A major container port processes millions of consignments annually. Physical inspection of each one is not operationally possible — the infrastructure, the officer time, and the trade disruption required would be prohibitive. Customs agencies operate at inspection rates of 2 to 5 percent of total consignment volumes at most major ports. Everything that clears without inspection either does so because it has been assessed as low-risk or because the volume made assessment impossible. The distinction between those two categories is the central problem of cargo risk targeting.
The quality of the targeting decision determines the security and revenue outcome of the entire customs operation. A model that achieves a 20 percent inspection yield — meaning one in five physically inspected consignments produces a seizure or enforcement finding — is producing five times the enforcement impact per inspector-day as a model achieving a 4 percent yield. For the same physical inspection resource, the difference between a 5 percent yield and a 20 percent yield is the difference between a programme that is managing risk and one that is primarily creating delays for legitimate traders while the fraud and prohibited goods pass through.
Where the targeting decision breaks down
Cargo risk targeting is currently based on a combination of rules profiles and officer judgment. Risk profiles are built around known high-risk trade routes, commodity types, shipper histories, and declaration characteristics. These profiles are effective at identifying consignments that match known risk patterns. They are less effective at identifying consignments from traders who have adapted their shipping patterns to avoid the profiles, commodities misdescribed to fall outside the flagged categories, and routes and origin combinations that have changed following enforcement actions elsewhere.
Static rules profiles are inherently backward-looking. They encode the patterns that were identified in prior enforcement and intelligence activity. The smuggler or duty evader who has read the enforcement environment and structured their operations to avoid those patterns will not appear in the rule-based targeting output. The AI model that identifies statistical anomalies in current trade patterns — consignments that deviate from the established behaviour of a specific shipper, route, or commodity category in ways that are not explained by legitimate trade variation — identifies emerging patterns that rules cannot.
The data available for cargo risk targeting has expanded significantly with mandatory advance declaration requirements. Container shipping data, shipper history, consignee profiles, transit patterns, and commodity classification data all arrive before the vessel. The question is whether the agency’s targeting infrastructure can use that data at the volume and speed required to produce a targeting decision before the consignment arrives.
The two objectives that the same model serves
Cargo risk targeting is simultaneously a revenue protection function and a security function, and the targeting model that addresses both is essentially the same. A consignment that appears anomalous — a declared value inconsistent with market prices for the commodity, a shipper profile that has not previously traded in the declared commodity category, a route that is atypical for the declared origin — is a candidate for physical examination for both duty evasion and prohibited goods. The features that predict duty fraud and the features that predict prohibited goods often overlap: unusual trade patterns, opaque supply chains, first-time traders, and declaration anomalies are risk indicators for both.
The risk scoring model should produce a combined risk score that integrates both revenue and security signals, with the ability to surface the specific indicators that drove the score for officer use. A targeting system that produces a risk score without supporting evidence gives the officer a directive to inspect without the information needed to focus the inspection effectively. A system that identifies the specific anomalies — this shipper has never previously declared this commodity category, the declared value is 40 percent below the median for this commodity from this origin — equips the officer to conduct a targeted examination rather than a general search.
The trade facilitation dimension
The case for better targeting is not only made on enforcement grounds. Every unnecessary physical examination of a legitimate low-risk consignment costs the importer time and money, creates port congestion, and damages the agency’s relationship with the compliant trader community that the authorised economic operator programme is designed to reward. A targeting model that concentrates examination on the highest-risk consignments produces both more enforcement per inspection and less friction for legitimate trade. The two objectives are aligned, not in tension.
The inspection yield rate is the single metric that captures both dimensions simultaneously. A higher yield rate means more enforcement per examination — the targeting is more precise. It also means fewer examinations of legitimate consignments — the false positive rate is lower. A programme that tracks inspection yield and uses it to evaluate and improve the targeting model is simultaneously improving its enforcement impact and its trade facilitation performance.
The technology dimension
Advance cargo declaration data for major trading nations is processed through customs management systems that at many large revenue and customs authorities run on IBM Z. The risk scoring model that evaluates each declaration against historical shipper and commodity data, benchmarks declared values, and identifies anomalous trade patterns can run on IBM Z via IBM Machine Learning for z/OS, scoring each declaration against the full historical data estate within the advance notification window. The targeting decision is available to port officers before the vessel arrives, allowing examination resources to be pre-positioned for the flagged consignments.
What success looks like
The primary metrics are inspection yield rate — separated into duty evasion yield and prohibited goods yield — trade facilitation rate for AEO and low-risk traders, and the proportion of total consignment volume covered by automated risk scoring. The last metric matters because manual or partial coverage creates the same unscreened population problem in cargo targeting as in passenger screening. A targeting system that scores the full advance declaration population, not just the declarations that match known profiles, is addressing the coverage problem that static rules cannot solve.
