Skip to content

The missing denominator

The previous page claimed AI risk is unpriced because no one defined the unit of exposure. This page is the proof — half of it from the market itself, in a Lloyd’s underwriter’s own words.

Traditional liability insurance counts policies, revenues, employees, or locations as exposure proxies. None of them capture what actually generates AI liability: the inference — a single model decision in production. Every AI decision in production is a discrete exposure event. Without capturing those events “there is no denominator, no loss frequency, and no actuarially defensible price.” GROUNDED · Brochure

Exposure proxy What it prices well Why it fails for AI
Policies / limits Book size Says nothing about how often the model acts
Revenue Business scale A small vendor can serve millions of inferences
Employees Human-error exposure The “worker” here is the model, not headcount
Locations Property concentration Inference is placeless
Attested inferences Actual decision volume, per deployment, per period — this is the missing base

The brochure’s sharpest line, worth memorizing verbatim: “Every carrier writing AI risk today is pricing a fraction whose denominator is unknown. Inference Re defines and counts that denominator for the first time.” GROUNDED · Brochure (That second sentence is a tier-ii claim — attribute it to the draft when you repeat it; see the claim taxonomy.)

The strongest validation of the denominator thesis came unprompted, from a Lloyd’s underwriter (recorded in the Glacis knowledge base as “Mark”): GROUNDED · KB

“They come to me with what I call a numerator. They never come to me with a denominator, and it’s no good. You cannot price.”

He went further in both directions — endorsing exactly the trigger shape Glacis’s stack emits, and rejecting the alternatives on offer:

  • Endorsed: “If a PII filter has been triggered X amount of times … you could probably put a parametric trigger on that.” A specific, countable, measurable signal. GROUNDED · KB
  • Rejected: KPI-based assurance of the kind proposed elsewhere in the reinsurance market (“KPIs no good … basis risk”) and ISO/SOC-style certifications as a pricing basis — a certificate is a point-in-time document, not a count. GROUNDED · KB

Notice what this buys the thesis: an underwriter independently articulated the problem (no denominator), the failure mode of the competition (basis risk in vague KPIs — see parametric triggers for why basis risk kills parametrics), and the acceptable answer (a counted, attested signal).

You will use these in nearly every conversation on the floor. Both are precedents for the same move: put a sensor in the wire, and behavior becomes a rate. GROUNDED · KB

  1. Metromile (telematics). Car insurance priced per-mile, from an accelerometer in the vehicle. The sensor turned driving behavior — previously invisible to the carrier — into a parametric exposure base. The Arbiter is the accelerometer; the inference is the mile.
  2. Calmwave (hospital alarms). ML-based alarm signal processing at a Georgia health system demonstrably reduced risk — and the evidence was used to renegotiate the system’s Lloyd’s policy. Verified operational signal moved a real premium. GROUNDED · KB

Both precedents sit closer to the house than they look: Joe is friends with the CEOs of Calmwave and Metromile, and both are Glacis advisors — the accounts can be confirmed at the source. GROUNDED · Glacis

Why nobody has loss history — and why that is the point

Section titled “Why nobody has loss history — and why that is the point”

The standard objection is “you have no loss history — I can’t price this.” The structural answer: no one has AI loss history, because loss history requires a denominator and the denominator has never been counted. That is why the entry product is parametric — triggers “require no loss history and price from attestation data alone” — and why an experience base only begins once attested deployments accumulate claims. GROUNDED · Brochure The full pricing ladder is section 20; the honest version of what exists today is insurance signals.

02 The accumulation problem

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