actually needs to be solved;
why it needs to be solved;
what physical, economic, regulatory, and safety constraints exist;
whether the generated result is correct;
whether it can be implemented in the real system;
who is responsible when the result is wrong.
AI dramatically reduces the cost of obtaining an answer.
It does not automatically eliminate the need to understand that answer.
This suggests an important relationship:
a competent specialist + AI can produce enormous productivity gains;
an incompetent specialist + AI can produce errors at enormous speed.
The relationship is not literally multiplicative in every task, but in many complex systems AI and human competence behave as complementary inputs.
A useful approximation is:
effective AI value ≈ technological capability × human competence × organizational capability.
A weakness in any one of these factors can limit the result.
Technology requires complementary capital
This problem is not unique to AI.
Economists studying general-purpose technologies have long observed that major technologies often require extensive complementary investments before their productivity potential becomes visible.
New technology may require:
redesigned processes;
new organizational structures;
new business models;
training;
new forms of human capital;
new infrastructure;
new methods of coordination.
Brynjolfsson, Rock, and Syverson describe this phenomenon through the idea of a Productivity J-Curve: firms may make large investments in a transformative technology while measured productivity initially improves only slowly because complementary intangible capital still has to be created.
Source: Brynjolfsson, Rock & Syverson, “The Productivity J-Curve”
AI may intensify this problem because the technological component is advancing extraordinarily quickly.
A company can purchase AI systems, computing capacity, robots, industrial software, digital twins, or automated engineering tools.
But between technological capability and actual productivity stand:
qualified workers;
engineering knowledge;
procedures;
documentation;
industrial data;
management systems;
cybersecurity;
regulation;
procurement;
certification;
maintenance;
organizational redesign.
An economy can therefore possess extremely advanced technology while lacking the complementary capabilities required to extract its full value.
The problem is more accurately described not as:
excess technology + lack of competence
but as:
rapid growth of technological capital + slower growth of complementary human and organizational capital.
Explicit knowledge becomes cheap; tacit competence may become expensive
The economic value of professional knowledge may therefore change.
AI sharply reduces the cost of retrieving and generating explicit knowledge.
It can explain a formula.
It can generate PLC code.
It can summarize technical documentation.
It can produce a circuit diagram.
It can explain how a variable-frequency drive works.
But professional competence consists of more than explicit knowledge.
It also contains:
knowledge + experience + judgment + context + responsibility.
A person may understand how a variable-frequency drive works and still be unable to commission a real industrial installation.
A person may know PLC syntax and still fail to understand the physical process controlled by the PLC.
A person may generate formally correct code while completely misunderstanding the machine on which that code will run.
Much of professional expertise is tacit.
It is accumulated through:
troubleshooting;
observing real systems;
making mistakes;
identifying abnormal behavior;
correcting wrong assumptions;
comparing expected and actual outcomes;
working under uncertainty.
As explicit knowledge becomes cheaper, tacit competence may become relatively more valuable.
The scarce skill may increasingly be not the ability to retrieve an answer, but the ability to determine whether the answer makes sense.
The important shortage may not be “AI programmers”
There will certainly be demand for specialists who build AI systems.
But a more consequential bottleneck may emerge among people working at the interface between digital intelligence and physical, regulated, or safety-critical systems.
These may include:
controls engineers;
commissioning engineers;
process engineers;
power engineers;
manufacturing specialists;
electricians;
maintenance technicians;
mechanical engineers;
infrastructure specialists;
cybersecurity specialists;
industrial network specialists;
medical professionals;
specialists who simultaneously understand a physical domain, software, and AI.
The causes of shortages differ between professions.
A shortage of electricians does not arise for the same reason as a shortage of cybersecurity specialists.
Labor-market conditions also vary by country and industry.
The common feature is different:
AI output must eventually interact with the real world.
For example, AI can generate a PLC program.
But someone must understand the entire chain:
sensor
↓
electrical diagram
↓
PLC
↓
program
↓
variable-frequency drive
↓
motor
↓
mechanics
↓
technological process
↓
safetyA formally correct piece of generated code may solve very little if nobody understands the complete system.
A machine may fail to start because of:
a safety interlock;
incorrect field wiring;
a sensor failure;
a drive parameter;
a network fault;
a mechanical problem;
an incorrect process assumption.
The code may be perfect.
The system may still not work.
The same principle applies in power systems, medicine, chemical plants, transport infrastructure, construction, manufacturing, and other domains in which errors have physical consequences.
Foundational skills were under pressure before generative AI
Educational data do not show a universal decline across every country, subject, and population group.
They do, however, show significant deterioration in several foundational competencies that began before generative AI became widely available.
Germany’s PISA 2022 results were the lowest recorded for the country in mathematics, reading, and science.
Compared with 2012, the share of German 15-year-olds performing below the basic proficiency level increased by:
12 percentage points in mathematics;
11 points in reading;
11 points in science.
Source: OECD, PISA 2022 Germany
The IQB Bildungstrend shows a similar pattern.
Between 2015 and 2022, average scores among German ninth-graders fell by:
25 points in reading;
44 points in listening;
31 points in spelling.
English performance improved.
This indicates an uneven deterioration in specific competencies rather than evidence of a general decline in intelligence.
Source: IQB Bildungstrend 2022.
https://www.oecd.org/en/publications/pisa-2022-results-volume-i-and-ii-country-notes_ed6fbcc5-en/germany_1a2cf137-en.html?utm_source=chatgpt.comThe United States shows another version of the same problem.
According to the NAEP long-term assessment, average results among 13-year-olds declined over the previous decade by 7 points in reading and 14 points in mathematics.
More recent results suggest that these declines have not simply continued at the same pace, but neither have they been fully reversed.
Source: NAEP Long-Term Trend
https://www.nationsreportcard.gov/highlights/ltt/2023/?os=vbkn42___&utm_source=chatgpt.comhttps://www.kmk.org/aktuelles-1/artikelansicht/iqb-bildungstrend-2022-kompetenzrueckgaenge-in-deutsch-aber-weitere-fortschritte-in-englisch.html?utm_source=chatgpt.comThese assessments do not establish a single cause.
They do not prove that future engineers, doctors, or technicians will necessarily become less competent.
Professional expertise is created not only in schools but also through universities, vocational education, apprenticeships, mentoring, and years of practical work.
The evidence supports a narrower conclusion:
some foundational competencies were already under pressure before generative AI entered schools and workplaces.
That matters because AI is now entering an educational environment in which the formation of basic skills is already challenging.
AI may amplify the problem — or help solve it
Generative AI did not create these educational declines.
But it may affect what happens next.
To become an expert, a person traditionally had to perform a large amount of cognitive work:
read documentation;
search for information;
calculate;
make mistakes;
identify those mistakes;
write programs;
troubleshoot systems;
compare alternatives;
make decisions independently.
These activities were not merely a way of producing work.
They were also a way of producing competence.
Professional intuition emerged from repeated exposure to problems.
If AI substitutes for this practice rather than augmenting it, a destructive feedback loop may emerge:
AI performs the difficult task
↓
the person practices less
↓
the person develops less independent understanding
↓
verification of AI becomes more difficult
↓
dependence on AI increases
↓
independent practice declines further
This is one possible mechanism of:
skill degradation through the automation of cognitive practice.
But the opposite outcome is equally possible.
AI can:
explain difficult concepts repeatedly;
generate exercises;
create simulations;
provide immediate feedback;
compare alternative solutions;
simulate faults;
help students explore systems faster.
Then the cycle becomes:
AI accelerates feedback
↓
the person experiments more
↓
learning becomes faster
↓
competence grows
↓
AI becomes an increasingly effective professional tool.The critical distinction is therefore not whether AI is used.
It is how AI is used.
AI may accelerate skill formation when it supports deliberate practice.
It may cause skill atrophy when it replaces deliberate practice.
The paradox of expert formation
This may be the most important long-term issue.
AI can make existing experts dramatically more productive while simultaneously weakening some of the mechanisms through which new experts are created.
An experienced engineer who already understands a process can use AI to:
analyze documentation;
generate code;
compare designs;
search for faults;
prepare reports;
perform calculations;
explore alternative solutions.
These tools may multiply the productivity of someone who already possesses strong domain knowledge.
But junior engineers traditionally develop that domain knowledge by doing many of those same tasks themselves.
They calculate.
They search manuals.
They make design mistakes.
They write poor code.
They debug it.
They misdiagnose faults.
They correct their reasoning.
They slowly build mental models of real systems.
If these developmental tasks are increasingly delegated to AI, an important question appears:
where will the experienced engineer come from ten years later who is capable of correcting the AI itself?
Professional competence is not merely a flow of newly trained workers.
It is a stock accumulated over time.
A simplified representation might be:
H(t+1) = H(t) + training + experience − retirement − obsolescence − deskillingwhere H represents the stock of human competence.
AI may simultaneously:
increase training efficiency;
increase the productivity of experienced specialists;
reduce the amount of routine work;
reduce some forms of professional practice;
increase the risk of deskilling.
The central question is therefore not simply:
Does AI increase productivity today?
It is also:
Does the system continue to reproduce the human competence required to operate tomorrow?
A society could theoretically consume the professional competence accumulated by previous generations faster than it creates a new generation of experts.
That is a much more serious problem than a temporary shortage of workers.
Skills shortages are already constraining transformation
There is already evidence that companies consider skills shortages an important obstacle to technological change.
In the World Economic Forum’s Future of Jobs Report 2025:
63% of employers identify skills gaps as the main barrier to business transformation;
around half of surveyed executives identify insufficient skills as a major barrier to AI adoption;
77% of employers plan to reskill or upskill employees in response to AI.
Source: World Economic Forum, Future of Jobs Report 2025.
https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/4-workforce-strategies/?utm_source=chatgpt.comOECD research also suggests that firms with stronger digital skills, ICT infrastructure, and complementary capabilities are more likely to adopt and benefit from AI.
Source: OECD research on AI adoption and complementary assets
These findings do not prove that a future competence crisis is inevitable.
They establish something narrower:
the ability to absorb and deploy technology already differs significantly between organizations, and skills are already one important constraint.
The open question is whether technological capability will continue to grow faster than deployment capability.
A more realistic model of AI productivity
A simplified production function might be written as:
Y = A × F(K, L, H)where:
K — physical and technological capital;
L — labor;
H — human capital;
A — productivity.
But the productivity effect of AI does not depend only on the existence of AI itself.
It may be more useful to treat productivity as a function of:
A = A(T, H, O)
where:
T — technological capability;
H — human competence;
O — organizational capability.
AI sharply increases T.
Investment in data centers, machinery, energy systems, semiconductor capacity, robotics, and computing infrastructure increases technological and physical capital.
But if H and O grow more slowly, the economy encounters a complementary-capital bottleneck.
The technology remains valuable.
The problem is that additional technological investment may produce less realized output until the complementary system catches up.
This is especially important because human and organizational capital often take much longer to build than computing infrastructure.
A data center can be constructed in years.
A highly competent engineer may require a decade of education and professional experience.
An effective industrial organization may require years of accumulated routines, documentation, culture, and institutional knowledge.
The adjustment speeds are different.
High AI investment may therefore produce temporarily weak realized returns
Suppose a company spends $10 million on AI infrastructure.
Management expects a major productivity increase.
But the company may lack:
clean data;
trained employees;
redesigned workflows;
integration with existing systems;
engineering knowledge;
reliable evaluation procedures;
internal accountability structures.
The theoretical capability of the system may therefore be much larger than its realized economic value.
This does not necessarily mean that the technology is overvalued or ineffective.
It may mean that complementary investment is incomplete.
This distinction matters.
Weak measured returns during the early stages of adoption may represent:
technological failure;
poor implementation;
missing complementary capital;
an immature organizational structure;
or simply a delay between investment and measurable output.
Stanford reports that AI adoption is already widespread, but the financial effects reported within individual business functions remain relatively modest in many firms.
Among companies reporting cost savings, most estimate them at less than 10%.
The most commonly reported revenue increase is below 5%.
Source: Stanford AI Index 2025
https://www.aeaweb.org/articles?id=10.1257%2Fmac.20180386&utm_source=chatgpt.comThese figures are not evidence that AI will ultimately generate small productivity gains.
They are evidence that technological capability and realized business value are not the same thing.
The United States has an important advantage
If human capital becomes a major bottleneck, the United States possesses one structural advantage: its historical ability to attract highly qualified specialists from abroad.
The American response can partly be:
insufficient domestic human capital → import human capital.
Researchers, engineers, doctors, software specialists, and entrepreneurs have long moved to the United States from:
India;
China;
Europe;
Eastern Europe;
Latin America;
other regions.
That may make a competence shortage less damaging for the United States than for economies with less attractive or more closed labor markets.
But professional competence is not equally mobile.
A software researcher may relocate comparatively easily.
A commissioning engineer, electrician, doctor, infrastructure specialist, or industrial technician may face:
licensing requirements;
recognition of qualifications;
language requirements;
security restrictions;
geographic constraints;
housing costs;
local regulations;
knowledge of national technical standards.
The more a profession is embedded in physical infrastructure and local regulation, the harder it is to solve shortages simply through migration.
Economic concentration could increase — but AI also pushes in the opposite direction
If complementary competence becomes scarce, large firms may gain an advantage.
Large technology and industrial corporations can recruit highly qualified specialists globally.
Small firms often cannot. This may create a reinforcing loop:
the strongest specialists
↓
the largest companies
↓
better AI deployment
↓
higher productivity
↓
more capital
↓
greater ability to recruit scarce specialistsThis could widen productivity differences between large corporations and smaller companies.
But AI also creates an opposing force.
Small firms now gain inexpensive access to capabilities that once required large internal departments:
programming;
translation;
data analysis;
legal drafting;
marketing;
technical documentation;
research support.
AI therefore creates two opposing tendencies:
democratization of cognitive capability
and
concentration of complementary capital.
The long-term economic outcome may depend on which effect dominates.
A useful — but limited — analogy with the late 1990s
There is a useful analogy with the internet boom, but it should not be overstated.
In the late 1990s, investors correctly recognized the internet as a transformative technology.
What many misjudged was not the importance of the technology itself, but the speed at which technological potential would become sustainable profit.
Capital flowed into internet companies and telecommunications infrastructure.
Valuations rose rapidly.
The underlying technological transformation was real.
But many investment assumptions proved too optimistic.
A Federal Reserve analysis published after the boom described technological investment followed by excess capacity, reduced profitability, corrections in equity values, business failures, and retrenchment, while the underlying technology continued contributing to productivity.
Source: Federal Reserve, “What Happened to the New Economy?”
The strongest analogy with AI is therefore not:
AI today = dot-com companies in 2000.
It is:
markets can correctly identify a revolutionary technology while incorrectly estimating the speed of diffusion and the timing of economic returns.
The differences are substantial.
AI software can diffuse extremely quickly.
Many AI investments support assets with real and potentially durable economic value:
semiconductor infrastructure;
data centers;
energy systems;
cloud infrastructure;
robotics.
But diffusion may remain much slower in physical industries such as:
-manufacturing;
-energy;
-medicine;
-construction;
-infrastructure.
-These sectors require interaction with machines, regulation, safety systems, existing equipment, and skilled professionals.
A potentially unstable combination therefore exists when:
-technological investment rises rapidly;
-expectations rise even faster;
-professional skills remain scarce;
-complementary organizational change proceeds slowly;
-educational and training systems adapt more gradually.
If markets price companies according to expected future AI productivity, while the productivity effect takes much longer to materialize, valuations can temporarily outrun realized economic returns. The revolutionary technology can still be real. The timing assumptions can still be wrong.
Why the competence-bottleneck scenario may not occur.
There are strong counterarguments.
AI may dramatically accelerate professional education.
Expert systems may reduce the amount of information every worker must personally remember.
Robotics may compensate for some labor shortages.
Higher wages may attract workers into technically demanding professions.
Education systems may adapt.
Companies may redesign apprenticeships and junior professional roles around AI.
International migration may partly compensate for demographic shortages.
AI systems may become much more reliable at verifying other AI systems.
Formal verification, simulation, automated testing, digital twins, standardized interfaces, and safety systems may reduce the amount of human intervention required.
The task itself may also change.
Perhaps future engineers will not need to know everything previous generations knew.
The important skill may become the ability to supervise networks of automated tools rather than perform every intermediate task manually.
In this scenario, AI would not consume human competence.
It would amplify and reproduce it more efficiently.
The outcome is therefore not predetermined.
The critical institutional question is whether education and professional training adapt quickly enough to ensure that AI:
augments the formation of expertise
rather than:
replacing the process through which expertise is formed.
The deeper problem
AI may be creating a new imbalance between different forms of capital.
Technological capability can now increase extremely quickly.
Computing capacity can be purchased.
Software can be copied.
Models can be deployed globally.
But human competence is accumulated more slowly.
A highly capable professional must often pass through:education
↓
practice
↓
failure
↓
correction
↓
experience
↓
judgment
↓
expertiseOrganizational competence develops slowly as well.
Companies accumulate effective procedures, internal standards, documentation, culture, and institutional knowledge over years.
Technology can therefore scale faster than the system required to use it.
This is the central hypothesis:
the next major economic bottleneck may not be technological capability itself, but the complementary human and organizational capital required to convert technological capability into reliable output.
The greatest long-term risk appears if several processes occur simultaneously:
experienced specialists retire;
foundational competencies remain weak;
junior workers delegate difficult cognitive work too early;
organizations fail to redesign training;
technological investment continues accelerating.
Then society could begin consuming the stock of professional competence accumulated by previous generations faster than it reproduces it.
That would create a paradox.
Humanity would possess extraordinarily powerful machines.
It would possess enormous computing capacity.
It would possess advanced AI systems capable of producing answers at almost zero marginal cost.
But the scarce resource would become the person capable of asking:
Is this the correct problem?
Does this answer make physical sense?
What happens when the real system behaves differently?
What are the safety consequences?
What has the machine failed to understand?
The most important shortage of the twenty-first century may therefore not be artificial intelligence.
It may be:
a human being capable of understanding what needs to be done, why it needs to be done, how the real system works, and how to determine when the machine is wrong.
The deeper risk is that these processes may not merely develop at different speeds, but may increasingly move in opposite directions.
Technological capability is accelerating at a historically unusual rate, while the formation of complementary human competence may be slowing, stagnating, or deteriorating in some areas.
At the same time, a large part of today’s professional competence is inherited capital:
knowledge accumulated over decades by engineers, technicians, doctors, operators, and other experienced specialists who are now progressively leaving the workforce through retirement. Their departure does not simply reduce the number of workers.
It removes tacit knowledge, judgment, troubleshooting experience, and institutional memory that cannot be reproduced quickly through software, documentation, or formal education. This makes the adjustment process far less controllable than a conventional technology-adoption delay. In the late-1990s technology boom, one important problem was that investment and expectations moved faster than the economy’s ability to convert new technology into sustainable returns. The present situation may contain the same mechanism, but with an additional force acting in the opposite direction: technological capability may be accelerating precisely while parts of the human-capital stock required to absorb it are being depleted. If these vectors continue to diverge, the resulting mismatch could be larger, more persistent, and economically more disruptive than a simple delay in technology diffusion. The problem would no longer be that society needs time to catch up with its machines.
It would be that the machines are accelerating while part of the accumulated human competence required to understand, deploy, verify, and maintain them is simultaneously disappearing.
The most dangerous scenario is not:
“AI replaced the human.”
It is:
“Humanity created extraordinarily powerful technologies, but failed to reproduce enough of the human competence required to understand, deploy, verify, and control them.”
https://www.oecd.org/en/publications/a-portrait-of-ai-adopters-across-countries_0fb79bb9-en.html?utm_source=chatgpt.comhttps://www.oecd.org/en/publications/ai-and-skills_f843b352-en/full-report.html?utm_source=chatgpt.com I think a massive short opportunity is brewing.
