Opinion
The US, UK and the EU have begun adjusting their labour market projections for artificial intelligence (AI), that is forecasts that try to take account of future impacts of AI. Australia is yet to do this based on current JSA disclaimers and DEWR’s ‘AI and Employment in Australia’ report that monitors the past rather than forecasting the future. Whilst difficult to implement economy-wide, mega-projects like the nuclear submarine build at Osborne crystallises why it’s practically needed.
Every occupation shortage list published anywhere in the world today seeks to answer the same implicit question: how many extra humans does each occupation and labour market need? None yet answer the question that increasingly impacts that number: how much of any projected shortfall will be absorbed not by humans at all, but by software agents, robotic cells and AI-controlled systems performing the work directly?
In other words, setting a novel supply-side framing by treating AI capacity as effective labour supply against projected vacancies as well as modelling it as a demand side jobs dampener. Call it the wicked “net-of-AI” question. A few counties and economies are working their way towards a solution, but Australia is not an early mover.
How other nations are progressing AI forecasting
The most consequential move has come from the US. The Bureau of Labor Statistics (BLS) now builds AI‑related productivity assumptions directly into its baseline occupational employment projections. In the 2024–34 cycle, adoption of generative AI and its associated productivity gains explicitly dampen projected labour demand in specific fields. Sales, design and administrative support, interpreters and translators, technical writers and special effects artists are marked down on the ground that automation/AI generative systems increasingly replicate their core tasks.
The BLS has a conservative methodology with adjustments made only after concrete, market-driven evidence of technology adoption before lowering demand projections, refreshing these annually as the technology and the evidence changes. One of the world's most-watched employment projection sources is quietly becoming a net-of-AI number, occupation by occupation.
The UK has taken a different but equally deliberate route. The Department for Education's AI Skills for Life and Work projections, (Jan 2026), formally overlay AI-adoption scenarios onto the Working Futures occupation-by-sector matrices, using five, seven and nine-year technology diffusion lags. These scenarios project accelerated decline in administrative, secretarial as well as skilled trades employment, the latter arising from greater automation.
Crucially for skills policy planning, replacement demand outpaces the net structural job losses across every single major occupational group through to 2035. Even where AI erodes an occupation, factors like retirements, career changes and churn keep the training requirement positive. That is the kind of finding a gross-of-AI methodology can’t answer, because it doesn’t ask the key question.
At the EU level, Cedefop's skills forecast and transition to 2035 now carries a published digital transition scenario combining sector-level automation risk with productivity assumptions for generative AI, under which rapid deployment scenarios reduces EU employment by around 5 per cent – up to 7.5 million workers – with roughly three-quarters of automation-related displacements materialising by 2035. It is an overlay rather than an adjusted baseline, but the assumptions are explicit, quantified and public. A national training authority anywhere in Europe can now read its own occupational mix against a stated AI trajectory.
Australia's approach is a neutral disclaimer
JSA’s employment projections to 2035 – using the Victoria University model outputs – carry an explicit caveat: “These projections are based on existing trends that are being observed in the economy. They do not currently reflect the labour market implications of the adoption of generative AI and other emerging technologies”.
It refers instead to JSA's Gen AI Capacity Study (2025). This scored all ANZSCO occupations for automation and augmentation potential, finding ~3–5 percent (depending on classification) of occupations are highly exposed to automation, and that persistent shortages e.g. trades, construction, care, engineering sit at the low-end of automation.
The Occupation Shortage List, is methodologically backward-looking, constructed from fill rates, vacancy data and employer surveys. AI substitution enters only to the extent that it has already suppressed observed vacancies. Nothing in the analysis looks forward at AI technology impact.
The new report AI and Employment in Australia by express design is also retrospective. It “does not make predictions about potential future effects of AI”. Between Nov.2022 and Feb.2026, employment in the most-exposed fifth of occupations grew by 5.6 percent, compared with 9.5 percent in the least-exposed fifth, a suggestive measure of past AI impact in the labour market. So, Australia does not use such rich data prospectively, unlike approaches the US, UK and EU are now attempting.
Why the approach needs rethinking
Readers studying JSA’s skills gaps and employment forecasts into the next decade need to absorb this caveat and know what they are looking at has no in-built projections of future AI impacts. That’s a problem: projections carry policy weight being widely relied upon e.g. skilled migration occupation lists, JSA-informed VET funding priorities, university places in “national priority” fields, etc. It is also illogical as AI fast becomes the most significant general-purpose technology since electrification.
Demand dampener or supply supplement
Existing adjustments, be it BLS, UK, Cedefop, treats AI as a demand shifter: fewer workers needed per unit of output. None yet directly treats AI as effective labour supply: agents and autonomous systems counted as capacity that partially fills a projected vacancy pool. Yet this is what employers will be asking. Whether a “shortage” of bookkeepers, paralegals or schedulers is a shortage at all, if agentic systems can absorb the marginal vacancy at prevailing wages.
A skills shortage methodology that does not include ‘net-of-AI’ also neatly avoids inflaming even larger emotional and political issues: ‘AI is taking our jobs’. But that’s what must be grappled with. The UK's replacement-demand finding shows this fear is overdone – netting does not collapse the training task. It redirects it, something which a forward training system needs to know about a decade in advance.
Submerged in the too hard basket
Yes, integrating any future forecast of AI impact is neither easy nor reliable. So, prudent status quo says: “We can't net out AI because nobody knows exactly where and how fast it will progress. Our published numbers will be wrong, so it's more responsible to publish nothing."
But that’s illogical, as it then sets an assumption that AI's effect is zero, which is the one value everyone agrees is wrong. This posture now spans two agencies: JSA disclaims AI in its projections, while DEWR expressly declines to predict. It’s reassuring headline of ‘no broad upheaval to date’ risks being read as licence to keep it that way.
So, the real question isn't ‘can we forecast AI accurately?’ It's ‘given we'll be wrong either way, which direction of being wrong is more bearable?’, best illustrated by a mega-project example.
SSN-AUKUS build: Net-of-AI is rational and tractable
The SSN-AUKUS build at Osborne is a mega project of national significance. Over 2026–2036 it combines up to 4,000 workers constructing the Submarine Construction Yard, a production and program workforce building toward a projected peak of 4,000–5,500 direct shipbuilding jobs, early component manufacture for the UK and US production lines, and a Skills and Training Academy first intake in 2028, all within a program frame of some 20,000 jobs (direct and indirect) over thirty years.
The chosen 2026-2036 decade stops short of the boats themselves: on current public scheduling, construction of the first SSN-AUKUS is expected to begin in the early 2030s for delivery in the early 2040s, and production peak lies beyond it. The decade to 2036 is when the workforce is recruited, trained, cleared and certified; which makes it exactly the decade over which net-of-AI assumptions must be made, because the training pipelines are longer than any credible AI forecast horizon.
“Net-of-AI” – cohort by cohort
The SSN-AUKUS project contains, within one postcode, both ends of JSA's exposure distribution: nuclear-grade trades that score among the least automatable occupations in the economy, and large volumes of complex program-office work like scheduling, document control, procurement, quality documentation that sit among the most exposed.
The civil construction workforce nets out almost nothing this decade; that shortage stands gross and must be achieved with migration, interstate attraction and wages.
The shipbuilding trades net out perhaps 10-20 per cent through robotic welding cells and AI-assisted non-destructive testing – material but bounded – because nuclear pressure-boundary work remains human-certified and regulator-governed.
Engineering likely nets out little in headcount terms even at 20-30 per cent task absorption, because first-of-class programs consume productivity gains as rework and design-changes have probable impact, mitigating against the release of people.
Nuclear stewardship – safety case, radiation protection, regulatory interface – nets out essentially nothing, because the accountability is legally and culturally human.
Maximum net-of-AI impact is concentrated in project controls, procurement and program administration, where agentic systems can plausibly absorb 30-50 per cent of tasks by the mid-2030s, and in corporate overheads, which a disciplined program holds flat while production triples.
There are two opposing overlays. Classified submarine work will avoid frontier cloud-delivered AI; accredited sovereign capability will lag the commercial frontier by ~two to four years, so whatever AI does in a commercial law firm in 2033, assume Osborne lags. And: Net-of-AI creates its own demand: several hundred of roles e.g. digital-thread engineers, model-based systems specialists, secure AI operations, AI assurance/verification – with few of these roles found in any AI gross baseline.
The estimate aggregate effect is a net workforce requirement perhaps 8-12 per cent below gross by 2036: less than naive exposure scores would imply, because the netting exposes where the shortage is not, and vanishes where the shortage is most prevalent.
Uncertainty requires asymmetric provisioning
Scenario A – Agentic AI in accredited environments moves far faster than assumed.
Risk- program-overhead netting could reach 50-60 per cent, and the Osbourne plan will excessively over-train white-collar roles. Result: Too many document controllers, procurement officers and schedulers employed. Cost: loss with some recoverable, Osbourne excess hires ideally redeploy into the general labour market. There are unproductive employment costs/redundancy and training waste. Correction time: multiples of months.
Scenario B – AI (or its regulatory acceptance) moves far slower than assumed.
Risk – Robotic welding certification stalls at the nuclear regulator, the trades shortfall may worsen by several hundred full-time equivalents. Result: Too few certified welders and nuclear specialists trained on the assumption robots would cover the gap. Cost: severe and slow to fix. These pipelines run five to ten years through training, certification and security clearance, and there's no spot market to buy from. Correction time is major impact: Under-supplying certified trades and nuclear specialists is an expensive, slow-to-correct error with no AI relief coming.
Mega- project rational risk management
Any rational plan wouldn’t try to track the AI trajectory precisely. It deliberately over-provisions the irreducible cohorts where being short is catastrophic (trades, nuclear stewardship) and deliberately under-provisions exposed cohorts where being short is more fixable (program administration).
This wide uncertainty band, far from being a reason not to plan, is exactly why a plan is needed. It’s reason to make netting explicit, so that the error budget is spent where errors are best recoverable.
The Australian Submarine Agency would be negligent, indeed derelict, if did not annually prepare an AI-adjusted workforce demand statement for the Osborne program. This would be a first of its kind by any Australian public entity, done at an enterprise level, within a specified deployment environment.
Harder to apply economy wide – yet other nations are on to it
The jump to whole of economy is admittedly harder and is perhaps why JSA/DEWR hesitates behind present caveats, unwilling to chart a way to cross this chasm. Net-of-AI requires details of the deployment environment: what capital stock the AI runs on, what regulation governs it, what data it can touch, how fast an organisation can absorb it.
Notably, the BLS started with occupational empirical case studies and worked outward. Australia's case studies are waiting at Osborne, SSN-AUKUS is an opportune proving ground. It would create the methodological precedent that national forecasting needs: an accumulating library of program-level AI nettings from which JSA might then build beyond what its disclaimer presently concedes is missing.
The question is whether we count the machines before, or after, we have trained a decade of people for work AI does and people will not do – and failed to train them for the work only they can do.
Dr Craig Fowler runs JCSF Consulting and has held a wide range of roles across the Tertiary Sector, including former Managing Director of NCVER.