Trade credit and cargo insurance have resisted AI-native risk assessment longer than any other major commercial line. This is not because the actuarial math is harder or the data is thinner. It is because the risk object — a supply chain — is fundamentally a network, and insurance has traditionally priced individual nodes rather than network topology. That conceptual mismatch is now being resolved, and the resulting opportunity is substantial.
Why Node-Level Pricing Breaks for Supply Chain Risk
Consider how a conventional cargo insurer assesses a shipment. They examine the shipper, the carrier, the cargo category, the route, and the declared value. They apply historical loss rates for the corridor. They charge a premium. This works well enough for simple single-leg cargo movements with predictable counterparties. It works poorly for modern supply chains that involve six to eight tiers of suppliers, multiple logistics providers, and conditional dependencies that are not disclosed in any single policy application.
The 2021 Suez Canal blockage made this visible at scale. The direct cargo claims were manageable. The business interruption exposure downstream — manufacturing facilities that could not source components, retailers who could not restock — was orders of magnitude larger and substantially uninsured, because trade credit and contingent business interruption policies had been written without modelling the second- and third-order dependencies of modern just-in-time supply chains.
This is not a failure of individual underwriters. It is a failure of the data model. Underwriters did not have access to the network graph of supplier dependencies. They priced the node they could see, not the network risk they were actually carrying.
What an AI-Native Approach Actually Requires
Building a supply chain risk assessment system that improves on the node-level model requires three capabilities that are genuinely novel — and genuinely difficult.
Supply chain graph construction. This means ingesting, cleaning, and structuring data from shipping manifests, bill of lading records, import/export declarations, and company registration data to infer supplier-buyer relationships at a level of granularity that traditional trade finance databases do not provide. Several data providers — Panjiva (part of S&P Global), ImportGenius, and newer entrants — provide commodity-flow data that forms the raw material. Building the inference layer that translates commodity flows into dependency graphs is where proprietary model development matters.
Real-time stress signalling. A static supplier graph is insufficiently predictive. What matters for underwriting and for early loss warning is dynamic stress signals — port congestion metrics, carrier financial health indicators, geopolitical conflict proximity scores, extreme weather event probability overlays. The challenge is not the availability of these signals; it is the integration architecture that maintains a live risk-adjusted view across a portfolio of insured supply chains simultaneously.
Counterparty credit risk layering. Trade credit insurance cannot be cleanly separated from supply chain structure risk. When a major buyer defaults, the effect propagates upstream to suppliers who are themselves insured. Building a model that correctly prices the correlation between supply chain disruption risk and trade credit risk requires data sources and model architectures that do not naturally sit together in incumbent insurer systems.
Our investment in Anansi reflects a view that the team that solves the data integration problem — that builds the canonical supply chain risk graph rather than offering yet another analytics dashboard on top of third-party data — creates the most defensible position in this space.
The Incumbent Response and Why It Has Been Slow
The major trade credit insurers — Euler Hermes (now Allianz Trade), Coface, Atradius — have enormous portfolios of trade credit risk and consequently enormous incentives to improve risk modelling. Why has AI-native supply chain risk assessment been slow to emerge from these incumbents?
Part of the answer is organisational. Trade credit underwriting at large incumbents involves deep buyer credit analyst teams who hold relationship-based knowledge that is difficult to systematise. The incentive to replace manual analyst judgement with model output is structurally complicated when the analysts are senior, experienced, and central to carrier relationships.
Part of it is data architecture legacy. Incumbent trade credit data systems track buyer payment behaviour at the aggregate level — this company paid on time or did not — but do not capture the supply chain topology that determines downstream impact when a buyer becomes insolvent or a carrier fails. Retrofitting this capability onto systems built in the 1990s is genuinely hard.
We are not saying incumbents are not trying. They are. But the architectural decisions that enable AI-native supply chain risk assessment are more naturally made at founding — when the data model, the API infrastructure, and the underwriting process can be co-designed — than retrofitted onto decades of legacy system investment.
The Parametric Angle
One resolution to the complexity of supply chain risk assessment is parametric trigger design. Rather than attempting to assess the full downstream consequence of a supply chain disruption, a parametric supply chain policy pays on the occurrence of a defined index event: port closure exceeding X days, shipping index exceeding Y threshold, declared force majeure by a named carrier. The payout is pre-agreed and automatic; no loss assessment required.
Parametric structures work well where the correlation between the index trigger and the insured's actual economic loss is high. For certain supply chain risk types — port congestion, weather-induced shipping delays, specific carrier insolvency — this correlation is strong enough that parametric structures are genuinely superior to indemnity products: faster payout, lower operational cost, no basis risk if the trigger is well-designed.
The interesting space is at the intersection of parametric design and network risk modelling — policies where the trigger is not a simple single-index threshold but a derived composite signal from the supply chain graph. This is technically feasible and commercially appealing. The teams building it are working on insurance products that did not exist in this form five years ago.
What Founders Building Here Need to Understand
Supply chain insurance is a distribution business as much as it is a technology business. The buyers of trade credit and cargo coverage are CFOs and treasury functions at mid-sized to large importers and exporters. The broker relationships that sit between those buyers and the insurance market are deeply entrenched, particularly in Scandinavia, Germany, and the Netherlands — the key European trade finance markets.
Founders who approach this space from a pure technology angle — build the model, build an API, wait for distribution to follow — consistently find that carrier licensing, broker relationships, and claims-handling credibility are longer-lead-time problems than the technical architecture. The teams with the best probability of success are those who either bring a carrier partner into the founding structure, or who have explicit carrier commitments before they attempt to scale. Data without distribution authority is an analytics product, not an insurance company. The distinction matters for valuation, regulatory treatment, and ultimately for the durability of the competitive position.