As of the H1 2026, the AI jobs apocalypse hasn't arrived - but today's stable data won't predict tomorrow. What AI leaders now believe, and what it means for Australia and NZ work.
In the middle of 2025, the most striking voices on the future of work were not labour economists or union leaders.
They were the chief executives building the technology itself, and the message was alarming.
By the middle of 2026, several of those same people had changed their tune.
Understanding why, and understanding the large caveat that comes with it, is the best available guide to how AI is actually going to change the way we work.
The 2026 reversal
For most of 2025, Anthropic chief executive Dario Amodei was the rare frontier-lab founder willing to say out loud what many of his peers only said privately: that AI could eliminate up to half of entry-level white-collar jobs within one to five years, potentially pushing unemployment toward 20%.
He named the exposed sectors directly, including law, finance, consulting and software, and argued AI was advancing fastest precisely at the kind of cognitive office work those careers are built on.
OpenAI chief executive Sam Altman made similar warnings in 2025 about entry-level roles being at serious risk. Then, in May 2026, he reversed course, saying he had been "pretty wrong" about the economic and social impact even while the technical predictions held up. He framed it as good news he was happy to be wrong about.
Amodei, too, shifted his emphasis, moving from the stark wipeout framing toward the argument that automation may actually expand the total amount of work people do.
The reason for the softening is straightforward: the data through early 2026 did not back the apocalypse. The Budget Lab at Yale, tracking AI-exposed occupations against the monthly employment microdata, has repeatedly found no clear relationship between AI exposure and unemployment.
Employment stayed relatively stable while productivity rose. That single fact, more than any executive press release, is what reshaped the conversation.
Why "no change yet" tells you very little about what comes next
Here is the trap in that reassuring data: it is a rear-view mirror. The stability measured through early 2026 reflects the capability of the models that were actually deployed in 2024 and 2025, filtered through how slowly organisations change their hiring and their workflows.
It says almost nothing about what materially more capable, more autonomous models will do once they arrive and once firms have had time to restructure around them.
The researchers themselves are careful about this. The Budget Lab titled its analysis AI Is Probably Not (Yet) the Reason for Labor Market Weakening, and explicitly notes that the models have become increasingly powerful and that many observers expect them to eventually reshape the labour market.
Its own summary describes a picture of stability "at an economy-wide level" while flagging that better data is needed and that the analysis will be updated as capability advances. "No effect so far" is a statement about the past, not a forecast.
Three things make the lag particularly deceptive:
- Capability is still climbing steeply. Amodei has described the models progressing from roughly a smart high-schooler to a smart college student and beyond within a couple of years. If that curve continues, the share of work that is automatable in 2027 or 2028 is simply not visible in 2026 employment figures.
- Adoption lags capability, and restructuring lags adoption. McKinsey's surveys show most organisations have not yet scaled AI across the enterprise even with today's tools. The job-market effect of a capability shows up only after firms redesign around it, which can take years. So a flat chart today is consistent with a steep change later.
- Agentic execution changes the unit of automation. Earlier models automated tasks; agents that can carry out multi-step work begin to automate whole workflows. That is a step-change in exposure, and the 2026 data largely predates its mainstream deployment.
In other words, the honest reading of the reversal is not "the worry was wrong." It is "the worry was early." The current calm is exactly what you would expect to see in the lag between a fast capability curve and a slow institutional one. Treating it as the all-clear would be the costliest possible misreading.
What the leaders actually believe now
The headline view across the major AI leaders has converged on something more nuanced than either the doom or the hype. It is worth taking each position on its own terms, because they disagree in instructive ways.
Sam Altman (OpenAI) now separates the technical trajectory from the economic one. His position is that the models advanced roughly as expected, but the predicted social dislocation has not arrived on the timeline he assumed, and entry-level roles have proved more resilient so far than he feared.
Dario Amodei (Anthropic) has not abandoned his concern about a sharp, broad transition. His distinctive argument is about breadth: because AI acts as a general labour substitute with wide cognitive reach, it could affect finance, consulting, law and tech at once, leaving fewer adjacent fields for displaced workers to move into.
His revised framing allows that automation may grow the pie, while still pressing for policy preparation rather than complacency.
Demis Hassabis (Google DeepMind), a Nobel laureate, frames the long arc optimistically as a path toward what he calls "radical abundance," a near-future in which AI compresses scientific discovery so dramatically that medicine, energy and materials are transformed.
He is candid that the road there runs through a turbulent decade-long shakeout, and that the technology demands a new approach to learning and skills.
On jobs specifically, his view is the classic one: new tools historically create new and better work, and AI will make people "a little bit superhuman" rather than redundant.
Jensen Huang (Nvidia) is the most direct sceptic of the apocalypse narrative, having publicly disagreed with much of Amodei's framing. His position is that AI will replace some jobs, change many more, and create others, including changing his own role, but that the wholesale-collapse story is overstated.
The disagreement is real, but notice that even the optimists build in a hard transition: Hassabis's "golden era" sits on the far side of a decade of disruption. The debate is less about whether work changes profoundly and more about how painful the passage is and how it nets out.
What the firms and the data show
Step away from the founders and the picture from the big advisory firms and independent researchers fills in the middle ground.
McKinsey's research describes a landscape where AI use is now mainstream but deep, enterprise-wide impact remains uneven. The large majority of organisations report using AI in at least one function, yet most have not scaled it across the enterprise.
On agents specifically, roughly a quarter of organisations report scaling an agentic system somewhere in the business, with another large share experimenting. The firm's 2026 work pushes the theme that the winners are redesigning whole workflows, not bolting AI onto old ones.
McKinsey is also a live case study in itself. Its chief executive has described the firm operating with tens of thousands of AI agents working alongside its human consultants, with an ambition for agents to roughly match human headcount and for every employee to be supported by one or more agents.
Whatever one makes of the framing, it signals how quickly a knowledge-work business can restructure around agents.
On the hard numbers, three data points anchor the realistic case:
The tension between "AI can do a meaningful slice of tasks" and "jobs have not collapsed yet" is the single most important thing to hold in mind.
Tasks are not jobs. A role can lose 30% of its tasks to automation and still exist, reshaped toward judgment, exceptions and oversight, right up until the point where enough tasks go that the role itself is rethought. That tipping point is what the capability curve is driving toward.
The Australia and New Zealand picture
Almost all of the headline data above is American. The local picture matters for businesses here, and it tells the same story with sharper edges: a clear augmentation tilt for now, a genuine displacement risk later, and an institutional-readiness gap that is arguably wider than in the US.
On the augmentation side, Australia's Productivity Commission has been deliberately measured. Its chair pointed to Jobs and Skills Australia analysis finding only around 4% of jobs could be fully or mostly automated, while a much larger share, 30% or more, is subject to augmentation where some tasks shift to AI but a human stays firmly in the loop.
A forthcoming federal report on labour-market changes to February 2026 reaches a similar early read: employment outcomes for young tertiary graduates, the so-called canaries in the coalmine for white-collar automation, have so far held up.
But the same caveat applies locally, and more forcefully. The Commission puts Australia's likely AI productivity gain at just 0.4 percentage points a year, below many global peers, and is blunt that adoption alone does not convert into gains.
Realising them requires deep business transformation, new processes and new skills, which is precisely the slow institutional change that lags the technology. Australia's broader productivity growth is already at its weakest in six decades, so the gap between what AI could do and what local firms actually capture is the central risk.
The displacement scenario is not hypothetical here either. Commonwealth Bank drew headlines when it cut call-centre roles after introducing an AI chatbot, and not every forecaster shares the Commission's calm.
Yarra Capital's Tim Toohey has warned that if AI is deployed at scale over two years, his base case, Australian unemployment could climb above 6% from 4.3%, on the order of 293,000 job losses. A Mercer survey of Australian executives and HR leaders found striking pessimism about headcount, a reminder that intent inside firms can move faster than the aggregate statistics.
New Zealand shows the adoption-versus-readiness tension clearly. Research from Randstad found a majority of NZ employers already report AI lifting workforce productivity, yet a perception gap has opened up: leaders expect AI to affect a high proportion of tasks while fewer frontline staff see it the same way.
That gap, between executive expectation and employee experience, is exactly where botched rollouts and lost trust live. The local upside is real, with generative AI projected to add tens of billions to the NZ economy over the next decade, but only for organisations that pair the tools with genuine upskilling rather than treating access as the finish line.
The practical takeaway for AU and NZ organisations is that the global pattern holds, but the margin for complacency is thinner. The technology is the same; the constraint is local capacity to redesign work around it.
That favours businesses that can bring in specialist capability quickly to bridge the skills and transformation gap, rather than waiting to build it slowly in-house while the window narrows.
Why agentic AI is the real inflection point
The 2023 to 2024 wave of AI made small teams faster by generating drafts, code and analysis that a human still had to direct and stitch together.
The 2025 to 2026 wave is different in kind, because the agents increasingly execute multi-step tasks across tools, not merely produce text.
That is what moves AI from a smarter autocomplete to something closer to a digital co-worker, and it is the main reason the next few years of labour-market data may look nothing like the last few.
This is also where the genuine operational risk shifts. Once systems can take actions, the failure mode is no longer only that the model said something wrong but that the model did something wrong, such as misusing a tool or acting outside its guardrails.
That is why the better-run organisations are building human-in-the-loop validation by design, with humans supervising, auditing and handling exceptions rather than doing the repetitive core. Reliability, monitoring and accountability become first-class disciplines, not afterthoughts.
Small versus large, and the rise of the solo operator
One of the most consequential effects of agentic AI is on company size. If one person can direct a fleet of agents, the historical relationship between headcount and output breaks down.
Sam Altman popularised the provocation that the first one-person billion-dollar company was coming, something he framed as unimaginable before AI. Amodei has put a high probability on a solo-founded billion-dollar business emerging in the near term.
As of early 2026 no fully verified one-person billion-dollar company existed, so this remains a forecast rather than a fact, and much of the breathless commentary around it is marketing.
But the underlying trend is real and already visible in smaller numbers. Solo and near-solo founders are reaching revenue figures that previously required substantial teams, by renting existing infrastructure and using AI as a full-stack operator across coding, design, content, support and automation.
The most cited 2026 example is a telehealth commerce business that reportedly reached hundreds of millions in first-year sales with a headcount you could count on one hand, running margins well above larger incumbents.
The strategic implication cuts both ways. Large organisations gain enormous leverage but must fight their own structural inertia to realise it. Small players and individuals can now punch far above their weight, particularly in software and digital services, where a product is built once and updated continually.
The advantage increasingly goes to those who operationalise AI deeply, not those who merely experiment with many tools shallowly.
How the impact differs by field
AI is not landing evenly. A useful way to think about it is by how much of a role is routine cognitive work versus judgment, relationships and accountability.
- Software engineering: Among the most affected, with AI coding tools now handling substantial portions of routine implementation. The role is shifting toward architecture, review, and orchestrating agent output, while early-career hiring has tightened.
- Finance and professional services: High exposure on the analytical and synthesis-heavy tasks, which is exactly where MIT located much of the potential wage saving. Expect compression of the junior-analyst pyramid and more emphasis on client judgment and oversight.
- Law and consulting: Research, drafting and document synthesis are heavily automatable, but advice, negotiation, accountability and trust are not. The work moves up the value chain rather than disappearing, at least until agents close the reliability gap.
- Science, healthcare and R&D: This is the augmentation story Hassabis emphasises, where AI compresses discovery cycles and acts as a force multiplier on human expertise rather than a substitute for it.
- Operations, marketing and support: Among the fastest to adopt agents for execution, with humans increasingly handling exceptions, quality and the customer relationship rather than the repetitive throughput.
The four paths from here
Stripped of hype, there are four broad ways this plays out, and they are not mutually exclusive across different roles and industries.
- Augmentation: AI makes existing workers more productive without large net job loss. This is closest to what the early-2026 data shows, but it is also the path most dependent on capability plateauing or adoption staying gradual.
- Orchestration: The nature of work shifts from doing tasks to directing systems and agents. Fewer people produce far more, and the premium skill becomes managing AI rather than executing manually.
- Net displacement: Amodei's caution, in which AI's breadth removes work faster than the economy creates replacements, producing a painful transition that needs policy responses such as retraining and tax adjustments. This is the path the lagging data cannot yet rule out.
- Jevons-style growth: The optimistic case, where making cognitive work dramatically cheaper expands demand for it so much that total work grows. Cheaper analysis, code or design leads to far more of it being wanted, and headcount follows.
The honest conclusion is that the outcome is undetermined, and a meaningful part of it is a choice. The same technology can be deployed to hollow out a workforce or to amplify one.
Organisations that redesign workflows thoughtfully, keep humans accountable for judgment, and invest in skills will land closer to augmentation and growth. Those that treat AI as a headcount-reduction lever first will get the brittle, low-trust version.
What this means for the way we work
The lesson of the 2026 reversal is not that AI was overhyped. It is that the technical curve and the human-system curve move at different speeds. The models are advancing roughly as the builders predicted.
The economy, organisations and skills adapt more slowly and more unevenly, which is why the jobs apocalypse has not arrived on schedule, and exactly why today's stability is no guarantee about tomorrow. The capability is still climbing, and the institutions have barely begun to adjust.
For anyone planning a career, a team or a business, the durable bets are clear enough. Learn to direct AI rather than compete with it on routine output.
Build for judgment, accountability and relationships, the parts that do not automate. And operationalise the tools deeply in the workflows that matter, because the gap between those who experiment and those who execute is where the next few years of advantage will be won.
.webp)
