Skip to content

The unpriced risk

Volume 2 starts where Volume 1 ended: you know what an MGA is, what a loss ratio does, and why evidence moves it. Now the harder question — why is generative-AI risk not priced at all? The draft Inference Re application answers in one sentence: GROUNDED · Brochure

“Generative AI risk is unpriced — not because the losses are uncertain, but because no one has defined the unit of exposure, measured it, controlled it, or verified it independently.”

That sentence is the thesis of this entire volume. Read it as four missing capabilities — define, measure, control, verify — because the rest of the curriculum walks them in order: sections 1718 cover define and measure, 18 covers control, and 19 covers verify.

Every alternative-risk-transfer structure — parametric cover, accumulation facility, cat bond, sidecar — shares one requirement: the risk must be observable, measurable, and contractually definable before any transfer can be built around it. Generative AI risk fails all three tests today, not because the losses are uncertain but because the infrastructure to make AI behavior observable at the point of decision does not exist. GROUNDED · Brochure

Recall from Volume 1 that pricing is frequency × severity against an exposure base (see rating and pricing). No exposure base, no frequency; no frequency, no expected loss; no expected loss, no defensible price. STABLE

The market has not waited for evidence — it is responding anyway, in three mutually contradictory ways: GROUNDED · Brochure

Response Who Why it fails
Blanket AI exclusions filed across CGL and professional-liability forms Major carriers Excludes risk their insureds demonstrably carry — and forfeits the premium for it
Point-in-time design assessments as the underwriting basis A handful of dedicated AI MGAs The assessment “expires the moment a model updates in production”
Silent carriage — AI risk embedded in cyber, E&O, and professional-liability books Everyone else Unidentified, unpriced, unreserved — invisible until claims arrive

The brochure’s summary line: the market is “simultaneously excluding and writing AI risk without the evidence to do either rationally.” GROUNDED · Brochure

Exclusions matter for a second reason you will use later in the sales motion: every insured facing a new AI exclusion is a customer for the evidence that makes them insurable again. Exclusions are the demand signal — see the motion. GROUNDED · Brochure

At the same time, product liability is shifting the burden of proof onto AI developers and deployers to show their systems “were operating within declared parameters at the time of a loss.” That evidence does not exist today. Whoever produces it first is positioned to define the standard for AI liability across both underwriting and litigation. GROUNDED · Brochure The hedge matters: “whoever produces it first” is an argument in a draft application, not an accomplished fact — the honest framing discipline is taught in the claim taxonomy.

Inference Re is described in the draft as a Lloyd’s-market alternative-risk- transfer product for generative-AI liability — “the first ART structure built on continuously verified evidence of how an AI system behaves in production,” rather than a point-in-time design assessment. It is enabled by Glacis’s Control Platform and governed by OVERT, an open runtime-evidence standard. GROUNDED · Brochure

Two hedges to carry with you whenever you repeat that:

  • “Proposed.” The source document is a draft application to Lloyd’s Lab (Cohort 17, Experimental Innovation track). Program status, dates, and deliverables are aspirations of that draft, not established facts.
  • “First.” “First” claims are attributable opinions (“the draft application argues this is the first…”), never bare assertions. STABLE

Even before any program outcome, the tier-i core holds on its own logic: the unit of exposure that generates AI liability is the inference; without an attested inference population there is no denominator, no loss frequency, and no actuarially defensible price; and evidence produced at the inference boundary is the only kind that survives a model update. GROUNDED · Brochure The next page makes that argument in an underwriter’s own words.

01 The missing denominator

Drill this page →20 bank questions stand behind what you just read. Check it while it’s warm.