The legacy insurance tower — cyber, professional indemnity, general liability, product liability, directors-and-officers — was assembled, layer by layer, for a world in which decisions were made by people and machines did as they were told. That world is not the world we are pricing.
In the past three years, every major specialty carrier has either added an affirmative artificial-intelligence exclusion to its forms or, more quietly, declared the existing form silent on AI risk and left the silence to be litigated. Cyber carriers, having watched ransomware swallow a decade of their underwriting margins, now decline coverage for losses caused by an autonomous decision the insured's own system made. Professional-indemnity carriers exclude services delivered by or substantially assisted by a model. Product-liability carriers wait for the European Union's revised Product Liability Directive to define software as a product, and then exclude it. The result is a gap. Castra exists to underwrite that gap.
§ IWhat we underwrite.
We underwrite four lines, deliberately.
Autonomous Systems Liability — third-party bodily-injury and property-damage cover for organisations operating systems that act in the physical world without a human in the immediate decision loop. Counter-uncrewed-aerial-systems, autonomous warehouse robotics, autonomous vehicles, surgical robotics, autonomous industrial control.
AI Agent Errors and Omissions — first- and third-party cover for organisations whose agents act in the digital world: brokered transactions, regulated recommendations, automated underwriting and adjudication, retrieval-augmented assistants embedded in regulated workflows. Misclassification, mispricing, wrongful denial, regulatory non-compliance.
Model Performance Warranty — a first-party warranty that a model's measured behaviour will remain inside an agreed envelope, anchored to a continuous evaluation contract. We pay if the envelope is breached. This is not insurance against general harm; it is a performance bond against a defined statistical event.
Regulatory Defense — defence cost cover for an insured's response to a competent-authority inquiry, audit, or enforcement action arising from an AI system. The administrative penalty itself is not insurable. The litigation it triggers is.
The four lines are deliberately narrow. We do not write personal-lines AI. We do not write cyber. We do not write the underlying tower the AI cover sits beside; we write the AI carve-out the tower has now produced.
§ IIWhy specialty, not general.
A general carrier prices on the law of large numbers. A million policyholders, a stationary distribution, a quarter-century of loss data, and an actuarial estimate of next year's frequency and severity. For AI risk, the conditions of that calculation do not obtain. The portfolio is small. The distribution is non-stationary. The loss data is sparse and arrives in lumps. The peril is the model itself, and the model is updated more often than the policy is endorsed.
A specialty managing general agent — sized to the line, with technical specialisation behind it — can price on a different basis: telemetry of the actual deployed system, an evaluation contract that defines what the model is supposed to do, and a continuous adjustment of premium against drift. A general carrier cannot afford to do this for one in ten thousand of its policyholders. Specialty practice is built for that one.
§ IIIWhy continuous, not annual.
The traditional annual policy was assembled to price a risk that does not change between renewals. A factory, with the same machinery, the same workforce, the same fire-suppression system, twelve months later. AI deployments do not have that property. A model fine-tuned on Wednesday is a materially different risk on Friday. A new dataset added to a retrieval index changes the deployed behaviour overnight. A vendor-supplied dependency updates and the model's classification surface shifts.
We price the bind. We re-price the deployment quarterly against a telemetry contract the insured agrees to at submission. The methodology is described separately in Issue No. 003 of this Quarterly.
The discipline is old. The risk is new. The form is, of necessity, both.
§ IVWhat we deliberately do not write.
We do not write open-ended general AI cover. We do not write coverage for systems whose mode of operation is wholly undefined by the insured at bind. We do not write coverage where the insured will not provide the telemetry contract. We do not write coverage for prohibited practices under the EU AI Act, the corresponding U.S. state statutes, or the equivalent in any operative jurisdiction. We do not write coverage for systems whose deliberate purpose is causing harm.
These exclusions are not marketing. They are how the line stays priceable.
§ VThe capacity behind the binder.
Castra Risk Underwriters LLC is a managing general agent. We do not balance-sheet the risks we bind; we place them with specialty reinsurance paper rated A or better by A. M. Best, S&P, or the equivalent national rating agency at the seat of the placement. Counterparty names are disclosed at quote, subject to the customary placement-confidentiality undertakings.
We share loss data, catastrophe analytics, and the portfolio concentration model with our treaty partners quarterly. The discipline is mutual; it is the only basis on which a treaty can be priced.
§ VIThe Quarterly.
This Quarterly is the place we explain what we did and why. Issues will be technical: Anatomy of an Autonomous Loss, Underwriting the Out-of-Distribution, The EU AI Act and Your Insurance Tower, Modeling Correlated Autonomy Failure. Each issue is written by the underwriting practice and reviewed before publication. We make mistakes and we correct them; corrections are appended to the issue, not buried.
We do not write thought-leadership. We write working notes.
§ VIIThe standing offer.
If an organisation is deploying AI in production and finds that its existing tower has either excluded the risk or refused to confirm its position, write to us at the address below. We will tell you what we can offer, what we cannot, and what we would need to see to underwrite. Submissions take six business days to bind once telemetry is in place.
That is the disciplined infrastructure built ahead of the advance. The Roman legions did not improvise the camp; we do not improvise the form. Imperium per disciplinam.