On September 21, in PLOS Global Public Health, Jim Campbell of King’s College London proposed a new discipline: Workforce Science. For decades, he argues, health systems planned their workforce by counting licensed professionals, their full-time equivalents and their density per 10,000 people. That count no longer describes what a health system can do. Capability, he writes, “is no longer solely human.”

He replaces the head count with the conceptual construct of “effective capability”, understood as a function of education financing, education quality, labor, the human-AI dynamic, governance capacity, and the conditions, rights and realities of work. Education, in his definition, is “the formation, maintenance and adaptation of capability” across a working human life. The planner’s question changes from how many workers to train and employ to how hybrid human-AI capability is “created, distributed, and stewarded.” The stakes are large: health and social work account for at least 1 in 10 jobs in many OECD countries, and 1 in 6 in Scotland.

The human-AI dynamic can be measured, and it has been. In 2024, in Nature Human Behaviour, Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone of MIT pooled 106 experiments, each of which tested the human alone, the AI alone and the two together. On average, the pair did worse than the better of the two working alone. The average hides the finding that matters. When the human was stronger than the AI, the pair beat both. When the AI was stronger, the pair fell below the AI working alone.

The human half of a hybrid is not a constant. Its strength decides whether the pair adds value or subtracts it.

Campbell names the hinge himself. AI, he writes, has “enormous potential to augment human capability if managed appropriately.” The last 2 essays on nonalgorithmic.com gave the well-managed version a name, amplification: AI that extends a capability the clinician already brings. It cannot multiply an absence, and Vaccaro’s 106 experiments show it in numbers.

Yet the human half is the half we rarely measure. On the same day Campbell published, a University of Toronto team mapped the literature on AI for health care quality and patient safety in the Journal of Medical Internet Research. Of 275 records, 13 measured what happens when an output reaches a clinician: whether an alert was accepted or overridden, how long action took, how much documentation it added. The tools worked when the output mapped to one immediate clinical action. They varied when it needed interpretation across layers of care, as sepsis alerts do.

The AI half, meanwhile, is getting a price. On September 24, Medscape reported that federal officials are considering paying AI tools that provide diagnostic or clinical care between 60% and 80% of what a human would receive for the same function. Whether that figure is a policy platform or a back-of-the-envelope number is unclear, Medscape noted. Medicare pays for clinician time, and AI has the opposite cost structure: expensive to build, almost free to run. “Every time they deploy it, it costs almost 0 cents,” said Ateev Mehrotra of Brown. “How the heck do we price that?”

Ravi Parikh of Emory would sort AI into 3 payment buckets: scribes as overhead, decision support folded into the underlying service, and a separate, lower payment for AI that could replace a clinician service. Lee Fleisher, a former chief medical officer at CMS, warns that the physician fee schedule is cost-neutral: “The more you add into that, the more physicians get devalued.” Campbell measures value another way: by how hybrid capability “contributes to human flourishing and wellbeing for the health worker, the patient and the population.” A price set per function has no term for the first of the three.

Read quickly, Vaccaro’s data seem to argue for that third bucket: where the AI is stronger, take the human out. The Toronto review suggests why that reading is too quick. Most of medicine is not one immediate action. It is an output that someone must interpret, for a particular patient, across layers of care. There, the human half still decides what the pair is worth.

Yesterday’s Standing Order read Joshua Sharfstein’s layered model of oversight, from licensing boards to hospitals. Every layer in it assumes a clinician strong enough to catch the machine. Campbell wants a discipline that can measure hybrid capability. The first thing to measure is whether the people in it are still being formed.

So: in your institution’s AI budget, where is the line item for forming the humans the AI will depend on?

Sources. Jim Campbell, “Workforce Science and the Human–AI economy of health and care,” PLOS Global Public Health, September 21, 2026 · Michelle Vaccaro, Abdullah Almaatouq and Thomas W. Malone, “When combinations of humans and AI are useful: a systematic review and meta-analysis,” Nature Human Behaviour, 2024 · Xu, Veillard and Kong, “AI for Health Care Quality and Patient Safety: Scoping Review,” Journal of Medical Internet Research, September 21, 2026 · Carrie Arnold, “Reimbursement Puzzle: How Should Doctors Be Paid for Using AI Tools?,” Medscape, September 24, 2026