The care workforce surge

Three claims, graded separately

The projection contains three claims that do not stand or fall together. Each was checked against named sources — federal statistics and regulation directly, with professional-body, advocacy, and trade-press material labelled as such where it is used. The full working, including what could not be confirmed, is in the Validation Record.

Healthcare is not receiving a wave of new caregivers who stay. It is settling into a permanent high-volume, low-tenure workforce — and the people expected to orient, delegate to, verify, and rescue every entrant are themselves leaving.

Why the constraint is supervision

Read together: millions of entries, near-equal exits, a training floor that varies by state, and a shrinking pool of experienced nurses expected to supervise all of it. Every entrant consumes the one input that is not growing.

And why the obvious co-pilot is the wrong one

The case for building novice-first is strong. In a study of 5,172 customer-support agents, AI assistance raised resolved issues per hour by 15 percent on average — with gains concentrated among less-experienced workers, alongside improved retention, while the most experienced saw small quality declines (Brynjolfsson, Li & Raymond, QJE, 2025).

The case against the obvious implementation is just as strong, and it is specific to health:

• Capability does not survive the handoff. In a randomized trial with 1,298 participants, the models alone identified relevant conditions in 94.9 percent of cases. The same models, used by the participants, produced under 34.5 percent — no better than the control group using ordinary methods (Bean, Payne et al., Nature Medicine, February 2026). Benchmarks measured the model. The trial measured the pair. The pair failed.
• Assistance erodes unassisted skill. After AI was introduced across four endoscopy centres, adenoma detection on colonoscopies performed without AI fell from 28.4 to 22.4 percent (Lancet Gastroenterology & Hepatology, August 2025). That happened to experienced clinicians. A workforce whose skill is still forming is more exposed, not less.
• Confident wrong output flips right answers. Participants given incorrect decision support made 51.7 to 65.8 percent more commission errors than with no support at all (Lyell et al., 2017).

Build novice-first — but the novice-facing product is not an answer engine. Orient. Make the user commit. Then compare. Then escalate.

Supervision is the scarce resource

Thirteen principles govern this work. Six carry the most weight:

1. Supervision is the scarce resource. Optimize safe supervisory throughput per experienced nurse. A feature whose metric is “fewer questions reached a nurse” must prove the suppressed questions were ones a nurse did not need to see.
2. Never the last check. Every consequential output ends at a named human with the license and authority to act.
3. Scope is enforced, not inferred. Outside a user's declared role, setting, and jurisdiction, the answer is a person — never a hedged partial answer.
4. The system never wears a license. No nurse title, no clinician persona, no implication that a licensed person authored the output. Doctrine here, and law in Delaware, Oregon, California, and Washington.
5. Deskilling is a tracked harm. Unassisted practice is measured on a schedule. If unassisted performance declines, the response is less assistance, not more.
6. No alibi for understaffing. If a deployment coincides with less supervision, less orientation, or fewer staff, the deployment is out of doctrine and support is withdrawn.

The Ask Ladder

The single habit the whole first release exists to install:

1. Look. It is on the care plan, the assignment sheet, the label, the posted policy.
2. Ask the system. Your role, your scope, your process, your words, your rehearsal.
3. Ask a person. Anything about this person's condition, care, or any judgment call. The system helps form the question; it does not answer instead.
4. Escalate now. Never gated behind a form, a login, or “let me help you word that.” Latency measured in seconds.

When in doubt, go up a rung. Ambiguity resolves upward, always.

Person, then dyad, then unit, then institution

The wedge is narrow enough to state in one sentence: a browser-only, no-install, no-patient-data orientation pack that teaches one thing well — what is mine to do, and how do I raise a hand. Each stage must produce its evidence before the next begins. Adoption is not evidence; satisfaction is not safety.

1. Instrument before building. Metric definitions, an adversarial refusal test bundle, consent language, synthetic scenarios. Gate: every refusal class has a test that fails against a deliberately weakened prompt — proving the tests detect, rather than passing vacuously.
2. Entrant Orientation Pack. Scope Card, Escalation Card, the Ask Ladder, commit-then-compare rehearsal, the unassisted check, localized boundary text. Gate: out-of-scope attempts fall and appropriate escalation rises against each user's own baseline — or the negative result is published. Any clinical-answering leak is a hard fail.
3. The preceptor loop. Two consenting people, one shared record, questions shaped before they arrive — with urgent signals bypassing shaping entirely. Gate: fewer low-value interruptions with no increase in missed or delayed escalations. Both halves required.
4. Unit visibility, non-evaluative. Where escalations cluster; where entrants work without a supervisor. Aggregate only, never a per-person score. Gate: a hostile reviewer cannot re-identify an individual's performance, and a leader names a supervision decision they changed.
5. Institutional pilot, separately authorized. Named owners, written scope, data agreement, local review, stop switch, pre-registered outcomes. Gate: negative results published, with that commitment made before the pilot starts.

Stated in advance: what would falsify this

• Health-care job growth falling below the all-industry average across two projection rounds.
• Direct-care turnover falling durably below about 30 percent — which would mean this was a pay problem, not a software one.
• Experienced-nurse intent-to-leave dropping while faculty and clinical-placement capacity expand.
• A trial of a scoped, escalation-first assistant showing worse unassisted performance, worse escalation timeliness, or more scope violations than an unassisted control.
• Regulation classifying scope guidance to unlicensed workers as clinical decision support requiring authorization this project does not hold.

The failure mode to avoid is not being wrong. It is being unfalsifiable.

The five documents

Related records in this repository

• Nurse Formation doctrine and playbook — the learning method this inherits: commit-then-compare, human evaluative authority, AI never grading alone.
• NIN Knowledge Commons — the distribution contract: versioned, reviewed, provenance-bearing Knowledge Packs.
• The Directive and the governed harness — the risk, data, and action dimensions this plan sits inside.