AI and the clinician shortage: what works, what doesn't, and what to expect by 2028

The healthcare AI industry has a shortage narrative problem. The narrative goes: we face a clinician shortage of historic proportions; AI can multiply clinician capacity; therefore AI can close the shortage gap. Each premise is roughly true. The conclusion is not.

What AI is actually doing to the clinician workforce is more interesting, and more useful, than the closing-the-gap story. It is reshaping the structure of clinical work in ways that change which shortage is most acute, which clinician types are most constrained, and which tasks are the limiting factor on care delivery. Understanding that reshaping is more practically valuable for health system workforce planners than the ambient claim that AI will somehow make the shortage less serious.

The shortage numbers that matter

Workforce projections are notoriously unreliable, but the directional signals across specialties are consistent enough to plan against.

Primary care faces a shortage that is structural, not cyclical. The AAMC’s 2025 projections put the primary care physician gap at between twenty-two thousand and thirty-six thousand by 2036. The pipeline is not closing this gap — medical school class sizes have grown modestly, but primary care specialty selection among graduating physicians has continued its three-decade decline. The workforce expanding into primary care gaps consists largely of NPs and PAs, whose numbers are growing rapidly but whose scope of practice remains contested state by state.

Nursing shortages are geographically and specialty-concentrated rather than uniformly severe. ICU nursing, surgical nursing, and psychiatric nursing are the pinch points; home health nursing has chronic shortage conditions that predate the pandemic. The overall RN supply has recovered from the pandemic-era crisis, but the distribution problem — too few nurses in rural markets, specialized settings, and night/weekend shifts — has not improved and is unlikely to respond to aggregate supply increases.

Psychiatry presents the starkest numbers. The SAMHSA estimates place the behavioral health provider shortage at roughly one hundred forty thousand, spanning psychiatrists, psychologists, LCSWs, and other licensed mental health professionals. Telepsychiatry has meaningfully expanded access in rural settings, but wait times for outpatient psychiatric evaluation in most metropolitan markets remain measured in months. This is not a problem that technology alone can solve — it requires changes in reimbursement, training pipeline, and scope of practice that are moving slowly.

Medical imaging reads are a special case. Radiologist supply has not kept pace with imaging volume, and the volume growth shows no sign of slowing. This is the space where AI capacity extension is most directly applicable and most measurably deployed.

What AI-enabled capacity extension looks like in practice

The clearest real-world examples of AI capacity extension share a common pattern: the AI handles a defined, bounded task that previously required full clinician scope, allowing the clinician to supervise more outputs than they could generate individually.

In radiology, AI triage of emergency imaging studies — identifying intracranial hemorrhage, pulmonary embolism, aortic pathology — has been deployed widely enough to have real-world performance data. The effect is not to reduce the number of radiologists required but to allow on-call radiologists to prioritize time-sensitive findings. A radiologist who previously read studies in the order they arrived can now read the genuinely emergent cases first. This is valuable without being transformative, and it is the honest framing of what AI triage is doing.

The ThinkSono vascular ultrasound case is worth examining as a model of a more fundamental type of task shifting. ThinkSono’s AI-guided vascular ultrasound protocol allows a trained nurse or health worker with no prior sonography training to perform diagnostic-quality deep vein thrombosis studies under AI-guided instruction, with results that compare favorably to sonographer-performed studies in validation settings. The DVT evaluation, which previously required scheduling with a sonography department staffed by credentialed vascular technologists, can now be performed at the point of care by non-specialist staff. This does not eliminate the need for sonographers — it eliminates the need for a sonographer for this specific, bounded clinical task.

The pattern generalizes: AI-guided task shifting is most feasible when the task is well-defined, the decision boundary is clear, and the consequence of error is detectable and manageable. DVT assessment fits. Complex cardiac imaging does not. Structured symptom triage in primary care fits better than diagnosis of undifferentiated illness.

What doesn’t work

The failures cluster around tasks that require contextual judgment, which is most clinical tasks.

AI-based patient triage in primary care has been extensively piloted and has not delivered the capacity extension that was predicted. The patients who most need triage are those whose presentation is most ambiguous — and ambiguity is precisely what current AI handles least reliably. Health systems that have deployed AI triage tools have generally found that the tools are effective at identifying clearly low-acuity encounters but that these encounters were not the capacity constraint in the first place. The patients consuming disproportionate clinician time and judgment are the ones the triage tool routes to human review anyway.

Psychiatric AI has been the most aggressively overclaimed segment. Mental health chatbots and AI-assisted therapy tools have proliferated, with marketing claims that span from “adjunct to therapy” to implicit suggestions of substitution. The clinical evidence for these tools in severe mental illness is thin. The evidence in mild-to-moderate depression and anxiety is more encouraging but remains early. And the regulatory environment for AI-based mental health tools has tightened considerably following several adverse event reports from 2024 and 2025. The honest role of AI in behavioral health capacity is in administrative functions — scheduling, no-show prediction, documentation — not clinical care delivery.

What to expect by 2028

The two-year outlook is not a shortage-closing story but a shortage-reshaping one. The tasks that AI can reliably handle will be more consistently off-loaded from clinical professionals, reducing the cognitive overhead per patient encounter without reducing the need for clinical professionals. The primary care visit will be shorter and better-documented because ambient AI handles transcription. The radiology read will be faster because AI handles preliminary triage. The home health nurse will need fewer administrative interruptions because AI handles care coordination documentation.

What won’t change: the shortage of people who can hold complex, ambiguous clinical situations with a patient. The diagnostic generalist. The psychiatric clinician who can conduct a safety assessment. The bedside nurse who can recognize subtle deterioration that doesn’t appear in the monitoring data. These functions are not at risk of AI substitution in any planning horizon that health system workforce teams need to act within.

The practical implication is that workforce planning through 2028 should treat AI as a task-shedding tool for administrative and structured clinical work, not a headcount-reduction tool for clinical FTE. Health systems that plan for AI to reduce clinical staffing requirements will be caught short. Health systems that plan for AI to allow existing clinical staff to work at higher volumes of complex care — while shedding the tasks that don’t require their training — will be better positioned.

The shortage is real. AI is useful. Neither sentence implies the other solves the problem.