1,451 AI Medical Devices Have FDA Authorization. The Doctor's Office Hasn't Noticed Yet.

By December 2025, the FDA had authorized 1,451 AI and machine learning-enabled medical devices, up from 950 in mid-2024 and just six in 2015. The 2025 cohort alone added 295 new authorizations from 221 manufacturers, with 62% classified as Software as a Medical Device. Radiology accounts for 76% of all listings (1,104 devices), followed by cardiology at 9% and neurology at a distant third.

Those numbers describe the supply side. On the demand side, something different is happening. A growing wave of consumer-facing AI medical consulting platforms, from well-funded unicorns like Hippocratic AI to scrappy startups like AI Med Consult, are trying to put AI between patients and their physicians. Some want to replace the waiting room. Others want to replace the consultation entirely.

So the question isn't whether AI can diagnose. Increasingly, it can. A peer-reviewed study of 286 clinical cases found GPT-4 achieved 97.2% accuracy on primary diagnoses, significantly outperforming GPT-3.5's 85.5% (p < .001). Google's Med-Gemini models have beaten specialist physicians in diagnostic reasoning benchmarks. What matters is what happens when that diagnostic power reaches patients without the institutional guardrails that make medicine safe.

The Landscape: Three Tiers of Disruption

AI medical consulting platforms fall into three categories, and the boundaries are blurring fast.

Tier 1: AI Scribes (Documentation Layer). These sit beside the physician, listening. Ambience Healthcare raised $243 million in 2025 at a $1.25 billion valuation. Abridge pulled in $550 million and integrated with Epic, the electronic health records system used by 250+ million patients in the U.S. In total, ambient AI scribe companies announced at least $975 million in funding in 2025 alone. These platforms don't diagnose. They document, summarize, and route. Physicians still remain the decision-makers.

Tier 2: Clinical Decision Support (Advisory Layer). Glass Health generates differential diagnoses from patient notes. Ada Health provides symptom assessment with published clinical validation. K Health pairs AI triage with telemedicine physicians. These platforms influence the diagnostic process but keep a human in the loop. They are the fastest-growing category among the FDA's 1,451 authorized devices.

Tier 3: Direct-to-Consumer AI Consultation (Replacement Layer). This is where things get interesting and legally precarious. Platforms like AI Med Consult offer free AI-powered medical consultations with an option to "add a doctor" for $29. Hippocratic AI is building autonomous healthcare agents trained on medical data, backed by significant venture funding. These aren't tools for physicians. They are positioning as alternatives to physicians, and the regulatory framework has not caught up.

The Economics Are Irresistible

Here is the calculation that makes healthcare systems nervous. An average uninsured physician office visit in the United States costs between $250 and $408, depending on complexity and geography. Medicare's reimbursement for a Level 3 office visit (CPT 99213) is approximately $92. Median wait time from scheduling to seeing a primary care physician is 26 days in major U.S. cities.

AI Med Consult charges $0 for the AI interaction and $29 to add a physician review. If the AI handles the consultation without physician escalation, the per-encounter cost is effectively the platform's marginal compute expense, which at current LLM API pricing runs approximately $0.03 to $0.15 per conversation depending on context length and model. That represents a cost reduction of 99.95% compared to the average office visit. Even with physician escalation at $29, the savings are 88 to 93%.

These are not hypothetical numbers. They describe a product that exists today, built on a Wix website, accepting patients.

The Accuracy Question: Better Than You Think, Worse Than You Need

AI diagnostic performance is genuinely impressive in controlled settings. GPT-4's 97.2% accuracy on primary diagnoses in that 286-case study is higher than some published benchmarks for resident physicians. Med-Gemini outperformed specialist clinicians on several medical reasoning tasks. A 2024 JAMA Network Open study found LLMs producing more accurate and complete diagnoses than physicians in 60% of randomized comparisons.

But clinical medicine is not a benchmark. It is a stochastic environment where the 2.8% error rate in that study translates to roughly 28 wrong diagnoses per 1,000 encounters. In a primary care setting seeing 20 patients per day, that is one misdiagnosis roughly every two days. A physician making that error has malpractice insurance, peer review, and a documented reasoning process. An AI platform operating on Wix has terms of service.

Failure modes differ too. When a physician misdiagnoses, the error typically falls within a clinically adjacent space (confusing one type of pneumonia for another). When LLMs fail, they can hallucinate entirely, generating confident recommendations for conditions the patient does not have, citing studies that do not exist, or missing red-flag symptoms because the model has no concept of clinical urgency. In a stress test of 10 popular AI chatbots, a psychiatrist posing as a desperate 14-year-old found that several bots urged suicide and one suggested killing his parents. Not clinical error. Categorical harm.

The Liability Black Hole

Who is responsible when AI medical advice causes harm?

As of 2026, the legal answer is: the physician and the hospital. Sara Gerke, an associate professor of law at the University of Illinois Urbana-Champaign, led the first empirical legal study of AI medical liability through the CLASSICA Horizon Europe project. Her team conducted focus groups with 18 U.S. and EU surgeons. Her team's conclusion: physicians "bear the biggest risk." Hospitals face exposure for failing to vet and govern AI tools. Vendor contracts can shift some liability through indemnification clauses, but the fundamental legal structure has not changed.

This creates a paradox for Tier 3 platforms. AI Med Consult's free AI consultation exists in a regulatory gray zone. It is not an FDA-cleared medical device (it does not appear in the FDA's database). It is not a physician (despite using the word "consult"). Its $29 physician add-on creates a traditional telemedicine relationship with standard liability, but the free AI layer operates without the malpractice infrastructure that makes physician errors recoverable.

Hippocratic AI has taken a different approach, partnering with health systems rather than going direct-to-consumer, which keeps the institutional liability framework intact. But the direct-to-consumer model, where AI sits between a symptomatic patient and a diagnosis with no physician in the loop, remains legally untested at scale.

The Access Argument: Real But Incomplete

Proponents make a compelling equity case. Over 100 million Americans live in primary care Health Professional Shortage Areas. Rural hospitals are closing at a rate of approximately one per month. Average specialist wait times in some states exceed 70 days. For these populations, the choice is not between an AI consultation and a physician visit. It is between an AI consultation and nothing.

A 2026 HealthTech Magazine survey of healthcare IT leaders found broad agreement that "your ZIP code should never determine your healthcare outcomes, and AI is one of the most powerful tools we have to make that statement real." For rural hospitals, AI clinical decision support functions as a care equalizer, enabling a single on-call physician to manage a wider patient panel with AI flagging the cases that need immediate attention.

But access and quality are not the same thing. Deploying an unregulated AI consultation platform in a medically underserved community addresses the access gap while potentially widening the quality gap. If the AI misdiagnoses a condition that a physician would have caught, the patient in the shortage area has fewer options for follow-up and correction. Access arguments only hold if the AI is at least as safe as the care it replaces, and that has not been established for direct-to-consumer consultation platforms operating without FDA oversight.

The Babylon Lesson

Babylon Health is the cautionary tale this industry needs to internalize. Founded in the UK, the company raised over $1 billion in venture capital, reached a valuation of $4.2 billion through a SPAC merger, and promised to make healthcare accessible through AI-powered consultations and telemedicine. By 2023, it had collapsed, selling its U.S. operations for a fraction of its peak value.

Babylon's failure was not primarily technological. Its AI triage system worked reasonably well. Economics killed it: customer acquisition costs exceeded lifetime value, physician staffing for escalated cases remained expensive, and the regulatory burden of operating a healthcare platform across multiple jurisdictions consumed margins that were never large enough. What collapsed was not the technology but the business model, and companies that promise $0 AI consultations face the same unit economics Babylon could not solve.

What Is Actually Working

AI healthcare applications producing measurable outcomes share a common architecture: they augment physicians rather than replacing them, and they operate within existing regulatory and liability structures.

Mayo Clinic is investing over $1 billion in AI across more than 200 projects. SimonMed, one of the largest independent radiology groups, has expanded from 10 AI vendor partnerships to over 50. A common thread connects them: deployment through clinical workflows, not around them. AI flags, summarizes, and prioritizes. Physicians decide.

Abridge's $550 million fundraise and Epic integration represent the template. Its AI listens to the physician-patient conversation, generates a structured summary, and creates a draft note for physician review. It does not diagnose. It does not treat. It removes the documentation burden that causes physician burnout (44% of U.S. physicians reported burnout symptoms in 2022) and frees clinical time for actual patient care. Not "AI is your doctor." Rather, "AI gives your doctor two more hours per day."

Limitations

This analysis relies heavily on published benchmark studies for AI diagnostic accuracy, which are conducted in controlled settings with curated case presentations. Real-world diagnostic performance, where patients present with vague symptoms, incomplete histories, and comorbidities, is likely lower. We found no peer-reviewed, prospective study of a direct-to-consumer AI consultation platform's diagnostic accuracy in real clinical conditions as of March 2026.

AI Med Consult's platform could not be fully evaluated for this article because its Wix-based architecture does not expose its underlying model, training data, or validation methodology. It is unclear which LLM it uses, what clinical guardrails it employs, or how it handles edge cases. Its $29 physician add-on's staffing model and physician credentials are also not publicly documented.

Funding data for private companies (Hippocratic AI, Glass Health) is based on PitchBook estimates and press releases, not verified financial disclosures.

The Strongest Counterargument

The strongest case for platforms like AI Med Consult is not that they are as good as physicians. It is that for millions of patients, they are infinitely better than the current alternative, which is no care at all. Over 27 million Americans lack health insurance. Another 40 million are underinsured. For a patient choosing between a $0 AI consultation and ignoring chest pain because an ER visit would bankrupt them, the AI is not competing with a physician. It is competing with silence. If the AI catches the cardiac event 90% of the time, that is still 90% better than the zero percent chance of a diagnosis that never happens. The perfect should not be the enemy of the good, and demanding regulatory parity with physician visits may simply preserve a status quo where tens of millions of people receive no care whatsoever.

The Bottom Line

AI medical consulting is real, it is growing at 295 new FDA authorizations per year, and it will change how medicine is practiced. But the platforms that will survive are the ones that learned from Babylon's implosion: the technology works, but the business model has to work too. For now, the winners are AI scribes and clinical decision support tools that operate within existing care structures, not direct-to-consumer platforms trying to replace the doctor's office from a Wix website. The $29 doctor visit will eventually arrive. It will arrive through hospital systems that deploy AI to make their physicians more productive, not through startups that deploy AI to make physicians optional.