Replacement, augmentation, time: How AI co-clinicians will redefine health care workforce economics

When Google DeepMind published its artificial intelligence (AI) co-clinician research, the usual arguments started up again: Will AI replace doctors? Will it not? Who wins, who loses? I’ve heard this debate enough times that it almost feels scripted at this point. But having spent years investing at the crossroads of technology and health care, I keep arriving at a different place from most: This isn’t a replacement story at all. It’s an augmentation story, and that distinction carries more weight than people realize, especially when you start thinking through what it actually does to workforce economics in health care.

The triadic model and what it actually means

The model DeepMind is building isn’t AI-instead-of-clinician, but a triadic structure in which the AI operates under physician supervision rather than beside or above it. The results of the new Google DeepMind benchmark are truly astonishing: Zero critical errors in 97 of 98 realistic primary care evaluations, and the system matched or outperformed physicians on 68 of 140 assessed consultation dimensions. In blind head-to-head testing, 67% of physicians preferred it over existing clinical tools, and when put directly against GPT-5.4 thinking with search, it came out 63 to 30. That last number I found particularly interesting, because it’s not a comparison against some older baseline. It’s against one of the most capable general-purpose models available at the time of evaluation.

What the AI does well is fairly specific, though: Fast evidence synthesis; structured patient intake; and real-time analysis across multimodal inputs, such as video, audio, images, documentation and routine guidance. Those are things that consume hours without necessarily requiring the judgment that took a decade in training to build. Human clinicians are still clearly better at the other stuff: reading a room, catching the detail that doesn’t fit the pattern, guiding a physical examination, the kind of contextual judgment you cannot encode in a model. So the division of labor here is grounded in where genuine complementarity exists.

The time problem is the real problem

Here’s what I keep coming back to — it’s less exciting than the AI benchmarks, but it’s more important. Landmark time-motion research found that physicians allocate just 27% of their working hours to direct patient contact, compared with 49% spent on electronic health records (EHRs) and desk work, which is nearly two hours of documentation for every one hour of patient care.

The burnout numbers follow from that ratio almost mechanically. For 2025, 41.9% of physicians reported experiencing at least one symptom of burnout, and EHR documentation and prior authorization continued to surface as the top drivers. This is a workforce that is already exhausted, and the supply side doesn’t offer relief: The Association of American Medical Colleges projects a U.S. physician shortfall of between 13,500 and 86,000 by 2036, while the World Health Organization is tracking a global health worker deficit of 11 million by 2030. The system is already running close to its limits. An AI co-clinician that handles triage, note generation, protocol lookups and follow-up coordination will actually save hours rather than put more burden on clinicians.

Where this works in practice

Primary care and chronic disease management are probably the clearest examples, and they’re also the areas under the most pressure right now. AI can run initial consultations via video, evaluate visible symptoms, reconcile medication histories and draft care plans before a physician reviews the case.

In wound care, a domain Niterra has invested in, the application is quite tangible. The AI analyzes serial images, tracks healing trajectories against clinical benchmarks and flags deviations that justify a specialist’s attention. Instead of a specialist doing a repetitive visual assessment at every visit, regardless of whether anything changed, they intervene when it actually matters. Remote monitoring works on a similar logic: continuous inputs from wearables and sensors feed the AI, enabling proactive adjustments without requiring a clinician to be in the loop for every data point.

Integration with existing clinical workflows is genuinely hard, and the real-world performance of these systems outside controlled evaluations is still being established. But the directional case is solid.

Staffing models will have to change

The staffing implications are significant, and I don’t think the industry has fully worked through them yet. If AI tools genuinely multiply clinician output (and the evidence suggests they can), then hospitals and clinics can sustain effective care with physician teams that are smaller in number but sharper in scope. Nurses, midlevel providers and technicians who currently hand off to physicians for things AI can now handle have a real opportunity to take on meaningfully expanded roles. This is a structural correction for roles that have historically been underutilized relative to the clinical judgment they actually carry.

Training will have to change as well. Not just adding an AI literacy module somewhere in year three of residency, but more fundamentally rethinking what clinical training is preparing people for. Judgment under genuine uncertainty. Collaboration with AI systems in live clinical environments. The empathic, relational work that no model is anywhere near replicating.

The investment angle

From an investment perspective, the transition creates opportunity in fairly specific areas. At Niterra, we’re focused on hardware-software convergence: advanced biosensors, edge computing infrastructure for low-latency AI inference at the point of care, secure and auditable data platforms and therapeutics that pair with intelligent monitoring. AI co-clinicians are only as good as the inputs they work with, and the companies that generate rich, real-world, high-fidelity data from remote monitoring, minimally invasive sensing and wound care are building the substrate the AI needs to perform.

Do the economics work?

The upside case is credible but still somewhat theoretical at the aggregate level, and I think it’s worth being honest about that. A Lancet Digital Health meta-analysis across 69 studies found that AI diagnostic systems matched or outperformed clinicians in nearly half the cases reviewed, across imaging, pathology and EHR analysis, supporting the missed diagnosis reduction argument at a systemic level. McKinsey has projected net savings of $200 billion to $360 billion annually from AI and automation in U.S. health care, but that’s a projection contingent on implementation quality, not a guaranteed outcome. Validation, regulatory frameworks, clinician trust and equitable access — one of those is automatic, and the gap between potential and realized value in health care AI has historically been wide.

The human role doesn’t shrink in this model. The complex diagnostics, the ethical calls, the therapeutic relationship. That’s where training, experience and genuine clinical talent actually live. Offloading the routine preserves all of that.

The demographic backdrop makes the urgency hard to ignore. By 2034, the U.S. Census Bureau projects adults over 65 will outnumber children under 18 for the first time in American history. Globally, the United Nations estimates that there will be 1.6 billion people aged 65 and older by 2050. Clinician burnout, geographic maldistribution and demand curves that hiring pipelines alone cannot close are alarming problems.

AI co-clinicians won’t solve everything. But they offer a credible path toward bending the cost and capacity curves without sacrificing quality of care. The future here is clinicians who can genuinely do more. The organizations working that out now will have a structural advantage that compounds over time.

Thomas Kluz is a distinguished venture capitalist with over a decade of experience. He’s the managing director at Niterra Ventures, where his investments focus on energy, mobility and health care. With deep expertise in health care-focused venture capital, he has a proven track record of success with various organizations, such as Qualcomm Ventures and Providence Ventures.