Effective January 1, 2026, the insurance market began withdrawing from generative artificial intelligence risk. The withdrawal was not announced in a speech or projected in a white paper. It arrived as policy language. Insurance Services Office endorsement CG 40 47 removes bodily injury, property damage, and personal and advertising injury arising out of generative AI from the standard Commercial General Liability form. W. R. Berkley Form PC 51380 is an absolute artificial intelligence exclusion written for the directors and officers, errors and omissions, and fiduciary liability lines. The first form reaches the institution. The second reaches its executives personally.

Six months on, the withdrawal is no longer early. Major carriers including W. R. Berkley, Chubb, Travelers, and Berkshire Hathaway have filed to adopt the endorsements or proprietary exclusion language of their own, state regulators have approved more than 80% of the filings submitted, and one industry projection puts eventual adoption at 95% of carriers on the commercial general liability line. The operative phrase in the exclusion, arising out of, is read broadly in insurance law. It does not require the AI to be the direct cause of a loss, only a link in the chain. A clinical recommendation that a generative system shaped anywhere in its path can be enough.

Where the exclusions apply, the health system holds its clinical AI exposure itself: unindemnified, and in most deployments, unowned. That configuration is the subject of this paper. An excluded risk is survivable when someone is named to own it, and an owned risk is insurable when the governance behind it can be documented. A risk that is neither covered nor owned sits at the exact point in the workflow where the accountability gap already lives, which means the coverage gap and The Accountability Gap™ (TAG™) now occupy the same institutional space. The exclusions did not create that space. They priced it.

This paper states what the two instruments exclude and whom each one reaches, why the underwriters moved before the institutions did, where the excluded risk now sits inside the clinical workflow, and what a renewal now requires: two seats named, six functions held, the Handoff documented, and a Gap Score™ the institution can put in front of an underwriter.

The two forms, and whom each one reaches

The two instruments are not parallel. They exclude different things, on different lines, and they land on different people. Reading them together is what reveals the shape of the exposure.

CG 40 47 reaches the institution. The endorsement attaches to the standard Commercial General Liability form, the base layer of a health system’s casualty program, and removes bodily injury, property damage, and personal and advertising injury arising out of generative artificial intelligence. It is the broad member of a modular family: CG 40 48 removes only the personal and advertising injury coverage, and CG 35 08 performs the same exclusion inside the products and completed operations coverage part. A carrier selects among them according to appetite. What the family shares is the trigger language. Arising out of does not require the AI to have caused the loss; it requires the AI to appear in the chain that produced it. A deterioration alert that fired late, a sepsis model that shaped a recommendation a clinician acted on, a generative summary that omitted the finding a lawsuit later turns on: under the broad form, each is arguably enough to place the claim outside the policy. When the exclusion applies, the loss lands on the institution’s own balance sheet.

PC 51380 reaches the executives personally. The W. R. Berkley form is an absolute artificial intelligence exclusion written for the directors and officers, errors and omissions, and fiduciary liability lines. Those lines exist for one purpose: they protect the individuals who make institutional decisions when those decisions are challenged. Directors and officers coverage is what stands between a named executive and personal exposure in a shareholder action, a regulatory proceeding, or a negligence claim aimed at the decision to deploy. An absolute AI exclusion on that line means the executive who signed the clinical AI charter, certified the deployment to the board, and answered the underwriter’s renewal questionnaire may do all of it without the indemnification those answers assumed. The Governance Owner holds Cover, the function that carries institutional responsibility when the deployment is challenged. PC 51380 is what happens when the institution’s own coverage for that seat is withdrawn: the person who holds Cover discovers that nothing covers them.

The asymmetry maps precisely onto the two seats of the Named Owner Principle. The CGL exclusion prices the institutional exposure, the exposure the Governance Owner accepts on the institution’s behalf at Commission. The Berkley form prices the personal exposure of holding that seat at all. A Decision Owner at the bedside remains protected, for now, by a medical professional liability line the exclusions have not yet reached. The seat the market moved on first is the one in the corner office.

One further feature of the withdrawal deserves stating plainly, because it defeats the most common institutional response. There is no single market position to check against. Some carriers attach the broad form at every renewal. Some adopt the narrow variants. Some have drafted proprietary exclusions that reach further than the standard language. No two renewals treat the exposure identically, which means the question a General Counsel must answer is not whether the market has excluded AI. It is which exclusion is on this institution’s schedule, on which lines, effective when. That answer exists only in the endorsement schedule of the policies the institution actually holds, and at most institutions, nobody has been named to read it.

Why the underwriters moved first

The exclusions are sometimes read as an overreaction to an unproven risk. The reading has it backwards. Underwriters are not futurists; they are readers of loss evidence, and the evidence was published before the forms were filed.

On Stanford’s MedAgentBench, run inside a realistic electronic health record environment under strict first-attempt scoring, the strongest clinical AI agents were wrong on close to one task in three. The NOHARM benchmark, released in January 2026 by the ARISE Network, measured the sharper number: across 31 large language models, the potential for severe harm from a model’s medical recommendation occurred in up to 22.2% of cases, and 76.6% of those harmful errors were errors of omission, the model failing to raise the diagnosis, the red flag, the necessary next step. Omission is the failure mode a confident output conceals and the one a coverage dispute later reconstructs. And the systems cannot reliably show their work: research published in Nature Communications found that between 50% and 90% of large language model medical responses were not fully supported by the sources the models themselves cited. A recommendation that cannot be traced to its basis cannot be defended after the fact, by a clinician, by an institution, or by a carrier deciding whether to pay.

Set that evidence beside what the institutions themselves report. In a Black Book Research survey of 182 United States hospital leaders, only 22% expressed high confidence that they could produce a complete, auditable AI explanation for regulators or payers within 30 days, and 33% named unclear internal ownership as a top barrier to audit readiness. An underwriter reading those two literatures together sees a system that fails silently one time in three, deployed by institutions that by their own account cannot document what happened or say who owns it. Pricing that risk is not alarmism. It is the profession. The market priced it the only way an unmeasurable exposure can be priced: excluded.

The institutions read the same evidence and drew a slower conclusion. The difference is structural. An underwriter who misprices risk loses money within a policy cycle; an institution that deploys with thin governance may not learn it for years, until an outcome is questioned and the reconstruction begins. The exclusions are best understood as the fast reader’s verdict on the slow reader’s posture. The verdict is not that clinical AI is uninsurable. It is that ungoverned clinical AI is uninsurable, and nobody had yet shown the market the difference.

One space, two gaps

An exclusion removes coverage from a category of loss. To know what the institution now holds, the category has to be located inside the workflow, and it locates precisely.

The place is the Shaped Decision. Of the three decision patterns a clinical AI deployment produces, a Clinician Decision made on clinical judgment alone, a Parallel Decision where clinician and system reached the same conclusion independently, and a Shaped Decision where an AI recommendation influenced the call, only the third carries the exposure the exclusions describe. It is the pattern where generative AI sits unambiguously in the chain of events, which is all the arising out of language requires. It is also, in most deployments, the pattern the record cannot isolate: all three decisions produce the same order in the chart, and without documentation built to distinguish them, every AI-adjacent decision becomes arguably a Shaped Decision after the fact. The exclusion’s breadth and the record’s silence compound each other. A carrier can contend the AI was in the chain; an institution that never captured the recommendation, the response, or the reasoning cannot demonstrate that it was not.

The Shaped Decision lives at a known address. It is Layer 4 of the governance cascade, Clinical AI Governance, the layer that protects the patient decision at the bedside, and it is the layer most institutions have not built. The four layers of the cascade do not sit in sequence only; they converge, and the convergence point is where all four must meet for the patient decision to hold. When the clinical layer has no named owner, the center of that convergence is exposed. The Accountability Gap™ (TAG™) lives in the exposed center. So, as of January 1, does the excluded risk. The two gaps are not analogous. They are co-located: the same decision, the same layer, the same undocumented junction where an AI-shaped recommendation reached a clinician and no named owner traveled with it.

The co-location is what makes the present configuration the worst available. An insured risk with no named owner is a governance failure the balance sheet absorbs. An uninsured risk with a named owner is an exposure someone is accountable for managing down. An uninsured risk with no named owner is neither absorbed nor managed: when the outcome is questioned, no policy responds, and no one answers. The chart supplies the one name it always contains, the clinician who acted, and the reconstruction works backward from the bedside through a junction nobody documented toward an institutional decision nobody currently owns, with the institution’s own counsel discovering the coverage position and the accountability position in the same review. Each gap was survivable alone. Institutions ran uninsured risks with clear owners long before AI, and insured risks with muddled governance. What January 1 ended was the ability to hold both gaps at once and call the result a tolerable posture.

There is one more property of the co-location, and it is the paper’s hinge: a single act closes both gaps, because both gaps have the same cause. The coverage gap is not, at bottom, a pricing dispute; the market excluded what it could not see governed. The accountability gap is not a staffing oversight; it is the absence of the named structure that governance requires. Name the two seats, activate the six functions, document the Handoff between the owners, and the institution has simultaneously built the accountability structure TAG™ requires and produced the governance evidence the market prices coverage on. What closes the gap is what reopens the coverage. That is the work of the final section.

What renewal now requires

Renewal used to be a procurement event. For a health system running clinical AI, it is now an audit, and the institution’s side of the audit must be built before the questionnaire arrives.

The market has already specified what it wants to see, in two forms. The first is the underwriting file. Across the carriers now willing to touch AI exposure, the questions cluster on the same controls: an inventory of where AI touches decisions, a human accountable in the loop, documented capacity to intervene and shut down, data provenance, and, on nearly every list, a named accountable executive for AI. The market is asking, in its own vocabulary, whether the Governance Owner seat is filled. The second form is the new coverage itself. An affirmative AI insurance market has begun to form precisely where the standard forms withdrew, and its defining feature is that it grants and prices coverage only on documented governance evidence. Coverage for AI risk has not disappeared. It has become conditional on exactly the artifacts most institutions have never produced.

Producing them is not a bespoke project, because the artifacts the underwriter requires and the artifacts TAG™ requires are the same artifacts. Before the next renewal cycle, the sequence is the one the architecture already defines. Name both seats for every clinical AI deployment: a Governance Owner who holds Charter, Commission, and Cover, and a Decision Owner, the treating clinician, who holds Decide, Document, and Defend. Document the Handoff between them, the junction where an AI-shaped recommendation leaves the system the institution authorized and reaches the clinician who must act, so that the record shows a named owner traveling with the information rather than information arriving alone. Build the record so the three decisions can be told apart, because an institution that can demonstrate which decisions the AI shaped, which it merely echoed, and which it never touched has converted the exclusion’s broadest weapon, the arising out of ambiguity, into a bounded and arguable question. Then score the deployment. The Gap Score™ rates the three layers that must converge, AI Governance, Clinical AI Governance, and Decision Accountability, each from 1 to 5. A total above 12 is defensible: the owners are named, the decision is traceable, the trail holds. A total below 9 is unpriced liability, and the phrase can now be read literally. Below 9, the institution is carrying an exposure the market has declined to price, on the market’s own evidence that no one could show it what was being insured.

The score’s function at renewal is worth stating exactly. A Gap Score™ is not a certificate the underwriter has asked for by name, and the paper does not predict that carriers will adopt the instrument. Its function is prior: it is how the institution knows, before the questionnaire arrives, whether it has a governance story or a governance gap, deployment by deployment, in a form leadership can read. The institution that scores above 12 walks into renewal holding the named owners, the documented junction, and the distinguishable decisions that the underwriting questions are groping toward. The institution that has never scored walks in holding assurances. The market has already told everyone which of the two it is willing to price.

None of this guarantees coverage, and the sequence is not a negotiation tactic. It is the same work the institution owes the patient, the board, and the regulator, now carrying an additional and unusually prompt payer of attention. The underwriter is simply the first party with the standing to ask the accountability question annually and the ability to attach a price to the answer. The exclusions made the question unavoidable. The architecture makes it answerable.

What this paper is not

Not coverage or legal advice. The exclusions described here are policy forms whose application turns on the endorsement schedule, the jurisdiction, and the facts of a claim. What an institution’s program actually covers is a question for its broker and counsel, read against the policies it holds, not against a paper. This paper’s subject is the accountability architecture the market’s withdrawal has exposed, and that architecture is the institution’s to build regardless of how any single claim resolves.

Not an argument against deployment. The evidence above describes ungoverned deployment, and the exclusions price ungoverned deployment. Nothing here counsels a health system to halt clinical AI; the clinical case for these systems is real and, in places, urgent. The counsel is narrower and harder: deploy with the two seats named, the six functions held, and the Handoff documented, because the alternative is no longer a covered risk. It is an unpriced one.

Not solved by vendor indemnification. A vendor can warrant the model, indemnify the contract, and stand behind its product in a dispute between companies. The vendor’s indemnity does not restore the institution’s excluded coverage, does not respond on the executive’s directors and officers line, and does not place a name in the institution’s own audit trail. The exclusion operates on the institution’s policies; the vendor’s promise operates on the vendor’s. Contractual recourse after a loss is not insurance before one, and it has never yet filled either seat.

Not a prediction that the market stays closed. Affirmative AI coverage is forming, carve-backs are being negotiated, and the exclusions themselves may narrow under litigation and regulatory pressure. The market position described here is the position as of this writing, and it will move. What is unlikely to move is the direction of the underwriting question, because it is the same question the regulator, the plaintiff’s counsel, and the board all ask: who owned this. Institutions that can answer it will find coverage first, on the best terms, whatever the forms then say.

@techreport{johnson2026unpriced,
  author      = {Johnson, Mo},
  title       = {Unpriced: What the AI Exclusions Mean for the Institutions Holding the Risk},
  institution = {GPe Research Publications},
  type        = {Position Paper},
  number      = {№02},
  year        = {2026},
  month       = {7},
  url         = {https://publications.gperesearch.com/papers/unpriced}
}