Accumulation and product design

AI can create common dependencies across otherwise different insureds. A model update, tool supplier, identity service, retrieval corpus, deployment channel, or shared policy defect may affect several workflows at once. The exposure question is which organisations depend on the same component for a consequential action. STABLE

Count something meaningful

An inference count measures model activity. A tool-action count measures attempted or completed actions under a defined convention. A decision count measures a workflow event. Affected persons, transaction values, permissions, and duration can add severity context. No single unit is automatically the right exposure base for every product. STABLE

Define the unit, the eligible population, the measurement source, and how it connects to the loss scenario. Then test whether the signal adds information beyond existing proxies. “More receipts” is not synonymous with more risk, and blocked attempts are not losses. STABLE

Fictional accumulation exercise

Three vendors use the same model, but one drafts low-stakes copy, one produces treatment suggestions, and one can release payments. A shared update is a common dependency; the likely loss mechanism and severity differ. Group by dependency and use, authority, controls, and reliance. STABLE

Keep the product ladder honest

Development step	Evidence needed
Scoped measurement	Defined events, integrity, capture scope, missingness
Useful market signal	Reviewer utility, operational feasibility, exposure relevance
Loss relationship	Suitable incident and loss data, severity, alternatives, uncertainty
Carrier-owned product design	Appetite, wording, legal review, claims handling, capital and pricing
Portfolio or risk transfer structure	Dependency model, correlated severity, event definition, aggregation and capacity

These are development questions, not a claim that each step has been completed. STABLE

Parametric design must address payment when a defined signal occurs without the expected loss, and loss without the signal: basis risk. A trigger still needs probability and severity analysis, reliable measurement, manipulation controls, an outage rule, and agreed dispute handling. Runtime records alone do not establish a price. STABLE

For indemnity, the policy’s claims and coverage terms remain central. For alternative risk transfer, the insurer or authorised product owner retains product and risk authority. Glacis contributes the defined control/evidence capability; a receipt does not itself create insurance. STABLE

Field drill

An identical update affects several insureds. What would you request before estimating accumulation? Name the shared dependency, deployed versions, business use, permitted actions, populations, timing, controls, and plausible losses, then state the missing data.

Continue

→ Underwriting fundamentals · Carrier economics

Drill this page →68 bank questions stand behind what you just read. Check it while it’s warm.