White Paper

The Insurability Problem in Healthcare AI

A Standards-Based Framework for Underwriting Risk Assessment

Depending on the facts and jurisdiction, a clinical AI failure may implicate product, professional, and enterprise liability. This research paper proposes a systematic way to organize underwriting diligence across those exposures.

Research proposal
A diligence framework, not an insurance product
Composite cases
Illustrations, not named customers or engagements
No coverage promise
Insurers define terms, pricing, and claim requirements
Evidence boundary
Operational records do not prove safety or compliance

The Standards-Proof Framework

The framework asks whether evidence produced under applicable medical-device and quality standards can inform underwriting diligence, then adds healthcare-specific validation and operational evidence. It does not assume every healthcare AI system is a medical device or that standards conformance determines coverage.

L1

Foundation

Product Liability

Risk management quality across ISO 14971, IEC 62304, ISO 13485, and AI-specific extensions. Scores implementation depth, not binary compliance.

L2

Healthcare-Specific Validation

Professional Liability

Validation by risk tier. Tier 1 (clinical decision support): prospective studies with subgroup stratification. Tier 2 (documentation, prior auth): harm-pathway-matched validation. Tier 3 (administrative): harm pathway analysis before accepting the classification.

L3

Continuous Operational Assurance

All Domains (Runtime)

Pre-deployment testing plus post-deployment operational records within a defined coverage boundary. A signed record can report the named control and recorded decision while supporting integrity and signer-attribution checks; it does not by itself prove that the control was effective or that the system was safe.

Four Illustrative Composite Cases

These scenarios are generalized teaching examples from the paper, not named customers, products, policies, or insurer decisions.

Case Study 1

Patient-Facing Triage Chatbot

Vendor reported 91% appropriate routing. Stratified by time-sensitive conditions, the undertriage rate was 14%. Prompt injection overrode scope boundaries in 7% of attempts.

Case Study 2

Prior Authorization AI

Classified as “administrative” but directly affected patient access to care. Assessment found 6% racial disparity in authorization rates, embedded in decades of training data.

Case Study 3

AI-Assisted Colonoscopy

FDA 510(k) clearance based on academic centers. Deployment population was community practices. In a scoping review of 692 public FDA authorization documents from 1995 to 2023, 62 (9.0%) contained a prospective study for post-market surveillance; that is a reporting finding about the reviewed documents, not a current percentage of all authorized AI devices.

Case Study 4

Clinical Documentation AI

“95% accuracy” across 40,000 encounters/month. Stratified: ~120 notes/month contained hallucinated allergies, invented symptoms, or omitted findings, concentrated in the most complex patients.

Get the Complete White Paper

36 pages including four case studies, the complete Standards-Proof framework, parametric coverage mechanisms, and recommendations for insurers, AI companies, health systems, and clinicians.

We respect your inbox. Unsubscribe anytime.

Jennifer Shannon, MD

Jennifer Shannon, MD

Chief Medical Officer, GLACIS Technologies

UW-trained psychiatrist. Previously helped develop the first FDA-authorized AI diagnostic device for autism at Cognoa.

SG

Sarah Gebauer, MD

Validara Health

Healthcare AI risk assessment and clinical validation.