The missing denominator

A frequency measure needs a defined exposure population. The useful question is which population relates to the loss scenario, and whether it was measured reliably. An attested inference count is one candidate, not the only defensible exposure base for all AI insurance. STABLE

Choose the unit for the workflow

Unit	What it captures	What else is needed
Model requests / inferences	Model activity under a declared counting convention	Consequence, tools, retries, dependence, missing paths
Tool actions	Attempted or completed external actions	Authority, values, effect, approval and duplicates
Workflow decisions	Events such as a reviewed claim classification	Reliance, population, review and outcome
Transactions / affected persons	Potential severity and who is exposed	Causal path, limits, timing and aggregation
Existing proxies	Revenue, customers, business use, limits	Test whether runtime measures add useful information

This is an academy framework, not a validated rating model. STABLE

Provenance is part of the count

Identify who counted the total and attested events, what period and versions they cover, and whether direct calls or alternate tools could escape observation. Reconcile them against an appropriate independent source where feasible. A population measured by the same optional hook as its receipts cannot establish deployment-wide completeness. STABLE

A blocked attempt is neither a successful external action nor an incurred loss. Changing traffic mix can alter rates. Exposure counts alone establish neither loss probability nor severity, correlation, causation or an insurance price. STABLE

The market exercise

Propose a unit for a payment agent and a clinical summary assistant. Explain why the same inference count may relate differently to their losses, then name the data an actuary or underwriter still needs.

Next

→ The accumulation problem

Drill this page →23 bank questions stand behind what you just read. Check it while it’s warm.