Part of the AI, Future and War series. Analysis and hypothetical examples are identified in the text.
Prediction is not the first problem to solve
An insurer can have an impressive geopolitical forecast and still misunderstand its own exposure. Locations may be incomplete, businesses may share critical suppliers and several insured interests may depend on the same infrastructure. Before asking AI to estimate the likelihood of a conflict event, the insurer should establish what could be affected together.
UNCTAD's maritime research illustrates how disruption can spread through transport networks. This article draws a broader analytical lesson for insurance: business connections can create common exposure even when insured assets are physically separated. That is an inference for risk analysis, not an insurance coverage statement from UNCTAD. UNCTAD: Review of Maritime Transport 2025
Separate the event from the contract
An economic loss and a covered insurance claim are not interchangeable. A scenario may include delayed deliveries, damaged property and lost revenue, but the treatment of each depends on the applicable wording and facts. AI can support document organisation; it should not silently turn an uncertain interpretation into an assumed payment.
A proposed exposure table would keep gross economic impact, potentially insured loss and unresolved coverage questions in separate fields. Legal and underwriting specialists should review material ambiguities. Avoid describing a general war-risk model as if it establishes cover or replaces the claims process.
Build scenarios around shared dependencies
Consider a hypothetical portfolio containing a port operator, a nearby warehouse and several import-dependent businesses. A common disruption could affect them through different channels and at different times. Counting each policy as independent would miss the connection; assuming every policy reaches its maximum loss would ignore meaningful differences.
Scenario analysis should identify the common event, the transmission channel and the basis for each estimate. Dependencies may include power, transport, communications and named counterparties. Where information is sparse, use clearly marked assumptions and ranges rather than presenting an uncalibrated probability as measured fact.
Assess what AI actually improves
Useful applications include finding inconsistent entity names, extracting relevant locations for review and identifying missing records. These tasks can improve the exposure inventory without making strong claims about predicting war. Their performance should be measured against independently checked samples.
NIST's voluntary AI framework offers a general reference for governing and evaluating AI use. In this context, a practical application is to document intended use, material errors and the authority retained by reviewers. The framework itself does not certify a risk model or validate its financial outputs. NIST: AI Risk Management Framework
Keep liquidity visible
A loss scenario also has a timing dimension. Claims payments, premium receipts and reinsurance recoveries need not occur together. A risk review should therefore examine the cash-flow pattern as well as the ultimate loss estimate, while avoiding assumptions that recoveries are immediate or uncontested.
The best near-term contribution of AI to war-risk insurance may be a more complete, auditable view of exposure. Prediction can be examined after data quality, contractual uncertainty and accumulation are understood. This article describes a public analytical approach; it does not disclose proprietary Quantica methods, claim a validated predictive capability or offer insurance terms.