The Liability Gap: Why Underwriting Is the Real Unlock for Clinical AI

Clinical AI is finally good enough to matter-but the real gating factor is not how impressive the accuracy metrics look on paper, it is underwriting. Whoever is willing to stand behind clinical AI with a clear medical malpractice and product liability framework is going to unlock adoption and win disproportionate value.

The Clinical AI Paradox

Dozens of AI-enabled devices and software tools have now been authorized by the FDA, spanning diagnostic imaging, pathology, clinical risk stratification, genomic interpretation, and early warning systems across acute and ambulatory care. Yet real-world use remains surprisingly narrow: with more than 1,200 AI medical tools FDA-cleared, fewer than 15% are used in daily practice despite strong performance metrics.

At the same time, the most visible progress in agentic AI is happening on the administrative side. Health systems are rapidly deploying AI agents in call centers, prior authorization, and revenue cycle management because the liability is lower and the ROI is immediate. 

This is the core paradox of clinical AI: health systems are convinced by the promise-earlier detection, fewer misses, more consistent care at lower cost-but remain hesitant to rely on these tools in the messy, high-stakes reality of clinical practice. When you ask executives and physicians why, they rarely point to model performance. They point to liability.

Promise Is No Longer the Problem

From an evidence perspective, clinical AI is increasingly compelling. AI-assisted interpretation of imaging and pathology is reducing diagnostic error in studies spanning radiology, dermatology, and ophthalmology. Clinical risk stratification models-predicting acute deterioration, disease progression, or treatment response-are outperforming legacy scoring systems in prospective validation. In oncology, AI tools are beginning to identify genomic signatures and treatment sensitivities that alter care decisions at the individual patient level.

If clinical AI were just another device with incremental benefit, you would expect adoption to follow the usual pattern: early KOLs, then guideline mentions, then payer support. Instead, we see pilots that work on the numbers but stall in deployment. 

Why Liability Is the Bottleneck

Clinicians today operate under a well-understood malpractice régime: they are liable when their care falls below the applicable standard of care for a reasonably prudent practitioner. AI upends that equilibrium in at least three ways:

• Ambiguous responsibility. When an AI system contributes to a harmful decision, it is often unclear whether liability should fall on the clinician, the hospital, the software vendor, or all of the above.
• Evolving standard of care. As AI tools become more capable, failing to use them could itself be framed as a deviation from the standard of care; yet relying on them too heavily risks being faulted when they err.
• Software as ​“product.” Courts have historically struggled to apply traditional product liability doctrines (design defect, failure to warn, breach of warranty) to software, especially when it is continuously learning and updated.

A recent New England Journal of Medicine–linked analysis makes the point bluntly: uncertainty about who will be sued when AI goes wrong is ​“daunting” and may chill adoption-even when the technology could reduce overall harm. Legal scholars reviewing AI in medicine reach similar conclusions, emphasizing that today’s frameworks were not designed for self-updating, data‑driven tools that blur the boundary between product and practice.

What Autonomous Cars Teach Us

We have seen this movie before with autonomous vehicles. Early data suggest that properly operated self-driving systems can reduce certain types of crashes caused by human errors, yet the legal and social response to an autonomous vehicle accident is existential for the autonomous car companies involved.

Traditional auto liability regimes were built around human negligence-drunk driving, inattention, speeding. With robotaxis, responsibility shifts toward manufacturers and software developers, triggering product liability theories, strict liability, and deep‑pocket targeting. Law firms already advertise around ​“who is liable when a self-driving car crashes?”, emphasizing product defects and software failures.

The result is hypocritical perfectionism: we accept tens of thousands of human road deaths annually, but expect self-driving systems to be nearly infallible. A single statistically rare but emotionally salient failure can generate lawsuits, regulatory scrutiny, and reputational damage out of proportion to the risk reduction provided across millions of miles.

Clinical AI is on a similar trajectory. Human clinicians make errors at measurable rates; AI systems can reduce those rates. But when an AI tool is involved in a miss-especially one it ​“should” have caught-the combination of deep pockets (vendors, health systems) and moral expectations of machines creates asymmetric legal downside. That asymmetry is a major headwind to adoption.

Why Underwriting Becomes the Wedge

If the technical case for clinical AI is increasingly solid and the legal environment is increasingly complex, the natural question is: who will absorb and price this new category of risk?

Today, liability for AI‑related harm is effectively distributed across:

• Individual clinicians (malpractice)
• Hospitals and health systems (enterprise and vicarious liability, negligent credentialing)
• Software developers and device manufacturers (product liability, failure to warn, breach of warranty)

In practice, physicians and hospital general counsel often assume they will end up holding most of the bag, because their malpractice and general liability policies are mature and visible, while coverage for AI‑specific risk is undefined or excluded.

This is the strategic opening. The entity that is willing to explicitly underwrite clinical AI liability-in a way that is transparent, priced, and contractually clear-will:

• Remove a major psychological and legal barrier for health systems evaluating AI tools
• Create a durable moat against point‑solution competitors that cannot match its risk appetite or actuarial sophistication
• Sit in the data exhaust of AI‑enabled care, improving its pricing over time and potentially becoming the coordinating layer for safety, monitoring, and performance improvement

Viewed through this lens, clinical AI is not ​“just software.” It is an insurance and risk‑transfer problem wrapped around a software asset.

The goal is to make the liability story boring. That means clear playbooks for when AI is followed or overruled, documented error rates, monitoring plans, and pre‑negotiated coverage so that a single incident does not become an existential event for the institution.

The Opportunity for Investors and Founders

For investors in healthcare, it is tempting to gravitate toward the ​“pure” AI companies: better models for imaging interpretation, risk stratification, diagnostic support, or clinical decision‑making. Those will remain important. But the strategic leverage may sit with platforms that combine:

• Strong clinical AI or aggregation of best‑of‑breed tools
• Embedded workflow and EHR integration
• And critically, embedded liability coverage-whether via captive insurance structures, partnerships with specialty carriers, or novel risk‑pooling mechanisms

The moat is not the model. It is the willingness to be wrong and stay solvent.

On the payer and policy side, early work suggests that failing to adopt high‑quality AI could itself be construed as negligence as the standard of care evolves. That flips the narrative: not using AI may become riskier than using it-if the liability framework is well‑designed.

Autonomous vehicles offer a cautionary tale. Companies that treated liability as an afterthought found themselves whipsawed by public opinion and legal uncertainty with each crash. In clinical AI, the winners will be those who start from the opposite premise: assume the technology will sometimes be wrong, design for that reality, and price it.

Where This Goes Next

We are moving toward a world in which hundreds of AI‑enabled tools shape diagnosis, risk stratification, and treatment selection. The question is not whether AI will be in the loop, but on whose balance sheet its failures will sit.

From a system perspective, the rational objective is to minimize population‑level harm, even if that means accepting that no technology-human or artificial-will be perfect. From a legal and financial perspective, however, we currently hold AI to a higher, almost mythical standard of infallibility. That mismatch is suppressing adoption of tools that could save lives.

The inflection point will come when a set of actors-insurers, health systems, AI vendors, or new intermediaries-are willing to say: we will underwrite this. When that happens, clinical AI can finally be judged on how much it reduces suffering across millions of patients, not whether it makes a single mistake.

Whoever is first, and best, at underwriting that risk will not just enable clinical AI-they will own the rails it runs on.