Construction Industry Capacity and Worker Shortages

  Infrastructure Australia's estimates of demand and supply 

 

In the 1930s, when economic statistics were being developed, manufacturing industries were of particular importance, because their share of the economy was two or three times larger than today and they drove the ups and downs of the business cycle. Two important factors in measuring manufacturing output were inventories and capacity utilisation. Before modern logistics and just-in-time management, increases and decreases in inventory levels indicated weakening or strengthening of demand and were important leading indicators. 

 

Capacity utilisation is the percentage of current capacity being used, calculated as the current level of output divided by the maximum possible output by 100 (actual output/potential output x 100). It will not be 100% because at any one time not all equipment will be used, so allowance is made for the setup, maintenance and downtime machinery and equipment needs. For capital intensive industries like manufacturing and utilities this is obviously an important factor. Other commonly used metrics are: production throughput (how much product can be made in a given time); equipment utilisation (how much of the available machinery is actively used); and overall equipment effectiveness (how effectively equipment is being utilised, considering availability, performance, and quality).

 

This post first looks at measuring industry capacity and the issues involved in doing this for construction. Then Infrastructure Australia’s Infrastructure Market Capacity report’s estimates of the five year project pipeline and supply side issues in materials and the workforce identified in their industry survey are given, before the construction workforce estimates of future supply and shortages are discussed. 

 

Measuring Industry Capacity 

 

Capacity is the level of output that can be produced with current resources over a set period. There is a long-run series for the US: ‘The Federal Reserve Board constructs estimates of capacity and capacity utilization for industries in manufacturing, mining, and electric and gas utilities. For a given industry, the capacity utilization rate is equal to an output index (seasonally adjusted) divided by a capacity index. The Federal Reserve Board's capacity indexes attempt to capture the concept of sustainable maximum output -- the greatest level of output a plant can maintain within the framework of a realistic work schedule, after factoring in normal downtime and assuming sufficient availability of inputs to operate the capital in place.’ There are separate indexes for each industry, and within manufacturing for iron and steel, automobiles, and semiconductors. Figure 1 has the indexes for the combined total and manufacturing, showing how utilisation rises and falls with the business cycle. 

 

Figure 1. US industry capacity utilisation

Source: FRED

 

The US series do not include service industries. Estimating total capacity and capacity utilisation for service industries such as health, professional services like accounting and legal, or personal and household services, is more difficult. In service industries the main limiting factor in output is typically taken to be the number of workers available or billable hours as a share of total hours worked. 

 

Capacity Utilisation in Construction

 

Construction is a hybrid industry where capacity is concerned. A lot of machinery and equipment is used, although how much varies across different types of work, and the industry is labour intensive compared to manufacturing, although the machinery and equipment capital stock per employee is among the highest of all Australian industries [1]. The industry requires a range of specialised skills and uses many subcontractors, so capacity is largely determined by the availability of labour, skills and expertise. This is very different to the physical limits of a manufacturing plant with its fixed and easily quantifiable physical capital. 

 

Construction is project-based and output is variable. Because output is affected by factors like weather, regulations, materials supply, and project timelines, it is difficult to define the maximum potential output for the industry compared to the continuous production processes in manufacturing. Further, the industry is fragmented, with many small firms and the subcontractors, which makes collecting consistent data on production capacity difficult. 

 

A 2015 Reserve Bank of Australia research paper on firm-level capacity reported on discussions with firms in the Bank's business liaison program that suggested the interpretation of ‘capacity utilisation’ varies considerably across different industries. The paper said: ‘construction contractors generally regarded ‘capacity utilisation’ to be of some use, primarily citing some form of labour utilisation. The focus of a ‘typical’ construction contractor in the liaison program is the time spent on each project, particularly in the detached housing market. Construction subcontractors that provide and operate capital equipment (e.g. cranes) tend to measure utilisation as the share of time that their equipment is in use.’

 

This brings us to Infrastructure Australia’s Infrastructure Market Capacity reports. The latest in November was the fifth, and looked at public infrastructure demand and market supply capacity over the five years 2024-25 to 2028-29. ‘It provides an updated health check and analysis of our national construction market’s capacity to deliver public infrastructure works’ by analysing and modelling total infrastructure demand by sector and project type, with a focus on public infrastructure, labour and skills supply and shortages, materials supply and costs, and industry productivity trends. Infrastructure Australia is an independent statutory body that provides research and advice to Australian Governments.

 

Because the Market Capacity report is primarily directed at policy makers and public sector decision-makers, it does not get the wider attention it deserves. It is a very good source of industry data, covering both demand-side and supply-side factors in a detailed and thorough analysis of the Australian construction industry that captures data on work done and the labour force from the Australian Bureau of Statistics (ABS), the project pipeline from government departments and GlobalData, a private provider, and modelling by consultants Nous on workforce and skills demand and supply. In the 2025 report, results from three surveys are included, one of 200 firms by Infrastructure Australia followed up with 20 interviews, and another of 134 members of the Civil Contractors Federation. There are six Appendices explaining the methodology and classification systems used.

 

The Project Pipeline

 

Since 2021, Infrastructure Australia (IA) has built a national project database with the location, cost, stage (pre-construction, under construction, completed), schedule, funding and type. Appendix A explains the bottom-up approach used by IA to produce their ‘portfolio’ of monthly project activity, based on expenditure at the stage of each project for all projects in a ‘project typecast’, with cost breakdowns for each project typecast over the four resource categories of plant, labour, equipment and materials. There are 83 typecasts that make up 22 ‘master sectors’ that are aggregated into the four ‘infrastructure’ sectors of buildings, transport, utilities and resources.

 

Figure 2 has the breakdown of project types. Between 2024–25 and 2028–29, IA estimates total construction work of $1.14 trillion, comparable to ABS total construction activity between 2020–21 and 2024–25 of $1.4 trillion. Buildings are 62% of expected expenditure, transport 17%, utilities 16%, and resources 5%. Public spending is 28% of the $1.14 trillion total. 

 

Figure 2. Construction pipeline by sector 2024–25 to 2028–29

Source: Infrastructure Market Capacity, p. 20.

 

The project database has six categories of projects: the Major Public Infrastructure Pipeline (MPIP, with projects over $100 million in New South Wales, Victoria, Queensland and Western Australia and over $50 million elsewhere), smaller public projects, road maintenance, mining, private, and housing. Of the $1.14 trillion total for 2024–25 to 2028–29, the MPIP is $242 billion (22%) and $66 billion (6%) is small capital projects. MPIP is up 14% from the 2023-24 report, driven by new housing, health and energy transmission projects. Within MPIP, transport at $192 billion is the largest category (53%), buildings are $77 billion (32%), and utilities are $36 billion (15%). Of the $716 billion in buildings, $116 billion is public investment, with $77 billion in the MPIP’s $242 billion and $48 billion in the $62 billion of smaller projects.

 

Using those six categories plus Defence, Figure 3 has IA’s forecast of annual construction spend from their project database, against ABS total construction activity for 2016-2025 (that includes private sector construction work not in the database). IA notes their forecast ‘uses cost estimates with limited certainty about future escalations’, with ‘forecast construction volumes peaking in 2027 at levels comparable to current ABS-reported activity.’ Peak investment is in 2027. There is $63 billion (27%) of activity outside the eight capital cities, with increasing regional demand in Queensland, Northern Territory, South Australia, and New South Wales. 

 

Figure 3. Forecast construction spend

Source: Infrastructure Market Capacity, p.19.

 

The demand-side project pipeline is for total construction, but the purpose of the report is about ‘capacity to deliver public infrastructure works’ and the MPIP. In the projected peak year of 2027, the MPIP is expected to be a bit more than $50 billion out of an industry total of around $250 billion of work. 

 

Figure 4. Major public infrastructure pipeline spend by sector 

Source: Infrastructure Market Capacity, p.21.

 

The Energy Pipeline

 

With the recent attention given to the politics of net zero, the IA’s forecast for renewable energy projects of transmission lines, solar, wind, and pumped hydro is relevant [2]. The report says: ‘As the net zero transition accelerates, the scale of energy infrastructure investment continues to grow. Irrespective of how it is funded, the pipeline for projects to build transmission, solar, wind and pumped hydro is now $163 billion for the five years from 2024-25 to 2028-29.’ This demand profile suggests workforce demand will be rising sharply from 2026.  Although most energy projects are privately funded and not counted in the MPIP, governments have $15 billion in transmission projects in the five year projection. 

 

Figure 5. Energy infrastructure pipeline

Source: Infrastructure Market Capacity, p. 28.

 

Reducing barriers to renewable energy projects by ‘accelerating approvals and smoothing supply chains’ is required. IA cites analysis from Infrastructure Partnerships Australia that 58% of the 298 energy projects included in their database ‘as having a low to moderate likelihood of being delivered to the project schedule.’ Of the organisations surveyed by IA in 2025, 47% said delays in obtaining planning and environmental approvals were among the greatest risks to project delivery. However, organisations reported disruptions to project delivery are driven primarily by cost of materials (64%), cost of labour (63%) and labour and skills shortages (59%).

 

Supply of Construction Materials 

 

IA’s Industry Confidence Survey suggests supply of key materials affects project delivery. Figure 6 shows respondents views on major or significant threats to successful delivery: 38% highlighted supply of timber and timber products, 32% steel or steel products, 30% sand or quarry products, 28% concrete or cement, 27% precast concrete, and 25% equipment and glass products.

 

Figure 6. Supply chain risk factors

Source: Infrastructure Market Capacity, p. 36.

 

There is a section that focuses on supply of fabricated steel products. IA estimates 26.6 million tonnes of structural steel will be needed over the five years 2024– 25 to 2028–29, of which the MPIP  will need 3.6 million tonnes. ‘As the estimated national steel fabrication capacity is approximately 1.4 million tonnes per annum, meeting this demand will require a combination of locally produced and imported steel.’ Imports are priced 15-50% below domestically produced products, and have been ‘rising rapidly in recent years’, but may not meet the quality and safety standards of locally made products. ‘Lack of traceability and certification makes it difficult to track material compositions, manufacturing processes and quality control procedures, which increases the risk of using substandard products.’ 

 

Workforce and Skills

 

This is the most important part of the report. In October 2025, Australia’s infrastructure workforce was 204,000 workers, with 62% trades workers and labourers, 26% engineers, architects and scientists, and 12% project management professionals. Note this is not restricted to employment of construction workers. IA estimates a shortage of 141,000 workers on public infrastructure works in October 2025. Figure 7 shows IA’s projection of demand versus supply with peak workforce demand of 521,000 in mid-2027, with an estimated shortage of 300,000 workers.

 

Figure 7. Demand, supply and shortage of infrastructure workers

Source: Infrastructure Market Capacity, p. 43. 

 

Of the estimated 300,000 worker shortage, engineers, architects and scientists will peak at 126,000 in late 2026 before gradually declining, shortages for trades workers and labourers peaks at 126,000 by mid-2027, and there will be a sustained demand for project management professionals, with a peak in mid-2027 at around 59,000. For firms IA surveyed in 2925, labour is a substantial delivery risk, with labour cost cited by 63% and labour and skills shortages by 59%. 

 

Figure 8. Worker shortage by occupational groups

Source: Infrastructure Market Capacity, p. 44. 

 

Shortages in capital cities are projected to rise from 131,700 in 2025 and peak at 148,000 in 2026. Regional locations are expected to have a much steeper increase, with the workforce shortage growing from 38,200 in October 2025 to a peak of 181,000 in 2027, because that is where the transmission, solar, wind and pumped hydro projects in the $163 billion energy infrastructure pipeline are.

 

Unpicking These Numbers

 

An explanation of how these estimates by Nous were arrived at is in Appendix E on Workforce and Skills Methodology.As the Appendix notes ‘The fundamental question addressed by this report is to what extent the current and projected supply of labour can support Australia’s proposed investment in public infrastructure.’  Nous defines the occupations and skills that underpin the workforce then estimates the ‘number of workers in or near the infrastructure workforce as determined by official statistics and our own forecasts or modelling based on those statistics’, plus ‘additional data (such as job advertisements) that provides extra information on variables (such as skills) not covered by the official statistics, and extra granularity.’ Figure 9 has their methodology.

 

Figure 9. Workforce quantification modelling methodology

Source: Appendix E: Workforce and Skills Methodology, p. 17. 

 

While the methodology looks good, there are some anomalies. Industry groups identified as directly linked to ‘construction of public infrastructure’ are 942 Equipment Repair and Maintenance and 529 Other Transport Support Services, along with the more obvious 692 Architectural, Engineering and Technical Services and the four industry groups in Construction Division E. Using job advertisement data, a  ‘share of non-project-management occupations are apportioned into project management occupations, to reflect that many project management roles on public infrastructure projects are undertaken by individuals captured under other occupations’, but occupations ‘that contained less than one per cent of project management professional roles in its job advertisements were excluded.’ ‘Weightings were developed to apportion the share of workers engaged and adjacent to public infrastructure’ using demand estimates and workforce-to-spend ratios provided by IA. 

 

The six demand side categories of major and minor public infrastructure, private infrastructure, private residential and non-residential buildings, and road maintenance are the basis of the capacity forecasts. In August 2025, there were 1.35 million people employed in the ABS Construction industry, of which 125,500 were employed in Engineering construction, which is 60% public work. There were another 330,000 people employed in Architectural, Engineering and Technical Services, and IA also adds project managers to their infrastructure workforce. The problem is the shift to the ‘public infrastructure’ component where, in 2025, IA estimated there were 204,000 workers and a 202,000 shortage. 

 

There is a conceptual problem here. How, in 2025, did the work get done if there was a shortage of half the required workforce? In 2027 the shortage will be 60% of the required workforce. Do shortages mean projects are not started because workers are not available, or lead times increase, or projects take longer to deliver because the workforce is spread over many projects? Are there more delays and bottlenecks due to such shortages? 

 

Conclusion

 

Assessing industry capacity requires identifying maximum potential output and the industry’s ability to meet current and future demand, based on supply of factors like the workforce and materials, and the effects of technology and market conditions. Other factors that are often considered are the number of active firms, trends in output and productivity, and capacity utilisation (i.e. the level of output compared to potential maximum output, often used to allow for downtime needed for maintenance of machinery and equipment). 

 

Infrastructure Australia (IA) produces an annual Infrastructure Market Capacity Report that provides forecasts of Australian infrastructure demand and supply of resources, and  makes recommendations to improve the capacity of the construction industry to meet forecast infrastructure demand across four areas: to actively manage demand; to increase material supply; to increase labour supply; and to improve construction productivity. 

 

Since the first report in 2021, Federal and state governments have adjusted the project pipeline, and the reports show a stretching out of work over more years and reductions of IA’s forecast peak in public infrastructure investment. IA notes this is ‘likely reflective of planned expenditure being pushed back as the market struggles to meet overly ambitious delivery targets.’ However, whether those decisions were based on the report or the result of delays in project funding, preparation or commencement is hard to tell. Compared to last year’s report, IA’s projected activity peak has fallen and has been shifted out a year, illustrating the uncertainty associated with long-term forecasts.

 

The report has a focus on projects in the Major Public Infrastructure Pipeline (MPIP). IA collects data on planned projects and labour and material supply, supported by interviews and surveys with key stakeholders, including state and territory governments and the Australian Department of Infrastructure, Transport, Regional Development, Communications and the Arts. Project data is classified by type, sector, phase and funding source. The report has evolved over five editions, to include current and emerging market conditions, availability of labour and construction materials, productivity trends, the supply of apprentices and trainees and other workforce issues. It includes regional demand, and the energy projects located in the regions have become an increasingly important component of total demand.

 

The research is not without challenges. In particular, the cost estimates for projects in the pipeline that are the basis for demand forecasts unavoidably have some uncertainty about future cost escalations. IA’s National Infrastructure Project Database aggregates data from public and private sources, including the ABS, but does not and is not intended to capture all private sector activity. How serious these issues are is a matter for debate. 

 

Industry capacity is the overall ability to deliver output, typically at a national and sectoral level. For IA, this is the capacity of the construction industry to deliver major infrastructure projects. There are two issues here. The first is the lack of a clear statement of what the maximum output possible with existing resources of the Australian industry actually is, whether for total work or for the three industry sectors of residential and non-residential building and engineering construction. These estimates should be clearly made.

 

The second is linking worker shortages to the numbers for the value of total public infrastructure and the major project pipeline, which is developed from the project database. There is no separate section in the report on how these estimates are derived. Although the methodology is in the Appendices, how the MPIP is used to estimate worker shortages is not explained. Given that managing the public infrastructure pipeline is the main point of producing the capacity forecasts, the analysis should be highlighted and be made more explicit. 

 

A serious problem is that there is no analysis in the report of the effect of increased demand on project duration. One of the characteristics of construction is that project delivery times increase during periods of high demand, because this the most important way the industry adjusts to increasing demand. As the available labour and materials are fixed in the short-run and capacity is limited by availability, these get spread over more work as new projects are started. Contractors bid for projects to ensure they have sufficient work in the future, adding to their workload, and the result is fewer people on a site and slower progress of ongoing work. Lead times and cycle times increase with workload, as project duration from order to delivery and from start of construction to completion increase. At high levels of activity, there is more potential for delays or bottlenecks in supply chains, as the reporting on industry survey results and discussion shows. 

 

Another problem is that, by assessing capacity against total demand, there is an underlying assumption that workers move between engineering and building work. However, there is no good evidence that there is much worker mobility between sectors. BuildSkills Australia 2025 Housing Workforce Capacity Study published in September 2025 found limited evidence that infrastructure activity is materially drawing labour away from residential construction [3]. How the shortages of architects, engineers and project managers are estimated is also not clear. 

 

A final point is that infrastructure is conventionally divided into physical (roads, rail, ports, energy etc.) and social (schools, hospitals, community centres etc.) projects. These divisions are not used by IA, instead they have four sectors of transport, utilities, buildings, and resources, and the portfolio breaks down ‘these three sectors’ [sic] across 22 Master Types and 83 separate typecasts (detailed in Appendix C). No reason is given for not using these categories and putting everything into the MPIP. 

 

The report could do with an edit and has a couple of basic errors, where different numbers are given for the same thing: transport is $129 billion on p.18 and $192 billion on p. 20, and the public infrastructure worker shortage in 2025 is 141,00 on p. 42 and 202,000 on p.43. However, the report has a detailed, a five-year forecast that covers sectors like transport, energy, water, and buildings, including housing. It analyses critical issues such as labour and skill shortages, material costs and equipment demand, productivity, and the challenges of the energy transition, including regional breakdowns. Insights from industry surveys and interviews add perspective on market conditions, risks, and challenges. It is an important source of data and provides a comprehensive overview of Australian construction.

 

                                                                          *

 

[1] This post looked at the capital stock of Australian industry.

 

[2] Opposition to net zero is bizarre. A 21^st century economy runs on abundant electricity, and today the cheapest source of electricity is solar with storage. The Ember 2025 Global Electricity Review found: ‘Renewable power sources added a record 858 TWh of generation in 2024 … brought low-carbon power to 40.9% (12,609 TWh) of the mix in 2024, compared with 39.4% in 2023. Hydro remained the largest source of low-carbon electricity (14.3%), followed by nuclear (9.0%), with wind (8.1%) and solar (6.9%). Solar generation has doubled over the last three years to reach over 2000 TWh. Solar was the largest source of new electricity generation globally for the third year in a row (+474 TWh) and the fastest growing source of electricity (+29%) for the 20th year in a row. More than half (53%) of the increase in solar generation in 2024 was in China.’

 

[3] A Residential Mobility Study is currently being done by BuildSkills Australia, supported by Jobs and Skills Australia and the ABS, to understand actual and potential labour flows between residential and other construction. 

 

 

 Subscribe on Substack

https://gerarddevalence.substack.com/

Recent Research on AI Effects on Employment and Work

 AI scenarios range from apocalypse to slow and gradual adoption

 

There has been and currently is a lot of discussion on the effects of artificial intelligence (AI) on employment. Across that research there is a wide range of views on the timing and extent of those effects, from a looming jobs apocalypse with high unemployment to slow and gradual adoption with low unemployment. The picture is clouded by tech firms self-serving promotion of AI solutions and the attention given to the possibility of an out-of-control unaligned AI killing everyone, or turning them into paperclips. 

 

A few examples will suffice. According to Goldman Sachs Research on How Will AI Affect the Global Workforce? in August 2025,  AI is unlikely to lead to a large increase in unemployment ‘because technological change tends to boost demand for workers in new occupations’, creates new jobs, and increases output and demand. They estimate unemployment will increase by half a percentage point during the AI transition period as displaced workers seek new positions and, if current AI use cases were expanded across the economy, 2.5% of US employment would be at risk of related job loss. So far, they believe the ‘vast majority of companies have not incorporated AI into regular workflows and the low adoption rate is limiting labour market effects.’ 

 

Massachusetts Institute of Technology professor Daron Acemoglu, who has written extensively on technology and work, believes only 5% of all jobs will be taken over, or at least heavily aided, by AI over the next decade. However, the World Economic Forum’s Future of Jobs Report 2025 estimates that, by 2030, new job creation and job displacement will be a ‘combined total of 22% of today’s total (formal) jobs.’ Their jobs outlook is based on the macrotrends of technology, economic uncertainty, demographics, and the energy transition, of which ‘AI and information processing technologies are expected to have the biggest impact – with 86% of respondents expecting these technologies to transform their business by 2030.’

 

On the threat of extinction, in April 2025 the AI Futures Project, a credible non-profit research group, released their AI 2027 scenario, when AI systems ‘become good enough to dramatically accelerate their research’ and start building their own superintelligent AI systems. Without human understanding of what’s happening, the system develops misaligned goals: ‘Previous AIs would lie to humans, but they weren’t systematically plotting to gain power over the humans.’ The superintelligent AI will manipulate humans and rapidly industrialise by manufacturing robots: ‘Once a sufficient number of robots have been built, the AI releases a bioweapon, killing all humans. Then, it continues the industrialization, and launches Von Neumann probes to colonize space.’ 

 

In September the US think tank RAND published a research paper on the potential for the proliferation of robotic embodiment of superintelligent AI called Averting a Robot Catastrophe, arguing for ‘the urgent need to proactively address this issue now rather than waiting until the technologies are fully deployed to ensure responsible governance and risk management.’ Another 2025 RAND paper on The Extinction Risk From AIconcluded ‘Although we could not show in any of our scenarios that AI could definitely create an extinction threat to humanity, we could not rule out the possibility… resources dedicated to extinction risk mitigation are most useful if they also contribute to mitigating global catastrophic risks and improving AI safety in general.’

 

While AI is developing rapidly, and there are examples of AI deception from Anthropic and OpenAI, a cautionary tale is US-based Builder.ai. The company claimed its product, an AI bot called Natasha, could help customers build software six times faster and 70% cheaper than humans. In 2023 it was ranked third by tech industry magazine Fast Company behind OpenAI and Google’s DeepMind in its innovative AI companies list, and was valued at $US1.5 billion. Builder.ai collapsed in May: ‘Alongside old-fashioned start-up dishonesty with dramatically overstating its revenue, allegations arose that the work of its Natasha neural network was actually the work of 700 human programmers in India.’ This is reminiscent of Elon Musk’s Optimus robots being remote controlled in a 2024 demonstration. 

 

Although it is still too early to say what the effect of AI on employment will be, there has been some useful recent research on the effect of AI on jobs and work, particularly in the US. This post surveys some of the research released over the last few months. 

 

Australian Research

 

The Productivity Commission’s August 2025 Harnessing Data and Digital Technology report said: ‘The economic potential of AI is clear, and we are still in the early stages of its development and adoption… multifactor productivity gains above 2.3% are likely over the next decade, though there is considerable uncertainty. This would translate into about 4.3% labour productivity  growth over the same period.’ The Commission argued data underpins growth and value in the digital economy. And a ‘mature data-sharing regime could add up to $10 billion to Australia’s annual economic output. Experience shows that we need a flexible approach to facilitating data access across the economy.’ In another report for the Economic Reform Roundtable on skills and employment, the Commission recommended improving education and training systems.

 

Grattan institute researchers Trent Wiltshire and Hui-Ling Chan’s September 2025 article AI is Coming: Prepare for the Worst argues ‘in the event of significant disruption, the federal government may need to consider how Australia’s safety net and retraining systems’, with better preparation and scenario planning for Australia for the possibility AI will cause mass unemployment. They suggest changes to income support should be considered, such as lifetime learning accounts, unemployment insurance (a time-limited payment linked to a person’s previous income widely used in Europe), easier access to superannuation when unemployed. They also recommend Denmark’s ‘flexicurity’ system where it is easy to retrench workers but there is a safety net that includes up to two years unemployment insurance, and education, retraining, and support programs. About 25% of Denmark’s private industry workers change jobs each year, and 67% of workers are union members. 

 

A June 2025 PwC AI jobs barometer ‘looked at close to a billion job adverts from 24 countries and 80 sectors to understand how the demand for workers is shifting in relation to AI adoption. The global study found that AI is making workers more valuable, not less. Industries most able to use AI have seen productivity growth nearly quadruple since 2022 and are seeing three times higher growth in revenue generated per employee. Jobs numbers and wages are also growing in virtually every AI-exposed occupation, with AI-skilled workers commanding a 56% wage premium, on average.’

 

The PwC survey found the Australian industry effect of AI was a surge in demand for AI skills in the overall jobs market, nearly doubling from 12,000 postings in 2020 to 23,000 in 2021. Since then there have been 23,000 postings a year, although this was only 1.8% of total job postings in 2024. As Figure 1 below from the report shows, Finance and Insurance was the leading industry, but there has been rapid growth in Construction industry AI job postings.

 

Figure 1. Job postings

 

Source: PwC

 

 

RBA Survey of Australian Businesses

 

The Reserve Bank Governor Michele Bullock gave a speech on September 3^rd which included results of an RBA survey of businesses about AI, robotics and technology adoption. Although not about employment, the speech had the Figure below from the RBA survey of businesses about technology adoption, with the striking finding that 80% of firms expect to be using AI in the next three years, up from 25% today. This is probably due to the RBA survey population being skewed toward larger firms. 

 

Figure 2. Australian businesses’ technology adoption

Source: RBA

 

In her speech she said: ‘ Firms mainly expect these tools to augment labour, automating repetitive tasks and redesigning the composition of roles. Firms thought they may initially see an increase in their headcount as they design and embed new technologies, though this may be followed by a small decline as they mature in their adoption of new technologies. Lower skilled roles may decline, while demand for higher skilled roles is expected to grow, continuing (and perhaps even fast-tracking) a decades-long trend away from routine manual work. While AI may eventually automate even some higher skilled tasks, firms tell us that it is too early to fully understand what this means for their workforce beyond the next few years. Some roles may change and the demand for different or new skills may in turn increase.’

 

US Research

 

An August 2025 paper by Eckhardt and Goldschlag called AI and Jobs: The Final Word (Until the Next One),found no detectable effect of AI on recent US employment trends using five measures of job exposure to AI. For three of their five measures there was no detectable difference in unemployment between the more exposed and the less exposed workers, and only a small difference, of 0.2 or 0.3 of a percentage point for two measures. They say ‘One pattern is clear in the data: highly exposed workers are doing better in the labor market than less exposed workers. Workers more exposed to AI are better paid, more likely to have Bachelor’s or graduate degrees, and less likely to be unemployed than less exposed workers.’

 

Their conclusion was that AI isn’t taking jobs yet, or the effect is very small. Figure 3 has their unemployment rate of workers with varying degrees of predicted AI exposure (1 is the least exposed, 5 is the most exposed), where there is no correlation between t AI exposure and unemployment.

 

Figure 3. Unemployment rate by AI exposure quintile

 

Source: Eckhardt and Goldschlag 2025.

 

In their Appendix they used US Census Bureau data, which in August had 9% of surveyed firms using AI, up from 5% a year and a half earlier, although 27% of firms in the information sector said they were using AI. The Appendix had the Figure below, showing Construction having one of the lowest levels of AI usage.

 

Figure 4. Percent of businesses using AI

 

Source: Eckhardt and Goldschlag 2025.

 

 

The next word on AI and Jobs came in a paper from the Stanford Digital Economy Lab by Brynjolfsson, Chandar, and Chen Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. This paper uses two measures of AI exposure, then compares recent employment trends for more and less exposed workers. Their conclusion is radically different to Eckhardt and Goldschlag. The abstract explains:

 

‘We find that since the widespread adoption of generative AI, early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 13 percent relative decline in employment even after controlling for firm-level shocks. In contrast, employment for workers in less exposed fields and more experienced workers in the same occupations has remained stable or continued to grow. We also find that adjustments occur primarily through employment rather than compensation. Furthermore, employment declines are concentrated in occupations where AI is more likely to automate, rather than augment, human labor. Our results are robust to alternative explanations, such as excluding technology-related firms and excluding occupations amenable to remote work. These six facts provide early, large-scale evidence consistent with the hypothesis that the AI revolution is beginning to have a significant and disproportionate impact on entry-level workers in the American labor market.’

 

Brynjolfsson et al. found ‘substantial declines in employment for early-career workers (ages 22-25) in occupations most exposed to AI’, such as software developers and customer service representatives. In jobs less exposed to AI, employment growth for young workers was comparable to older workers. Declining employment in AI-exposed jobs is driving ‘tepid overall employment growth for 22- to 25- year-olds as employment for older workers continues to grow.’ It should be noted that Eckhardt and Goldschlag use of three other metrics gives a broader perspective. Figure 5 shows growth in employment between October 2022 and July 2025 by age and GPT-4 based AI exposure where quintile 1 is least exposed and 5 the most exposed [1]. 

 

Figure 5. AI exposure group growth in employment.

 

Source: Brynjolfsson et al. 2025

 

 

Is AI a Complement or Substitute?

 

The different conclusions from this US research by throws into sharp relief what is, at this point, the core issue: is AI a substitute for workers, particularly skilled workers, or a complement? In other words, is AI replacing workers in some occupations, or is it being used as a tool to enhance productivity and pefomance? 

 

If AI is a substitute for workers, wages and employment fall, and because AI is substituting for human cognition, businesses will replace expensive humans with a skill or experience premium that is no longer valuable, probably older workers. On the other hand, if AI is a complement, wages and employment increase, and businesses will recruit humans with skill or experience, probably older workers.

 

What Figure 5 above shows is that young workers saw reductions in AI-exposed jobs, but for the other age groups it was positive. In particular, employment for older workers in AI-exposed jobs increased. Eckhardt and Goldschlag also found workers exposed to AI are doing better in the labour market than less exposed workers. This strongly suggests AI is a complement not a substitute. AI complements human skills and augments productivity of workers with the necessary skills and experience, so these people with tacit knowledge not available to an AI are the ones firms are employing. 

 

What About Construction?

 

The US Bureau of Labor Statistics (BLS) 2025 Occupational Outlook Handbook covers 600 occupations [2]. Based on that, the Employment Projections program develops US labour market estimates for 10 years in the future, based on the assumption that labour productivity and technological progress will be in line with historical experience which shows ‘technology impacts occupations, but that these changes tend to be gradual, not sudden. Occupations involve complex combinations of tasks, and even when technology advances rapidly, it can take time for employers and workers to figure out how to incorporate new technology.’ Because technological developments over the next 10 years are ‘impossible to predict with precision,’ new projections are released annually. Figure 6 has Construction employment growing at 4.4% to 2034.

 

Figure 6. US labour market 

 

Source: BLS Employment Projections, August 2025. 

 

In the BLS sector specific projections for the Infrastructure sector by 2030: ‘new job roles are expected to be created for Big Data Specialists and Organizational Development Specialists... Twenty-seven percent of employees in the sector are anticipated to be able to upskill in their current roles, with an additional 17% projected to be reskilled and redeployed. Almost 70% of respondents expect reskilling and upskilling to improve talent retention and enhance competitiveness and productivity of their company, with 50% planning to increase talent mobility through training programmes.’

 

An article in the February 2025 BLS Monthly Labor Review on Incorporating AI impacts in BLS employment projections: occupational case studies argued ‘GenAI can support many tasks involved in architecture and engineering occupations, potentially increasing worker productivity.‘ Their technical expertise and existing regulatory requirements create uncertainty about the extent and employment impact of AI adoption, and underlying demand is expected to remain strong, resulting in US employment growth of 6.8% for architects and engineers.

 

AI Development and Diffusion

 

An April 2025 paper by Arvind Narayanan and Sayash Kapoor from Princeton University’s Center for Information Technology Policy was called AI as Normal technology: An alternative to the vision of AI as a potential superintelligence. They view ‘AI as a tool that we can and should remain in control of,’ and argue this does not require drastic policy interventions. They do not think viewing AI as a humanlike intelligence is ‘currently accurate or useful for understanding its societal impacts.’ 

 

Their lengthy and sometimes digressive paper is based on the idea of a normal technology, where sudden economic impacts are implausible because ‘Innovation and diffusion happen in a feedback loop… With past general-purpose technologies such as electricity, computers, and the internet, the respective feedback loops unfolded over several decades, and we should expect the same to happen with AI.’ They dismiss catastrophic AI because it ‘relies on dubious assumptions about how technology is deployed in the real world. Long before a system would be granted access to consequential decisions, it would need to demonstrate reliable performance in less critical contexts.’

 

Narayanan and Kapoor’s ‘AI as normal technology is a worldview that stands in contrast to the worldview of AI as impending superintelligence.’ They don’t believe progress in generative AI is as fast as claimed, nor that AI diffusion will be much different to electricity or computers, because diffusion ‘occurs over decades, not years.’ This is very different to what they call the utopian and dystopian worldviews of AI, both based on the idea of superintelligence but with opposite consequences. Because the idea of immanent take-off superintelligence is so prevalent in the discussion about AI, as either the solution to many problems or as an extinction event, the suggestion that AI might just be the latest in a long series of powerful general purpose technologies and develop over time in a historically familiar way is both radical and unusual.

 

There is support for this slow adoption and diffusion view from the McKinsey 2025 State of AI report, which is somewhat ironic as McKinsey is one of the biggest boosters of corporate use of AI. Published in March 2025 but based on a mid-2024 survey sample of 1,491, it  found 75% of respondents using AI in at least one business function but only 1% ‘described their gen AI rollouts as mature.’ The survey showed a quarter of large organisations and 12% of smaller ones had an AI roadmap, 52% of large organisations but only 23% of small ones had a dedicated team to drive AI use, and only 28% and 23% respectively had effectively embedded gen AI into business processes. In McKinsey’s sample, 92% of companies plan to increase their AI investment over the next three years. However, that sample will not in any way be representative of most businesses. 

 

Figure 7. AI deployment

 

Source: McKinsey 2025 State of AI report, 42% of respondents  work for organizations with annual revenue over $500 million.

 

Chat GPT was launched in November 2022 by OpenAI. When GPT-4 was released in March 2023, AI went from being unreliable and error prone to being able to synthesise, summarise and interpret data. In August 2005 GPT-5 was released, which again improved performance but not by as much as the previous upgrades, so progress in AI models might be slowing down. The latest models still require supervision and checking of results. 

 

Conclusion

 

There are very many possible futures that could unfold over the next few decades as technologies like AI, automation and robotics develop. However, the key technology is intelligent machines operating in a connected but parallel digital world with varying degrees of autonomy. AI agents will be trained to use data in specific but limited ways, interacting with each other and working with humans. The tools, techniques and data sets needed for machine learning are becoming more accessible for experiment and model building, and as well as the cloud-based large language models like Gemini and ChatGPT, new AI systems like small language models and agentic AI are now appearing. 

 

So far, in many cases these technologies are not a substitute human labour. Generative design software does not replace architects or engineers, automated plan reading does not replace estimators, and optimization of logistics or maintenance by AI does not replace mechanics. Nevertheless, there is an immediate and important need for politicians and policy-makers to increase the urgency and attention given to the effects of AI on employment. Governments have to integrate AI literacy into school curriculums, provide learning subsidies for retraining, and ensure access to technology. 

 

The BLS employment projections show employment declines concentrated in occupations where AI is more likely to automate rather than augment human labour. The industries most affected are mining, retail, manufacturing and employment by government. For construction, between 2024 and 2035 in the US, the projection is for an increase of 4.4%, and for architects and engineers and increase of 6.8% in employment. How representative that is for other counties is impossible to know, but AI use in the US is probably more advanced than in most places.

 

Current employment data from the US shows that employment is steady or increasing for older workers with skills and experience, even in jobs that have high exposure to AI, although for younger workers with less experience there has been an increase in unemployment. At present, AI is affecting entry-level jobs but there are few wider employment effects, and the limited evidence suggests AI complements human skills and augments the productivity of workers with tacit knowledge not available to an AI. 

 

The picture is mixed. Surveys of companies, like the ones from the World Economic Forum, the RBA and McKinsey, report strong interest in AI and a high level of investment planned for the next few years. The share of job postings requiring AI skills is small but increasing. At the same time, employment in AI exposed jobs in the US is rising, not falling, with little or no difference in current unemployment levels between more exposed and less exposed workers. However, research shows  unemployment among 20 to 30 year old tech workers has risen. 

 

There are some other signs of AI effects in the US, with BLS data showing employment growth in marketing, graphic design, office administration, and telephone call centres in 2025 below trend, with reduced demand for workers attributed to AI-related efficiency gains. In Australia there are similar reports, like the use of chatbots by Origin Energy, insurer Suncorp, and banks cutting jobs (announced this week were 3,500 by ANZ and 400 by NAB). 

 

None of this data is conclusive. Survey results are primarily from large firms, micro and small size firms are missing, and surveys do not accurately capture most of the medium size ones. Employment and unemployment data is a lagging indicator that is variable and often revised over the following months, does not include many casual workers, and misses all informal workers completely. Many companies will retrain or relocate workers displaced by AI. The online jobs databases researchers use to estimate AI employment effects are a subset of the overall labour market, and they can only be partially representative of current conditions at best.   

 

On present trends and performance, more extreme AI scenarios are not plausible, such as AI superintelligence delivering annual economic growth of 20%, as a breakthrough problem solving, research and innovation bonanza, as a jobs apocalypse, or as an extinction event. Whether that means AI is a ‘normal’ general purpose technology that will take a few decades to become widely used across industries and the economy is not obvious. According to OpenAI, in mid-2025, ChatGPT had about 800 million weekly active users and 122-130 million daily active users, and 10 million paying users, including 92% of US Fortune 500 companies (N.B. these numbers are from a query on the OpenAI Research website). 

 

Another indicator is downloads of AI models. ChatGPT is averaging 45 million a month, and according to Wikipedia by ‘January 2023, ChatGPT had become the fastest-growing consumer software application in history, gaining over 100 million users in two months. As of May 2025, ChatGPT's website is among the 5 most visited websites.’ That level of uptake is a lot faster than the decades taken for previous technologies like electricity, the internal combustion engine or the internet to become widely used. This reinforces survey findings that many people use AI, including at work, but AI adoption by companies remains low, especially for small and medium size ones, and the great majority of companies have not incorporated AI into their workflows.

 

There are two key points that emerge from, what is at present, an unclear picture of the next decade. The first is that AI automates tasks not jobs, so jobs with structured workflows doing routine and repetitive tasks will be quickly and heavily affected. Examples are administration and data compilation, document processing, customer support, data management, note taking and drafting reports. Employers will do this because it is cost effective and relatively straightforward to train an AI agent for a specific task if the data is available. 

 

The second is the value of tacit knowledge and experience. One example is trade skills and tasks, where some can be automated but some will not, because of the  physical demands of the work. Construction trades will be one of the occupations least affected by AI. Another more pertinent example is in health, where AI assisted diagnostics require oversight by a knowledgeable human. For skilled workers like architects and engineers, using AI requires a high level of knowledge gained through learning by doing in the person responsible for supervision and checking the AI output, and AI could increase demand for these workers. The assumption is that some insight into how the AI works is required.

 

Ethan Mollick’s 2024 book Co-Intelligence outlined how humans can work with an AI chatbot as a co-worker, correcting its errors, checking its work, co-developing ideas, and guiding it in the right direction. This is a widely held view of the way AI will be used. However, in September 2025 he  wrote: ‘I have come to believe that co-intelligence is still important but that the nature of AI is starting to point in a different direction. We're moving from partners to audience, from collaboration to conjuring.’ Mollick suggests the newest most powerful AI models like GPT-5 Pro have ‘impressive output, opaque process’ and ‘for an increasing range of complex tasks, you get an amazing and sophisticated output in response to a vague request, but you have no part in the process….Magic gets done.’ 

 

In the twentieth century, the electrification of workplaces took several decades, well into the 1930s, as organisations restructured around the new technology, relocating and redesigning factories, creating new jobs and developing new products. Now, a hundred years later, AI is having the same effects, but it will not take decades for the restructuring of organisations and the jobs they provide. While the future is uncertain, within a decade AI will probably have become as ubiquitous as electricity and the internet, something we use all the time without thinking about where it comes from or how it works. 

 

                                                            *

 

[1] Eckhardt and Goldschlag, and Brynjolfsson  et al., use a metric based on queries to the Occupational Information Network (O*NET), an online database with hundreds of job definitions, using ChatGPT, developed by Felton, E., M. Raj and R. Seamans in Occupational, industry, and geographic exposure to artificial intelligence, Strategic Management Journal. It was also used by Eloundou, T., S. Manning, P. Mishkin, and D. Rock. 2023. GPTs are GPTs: Labor market impact potential of LLMs, arXiv. 

 

[2] The BLS 2025 Occupational Outlook Handbook includes information on about 600 detailed occupations in over 300 occupational profiles, covering about 4 out of 5 jobs in the US economy. Each profile features 2024–34 projections, along with assessments of the job outlook, work activities, wages, education and training requirements.

Subscribe on Substack here

https://gerarddevalence.substack.com/