The Unit of Analysis Is Wrong

When a customs agency deploys AI, it almost always does so at the level of the declaration. Each import record is scored for risk. Consignments above a threshold are flagged for examination. Officers inspect, seize, or clear. The model is retrained on outcomes. This is a sensible architecture for the problem as it is typically framed — find the risky shipment in the volume — and it has materially improved prohibited goods detection at borders that have invested in it.

The problem is that most systematic customs fraud is not a shipment-level phenomenon. It is a trader-level phenomenon. And the AI built to score individual declarations is structurally blind to patterns that only become visible when you look across a trading relationship over time.

Consider the mechanics of customs undervaluation — the most prevalent form of customs duty evasion, estimated to cost governments between 1–3% of theoretical import duty revenue annually. A trader who undervalues a commodity by 15% does not do so on one shipment and stop. They do so consistently, across every shipment, because the economics are straightforward: a 15% undervaluation on a 10% duty rate is a 1.5% cost saving per shipment, compounding indefinitely until detected. No single declaration looks anomalous. The declared value is plausible. The goods are genuine. The documentation is complete. The risk score is moderate at worst. The fraud is invisible at the declaration level and obvious at the relationship level — where 200 shipments of the same commodity from the same origin show values that sit systematically below every comparable market benchmark.

This is not a subtle distinction. It is the difference between the question AI is being asked — is this declaration suspicious? — and the question that would actually close the revenue gap — is this trader engaged in systematic fraud?

What Systematic Fraud Looks Like Across the Relationship

The undervaluation example is the most economically significant, but it is not the only customs fraud type that lives in the relationship rather than the declaration.

Origin fraud — declaring goods as originating from a country that qualifies for preferential tariff treatment when they in fact originate elsewhere — is rarely a one-time event. Circumventing a tariff through false origin requires an established supply chain arrangement, supplier documentation, and consistent misrepresentation across multiple shipments. A single declaration claiming preferential origin treatment from a qualifying country is unremarkable. A trader whose entire import history in a particular commodity consists of preferential origin claims, in a commodity category where the declared origin is implausible given the volumes involved, is a different matter entirely. The signal requires the history.

AEO abuse is the customs fraud type with the highest leverage and the most underdeveloped detection capability. Authorised Economic Operator status grants expedited clearance — precisely the condition that creates the maximum opportunity for fraud. AEO certification is awarded based on a compliance assessment conducted at a point in time. The trader’s compliance posture may have changed materially since certification. Ownership may have changed. Business model may have shifted. The volume and nature of imports may look nothing like what was assessed at certification. Most customs agencies monitor AEO compliance through periodic renewal cycles rather than continuous behavioural assessment. The result is that the expedited clearance channel — the highest-trust channel in the customs system — is monitored least frequently. Longitudinal compliance monitoring that tracks AEO trader behaviour against their certification baseline detects drift before it becomes systemic abuse rather than after the scheme has run for years.

Carousel and VAT fraud is the network expression of the same insight taken to its logical endpoint. Individual entities in a carousel transaction chain — the importer, the domestic seller, the exporter, the missing trader — can each appear entirely legitimate in isolation. The fraud is constructed from the relationships between them: the transaction flows, the timing, the reclaim submissions, the missing VAT remittances. No per-declaration model detects it because no individual declaration contains the evidence. The evidence is the network — the pattern of relationships across entities and transactions that reveals the scheme structure. Network analysis of transaction chains does for carousel fraud what longitudinal trader profiling does for undervaluation: it makes the fraud visible at the level where the fraud actually exists.

Why the Declaration-Level Approach Persists

The persistence of declaration-level AI is not irrational. It reflects genuine constraints that are worth acknowledging before explaining why they need to be overcome.

Declaration data is clean, structured, and already flows through customs processing systems. Each declaration is a discrete event with a defined set of fields, a clear outcome — cleared, queried, seized — and a feedback loop that makes model training tractable. Building a risk score on a single declaration is a well-understood machine learning problem with an accessible training dataset.

Longitudinal trader intelligence requires more. It requires connecting declarations to traders consistently over time — which sounds simple but involves entity resolution challenges when traders use different legal entities, agents, or declaration formats across shipments. It requires constructing commodity-level price benchmarks across origins, routes, and time periods that can be compared against declared values at scale. It requires maintaining trader compliance history in a form that is queryable in real time at the point of declaration processing. And it requires the analytical infrastructure to detect drift from baseline behaviour — a trader whose compliance posture is gradually shifting rather than suddenly collapsing.

These are harder problems than per-declaration scoring. They require data architecture investment that does not produce immediate, visible results. The per-declaration model generates seizures that appear in performance reports. The longitudinal intelligence model generates the understanding of systematic risk that makes the right seizures possible — a less visible output, even when the revenue impact is larger.

The Data Infrastructure Implication

The shift from declaration intelligence to trader intelligence has a specific infrastructure consequence that most customs agencies have not fully confronted: the relevant unit of data storage and retrieval needs to change.

Current customs data systems are typically optimised for declaration processing — receiving, validating, risk-scoring, and clearing individual records at high volume and low latency. The data lives in transaction tables, organised by declaration date and reference number. This architecture is well-suited to per-declaration AI. It is poorly suited to longitudinal analysis because answering the question “what does this trader’s full import history look like over the past three years, by commodity, by origin, by declared value?” requires aggregation across large volumes of historical records that declaration-processing systems were not built to serve efficiently.

Building trader intelligence capability therefore requires a parallel data layer — a trader profile store that maintains aggregated, continuously updated commercial history for each known trader, updated with each new declaration and queryable at declaration processing speed. This is not a novel architecture. It is the same pattern that retail banks use to maintain customer behavioural profiles for fraud detection and that health plans use to maintain member risk profiles for care management. In both cases, the insight was identical: the relevant signal lives in the accumulated history, not the individual event. The investment required to make that history queryable at transaction speed is the foundational infrastructure challenge.

Where This Changes the Investment Priority

The practical consequence of the relationship frame is a reordering of customs AI investment priorities that differs meaningfully from the declaration-centric view.

Undervaluation analytics — statistical benchmarking of declared values against market prices and historical transaction data for comparable commodity-origin combinations — becomes a higher priority than it currently receives in most customs AI portfolios. Undervaluation is the dominant revenue fraud type and the one most directly addressed by trader history analysis. A benchmarking model that flags a trader whose declared values for a specific commodity consistently sit 18% below comparable market prices is generating the most economically significant customs intelligence available, at a fraction of the cost of physical inspection programmes that will never surface this pattern.

AEO compliance monitoring — continuous behavioural assessment of certified traders against their certification baseline, tracking changes in import patterns, commodity mix, origin profiles, and declaration accuracy — becomes a distinct AI investment rather than a periodic manual review. The leverage is significant: AEO traders clear with minimal scrutiny, which means that detecting compliance drift early has both a direct revenue benefit and a systemic integrity benefit for the entire expedited clearance programme.

Network analysis of transaction chains — entity resolution and graph-based analysis of relationships between importers, exporters, agents, freight forwarders, and financial counterparties — becomes the primary tool for carousel and VAT fraud detection rather than per-declaration anomaly scoring. The fraud is in the network. The analysis needs to be at the network level.

The Recommendation

Build longitudinal trader intelligence as a distinct capability layer alongside per-declaration risk scoring — and evaluate each against the fraud type it is actually designed to address.

Per-declaration models are the right tool for prohibited goods detection, where the threat is a one-time event with physical evidence. They are the wrong tool for revenue fraud, where the evidence accumulates across dozens or hundreds of transactions and is only visible in aggregate. Deploying declaration-level models against systematic undervaluation is not just ineffective — it is producing the false assurance that the revenue problem is being managed when it is not.

The three investments that follow from the relationship frame are sequenced deliberately. First, build the trader profile data layer that makes longitudinal analysis possible at declaration processing speed — this is the infrastructure prerequisite for everything else. Second, deploy undervaluation benchmarking models against the existing declaration history to establish the baseline assessment of where systematic revenue fraud is concentrated in the current import population. Third, restructure AEO compliance monitoring from periodic renewal review to continuous behavioural surveillance, using the same trader profile infrastructure.

Each of these investments addresses a different point in the same underlying problem: systematic customs fraud that has been structured to look legitimate one declaration at a time. Seeing it requires looking across all of them at once.

The declaration is the event. The trading relationship is where the evidence lives.