Health care has been named in nearly every AI-and-jobs forecast — usually as one of the safest sectors, occasionally as the next to be automated. A perspective published in the New England Journal of Medicine on 12 September 2026 makes a more specific and more interesting argument: that AI agents, even ones capable of real cognitive work, could end up growing the clinical workforce rather than shrinking it.
🔍 THE BOTTOM LINE
Dr Dhruv Khullar, a physician and health policy researcher at Weill Cornell Medicine, does not deny that AI will automate clinical tasks — he takes it as given. His case is that automating tasks is not the same as automating jobs, and that three well-established economic ideas all point the same way in medicine: efficiency expands demand (Jevons paradox), the amount of work is not fixed (the “lump of labour” fallacy), and high-stakes work still needs humans to supervise it (O-ring theory). If he is right, the sector many forecasters treated as exposed may instead become one of AI’s clearest job-creation stories. It is a perspective essay — an argument, not a data set — and the honest read treats it that way.
The Three Arguments
Jevons paradox: cheaper capacity means more patients served. When a technology makes a resource more efficient, total use of that resource tends to rise, not fall. Khullar’s examples are cataract surgery and joint replacement: both became far more efficient over recent decades — less clinical effort, shorter recovery, better safety — and both ended up serving far more patients, supporting more surgical and nursing employment, not less. His claim is that AI could do the same for other services, especially where AI systems are priced near marginal cost and bring the cost of delivering care down. Lower cost per consultation, on this logic, means more consultations — not fewer clinicians.
The “lump of labour” fallacy: work is not a fixed pie. The fear that there is a fixed quantity of medical work to be divided between humans and machines is, Khullar argues, the same fallacy that has accompanied every general-purpose technology. AI may allow clinicians to prevent or treat conditions that are not currently treatable at all, creating “new treatments, care modes and opportunities for specialization” — and with them, demand for capabilities that did not previously exist.
O-ring theory: in high-stakes work, supervision is the job. Named for the faulty O-ring on the space shuttle Challenger, O-ring theory holds that when any single error can undermine an entire outcome, the value of reliable human oversight rises as systems get more powerful. Medicine is the canonical high-stakes case: interpreting results, negotiating treatment plans, and carrying legal and moral responsibility for decisions. “If AI automates some clinical tasks, the value of nonautomated, human tasks may increase,” Khullar wrote.
Where This Fits in the Evidence So Far
It is worth being precise about what kind of contribution this is. The NEJM piece is theory and history, not payroll data. But it lands in a year that has produced genuine data pointing in several directions at once:
- Revelio Labs’ September readout found employment in the most AI-exposed US occupations down about 6 percent relative to the least-exposed since late 2022 — and 19 percent among workers aged 22 to 25 — while also finding that AI-adopting firms keep growing headcount relative to non-adopters.
- The New York Fed’s survey work, which we covered earlier this month, found only a small share of AI-using firms had actually laid anyone off, with retraining the far more common response.
- Gusto’s payroll study of small US businesses, also this week, found AI adopters hiring more than non-adopters, with health care showing the strongest hiring effect of any sector — dental practices and mental health clinics growing headcount nearly 20 percent more than non-adopters by hiring more care providers.
The pattern across the firm-level data — AI adoption coinciding with more hiring, not less — is at least consistent with the Jevons logic Khullar spells out: as the cost of capacity falls, the use of capacity rises. The NEJM essay supplies the mechanism the small-business payroll data hints at.
On the other side, the entry-level data is real: Stanford’s ongoing canary research on the widening entry-level gap in AI-exposed roles, and Revelio’s 19 percent youth gap, both show that the distribution of AI’s effects is uneven even when the aggregate is not catastrophic. A growing clinical workforce and a squeezed graduate market can happen at the same time — and in several countries, already are.
The NZ Read
New Zealand has more riding on this question than most countries. Health care is consistently among the largest employing sectors, and the country runs a persistent shortage: successive workforce plans have flagged gaps in nursing, general practice and allied health, and offshore recruitment has carried much of the load. Te Whatu Ora’s own workforce planning has projected demand growth driven by an ageing population, entirely independent of AI.
That makes Khullar’s argument unusually relevant here. If AI agents reduce the administrative load on NZ clinicians — the documentation burden that Oracle Health’s nurse-facing tools and this country’s own shadow-AI clinical note adoption have already put on the record — the Jevons question becomes concrete: does freed clinician time become more patient appointments, or budget savings? A workforce that is already short changes the arithmetic. In a health system with unmet demand, efficiency gains have somewhere to go: into treating the people currently waiting.
The counterpoint also travels. AI clinical tools carry documented failure modes — from the Ontario medical-scribe audit that found errors and hallucinations in clinical notes to the governance questions health systems are still working out — and every efficiency argument in the NEJM piece assumes the tools are supervised well. O-ring theory cuts both ways: it is an argument for keeping clinicians in the loop, and an argument for why bad AI deployments could damage trust in exactly the sectors doing the hiring.
What It Means for Careers in Health
For students and career-changers weighing health careers against the AI headlines, the perspective offers a reasoned case that the sector’s long-run demand story survives the technology. It also sharpens the skill question. If automating tasks raises the value of the non-automated ones, the premium shifts to exactly what is hardest to automate: judgement under ambiguity, patient trust, supervision of machine output, and the interprofessional coordination that turns tasks into care. That reading matches what employers have signalled in job-ads data all year — NZ employers’ fastest-growing AI skill demand is ethics and governance, not coding.
None of this settles the argument. Khullar’s own framing is careful: clinical roles will change, some jobs will be replaced, and the expansion case is a long-run one grounded in economic theory and historical precedent. Readers weighing a health career should hold both halves: the theory is coherent, the history is real, and the data confirming it in practice is still being written.
❓ FAQ
Who wrote the NEJM perspective on AI and the clinical workforce? Dr Dhruv Khullar, associate professor of population health sciences at Weill Cornell Medicine and a hospitalist at NewYork-Presbyterian/Weill Cornell Medical Center, published 12 September 2026 in the New England Journal of Medicine.
What is Jevons paradox in the context of AI and health jobs? When technology makes a resource more efficient, total consumption of that resource tends to rise. Khullar cites cataract surgery and joint replacement: as they became more efficient, far more patients received them. He argues AI could have the same effect on health care services, increasing rather than reducing demand for clinicians.
Does AI threaten radiology jobs specifically? Radiology is often cited as highly exposed, but the NEJM perspective argues exposure to automation does not equal job loss, because oversight, integration and judgment remain human work. Concerns about radiology automation have circulated for years; the perspective is a counterweight argument rather than new employment data.
Is this relevant to New Zealand’s health workforce? Potentially — NZ runs persistent clinical workforce shortages in a sector that is among the country’s largest employers. If AI reduces administrative load and demand expands to meet it, the effect would be more clinical care delivered, not fewer clinical jobs.
Sources
- Dhruv Khullar — Artificial Intelligence and the Future of the Clinical Workforce, New England Journal of Medicine (12 September 2026), DOI: 10.1056/nejmp2607831
- Medical Xpress / Weill Cornell Medicine — How will AI impact the future of the clinical workforce? (14 September 2026)
- Revelio Labs — Monthly Workforce Update, August 2026
- Gusto — Small Businesses that Adopted AI Are Hiring Faster (10 September 2026)
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
