Price AI risk.
Don’t exclude it.
The difference between excluding AI and pricing it is per-deployment evidence. Verdict produces a tamper-evident, independently verifiable record of every AI agent action — the attestation primitive that turns “too opaque to price” into “priced on attested controls.”
Not a dashboard you have to trust. A cryptographic record anyone can verify — including the court, opposing counsel, and the reinsurer.
The loss curve
is already vertical.
Gallagher Re, with MIT and Testudo Global (2026). Claim mix: patent infringement 11.9%, copyright 11.2%, privacy & personal-injury 10.2%.
Courts treat AI as a tool, not a legal actor — so liability lands on the deployer, not the model vendor. Yet vendor contracts cap liability at roughly twelve months of fees with no performance warranty, and the deployer has the least visibility into how the model was built or whether it changed. The party holding the risk holds the least proof. That is the gap Verdict closes.
The market is splitting.
The split is the opportunity.
Incumbents are excluding it
AIG, Great American, and WR Berkley filed with U.S. regulators (Financial Times, Nov 2025) to exclude AI-related liabilities — WR Berkley proposed an “absolute AI exclusion” across D&O, E&O, and fiduciary lines, calling AI “too unpredictable, opaque, and difficult to price.” It is now standardized: ISO/Verisk generative-AI exclusion endorsements (CG 40 47, CG 40 48, CG 35 08) took effect January 2026, with 80%+ of carrier exclusion filings approved. The standard liability form now ships excluding AI.
Specialists are covering it
Armilla (Lloyd’s, via Chaucer) raised standalone AI-liability limits to $25M, covering generative AI and autonomous agents. Munich Re’s aiSure™ provides up to €15M performance-guarantee capacity via Mosaic. Testudo (Lloyd’s) writes up to $9.25M per insured. Coalition, Relm, and Vouch have launched affirmative AI endorsements.
The carrier that can distinguish a well-governed AI deployment from a reckless one can underwrite the first and decline the second — instead of excluding both.
It isn’t severity.
It’s correlated accumulation.
Insurers can absorb a $400–500 million loss from one company’s AI deployment, but not “a 1,000 or 10,000 losses, a systemic, correlated, aggregated risk.”
— Kevin Kalinich, Aon (global cyber lead)
No loss history
“Where these claims fit is still kind of being figured out.” Cyber took ~20 years to build usable actuarial data. AI liability is at year zero.
Black-box opacity
Insurers are “often unable to explain why an AI system generates a specific output” — so they cannot assign causation, and cannot price what they cannot attribute.
No per-insured signal
With nothing to separate a governed deployment from a reckless one, the rational response is a blanket exclusion — exactly what WR Berkley filed. Pricing requires differentiation.
And the market is already naming the missing input. Emerging underwriting requirements center on “disclosing, logging and auditing AI outputs” (IAPP). Munich Re’s aiSure already conditions coverage on passing technical due diligence. The instrument they need has not existed. It does now.
A Sealed Evidence Record
of every AI action.
Content-addressed
SHA-256 of a canonical record of the AI action — input, output, model, human-review state, citations, risk flags, timestamp. One byte changes, the hash changes. No silent edits.
Independently verifiable
Anyone re-derives the hash, the commitment, and the record id from the record itself. Zero trust in Verdict or the insured. The carrier’s own expert can verify it.
Transparency-anchored
The record commitment is posted to the public Sigstore Rekor log (Linux Foundation). A neutral, timestamped inclusion proof — proof the record existed, unchanged, at a moment.
Tamper-detecting
Post-loss, a verification check re-computes every committed value and flags any mismatch. Fraud becomes binary, not argued.
This is not slideware. Verify any record in your browser at verdict.systems/verify — change one byte and watch it fail. Seal a real, transparency-log-anchored record at verdict.systems/live-seal.
Verifiable controls
move price. They always have.
Stacked controls (MFA + EDR + backups) reach 20–50% combined. Absence is penalized: no MFA → 20–40% increase or non-renewal.
Verifiable evidence of a control is the entry ticket — not just a discount lever.
Insurers want “proof that you are carefully tending to the protection of your computing system” before discounting. AI liability has no equivalent primitive — yet.
Insurers rarely discount on a single control — so Verdict is positioned as one high-signal control in an AI-governance stack, alongside usage policies, access controls, and human-oversight attestation. The difference: AI liability today has no equivalent attestation primitive at all. Verdict is the first.
A statutory logging mandate
EU AI Act Article 12requires that high-risk AI systems “technically allow for the automatic recording of events (logs) over the lifetime of the system.” The Sealed Evidence Record is a purpose-built artifact for it.
Under the Digital Omnibus, the high-risk Annex III date was provisionally agreed (6 May 2026; Council 13 May 2026) to move to 2 December 2027 — pending formal adoption and publication in the Official Journal. The obligation is unchanged; only the date moves.
An answer to correlated risk
Verdict Edge binds a hash of the model weights into each record and refuses to seal if the loaded model does not match the attested hash.
Condition coverage on a specific, attested model — and detect a silent swap to an unvetted model at claim time. Correlated unknown-model risk becomes attested, monitored exposure. The difference between insuring a black box and insuring a box with a tamper-evident seal and a serial number.
A 90-day design pilot.
We instrument. You keep the data.
Verdict instruments 1–3 of your insured AI deployments (or your own AI underwriting and claims tooling) to seal every agent action into Sealed Evidence Records. You receive a verifiable evidence stream and a per-deployment evidentiary-completeness score. At pilot end: a joint readout on whether sealed evidence materially changes loss-adjustment cost, disputed-claim resolution time, and your confidence in pricing the risk — plus the beginnings of the loss-data series that makes AI liability price-able.
The ask is a 30-minute technical conversation. Start by verifying a record yourself — then tell us where it would have changed a claim you have already paid.
Verdict Systems is an independent company. We are not an endorsed vendor or panel partner of any carrier named on this page — Armilla, Munich Re, Testudo, Coalition, Relm, and others are referenced as market context, verified to public sources as of June 2026. Litigation-trend figures are from Gallagher Re, with MIT and Testudo Global (2026); the standardized exclusion endorsements are Verisk/ISO forms reported effective January 2026. Market figures, quotes, and regulatory dates are cited; regulatory dates reflect provisional agreements pending formal adoption. The Sigstore Rekor transparency log is operated by the Linux Foundation — not Verdict Systems. Three USPTO patent applications are filed. Nothing here is legal advice. If you underwrite AI risk and want to move from exclusion to pricing, we want to talk. Reach the founder directly →
The agent acted.
Now you can price it.
Underwriters, MGA founders, and reinsurance innovation leads evaluating a sealed-evidence layer — start with the live demo, then talk to the founder.