Insurance & Underwriting
AI risk is being carved out of standard lines faster than affirmative cover can form, because the market lacks the thing underwriting runs on: an independent record of what the system was allowed to do, and evidence of what it actually did. Triodian produces that record, a prevention layer the operator cannot bypass, a per-action attestation stream, and a certified bound on how often a violation gets through. Infrastructure for the carriers building this class, not a policy we sell.
Placing rather than writing the risk? Brokers & Placement → · Reading this for a treaty? Reinsurers & Accumulation →
Why now
On 2026 industry reporting, carriers have introduced absolute AI exclusions across D&O, E&O and cyber; ISO exclusion forms exist; standalone AI liability has appeared with limits reported between $2M and $50M; a Lloyd's-backed AI MGA began underwriting in early 2026; and practitioner consensus expects a cyber-like standalone class within five to ten years.
Exclusions are how a new class announces itself, cyber began the same way. What follows exclusions is questionnaire underwriting, and what ends questionnaire underwriting is loss experience. Cyber took fifteen years to reach telemetry-based pricing. AI will not get fifteen years: its losses are faster and more correlated, and from August 2026 the EU AI Act requires per-decision event logging on high-risk systems. The market will go looking for a verifiable per-action substrate. This page describes the one that exists.
US generative-AI lawsuits, 2021–2025, on Gallagher Re / MIT reporting.
Limits on early standalone AI liability, 2026 reporting.
EU AI Act high-risk per-decision logging obligations operational.
What an underwriter can rate on
A distribution-free ceiling on how often a violation gets through, at a stated confidence. Read it as a loss-frequency bound, with calibration monitored and re-certification forced on model change, measured shift, or calendar, the policy-condition machinery already built.
Ships now · Rules tierEvery governed action emits a signed token: rule version, constraint class, verdict. In aggregate it is telematics for decisions, a continuous exposure record priced the way motor books price the black box.
Established statisticsConcentration, correlation-load and margin-erosion across a stream of individually-compliant decisions: the early-warning layer for the accumulation problem, before it is on anyone's book.
The peril is gated before loss, by a control below the operator. Cyber insurance monitors its peril; this architecture can refuse it, including a bad model update, which is this class's systemic event. Prevention-led books run the loss ratios prevention-led books have always run.
The test
Your actuaries run a one-page protocol on a corpus you choose, inside your walls. We never see the data. The certified bound holds the realised miss-rate at or below your stated α, or you have falsified the frequency measure this class would be rated on, before a dollar of capacity is committed.
Run the protocol →The first rateable class
Award interpretation is deterministic, Rules-tier, an authored, versioned constraint family, precisely the certified action class the first product attaches to. The loss history is a matter of public record.
Recovered by the Fair Work Ombudsman for 249,000+ workers, 2024–25.
Across five years of recoveries; ~60% from large corporate employers.
Per-event severity, public, at the top of the range.
A burning cost from public records, a frequency bound that is certified rather than asserted, and an exposure count from the token stream, every element of the rating basis has a source you can check without trusting us.
The roadmap
Product-attached, modest limits, carrier-backed, no MGA required.
Endorsements to PI / D&O / E&O, premium credits rated on the attestation stream.
Systemic model-update events carried by reinsurance; FM Global structure as reference model.
Counterparty durability
The IP is held by an independent Foundation, beyond any single owner, so a frontier lab cannot acquire the company and shelve the control layer a book of business depends on. Builder and guardian are held in different hands, on a decades-horizon class.
Why a Foundation →The precedent
The last time an uninsurable technology became an insurable one, insurers didn't wait for the state. Hartford, 1866: a sealed valve below the operator, a certificate that travelled with the machine, and the rule that made the market, no inspection, no coverage. The company it built is still operating inside Munich Re, and is publicly working this exact market today.
The modern form of no inspection, no coverage is a coverage boundary drawn at the governance boundary: attested actions inside cover, everything routed around the valve outside it. Nobody mandates the appliance. The policy does.
Read the precedent →What we're looking for
A carrier or reinsurer to co-design a first product, a remediation-cost warranty on certified action classes, rated against Australia's public underpayment loss history.
An MGA partner for telemetry-priced endorsements as token volume accumulates.
Actuarial and broker conversations now, structured around blind-validation results rather than a prospectus.
The honest limit
What is insurable maps to the ladder. Certified action classes today are the Rules and Distribution tiers, deterministic constraints and calibrated bounds. The Meaning tier joins when its pre-registered experiment passes, not before, and any policy schedule built on this architecture should say the same.
Correlated model-update loss is this class's hard problem; the valve's pre-loss gating is our answer to it, and reinsurers should test that answer rather than take it.
Triodian Pty Ltd is not an insurer, does not hold an Australian Financial Services Licence, and does not issue, arrange, or advise on insurance products. The pages in this section describe technical infrastructure and invite partnership discussions with licensed carriers, reinsurers and underwriting agencies.