AI case study
National Customs Targeting Centre for Cargo, Directorate General of Analytics and Risk Management of India
As cross-border e-commerce drives rapid growth in low-value courier and postal consignments, the National Customs Targeting Centre for Cargo is piloting machine learning across anomaly detection, pattern recognition, and X-ray image analytics with the aim of helping Indian Customs target fraud, smuggling and revenue risks with more precision, while continuing to facilitate legitimate trade.
The Problem
The National Customs Targeting Centre for Cargo, or NCTC-Cargo, is part of the Directorate General of Analytics and Risk Management (DGARM), which was established in 2017 for data analytics and risk management within India's Central Board of Indirect Taxes and Customs (CBIC). Based in Mumbai, it is responsible for the national risk management system (RMS) that profiles and assesses shipments by risk. Its remit spans express courier, postal and cargo channels. The courier and postal channels alone accounted for approximately 23 million declarations (import and export) in the 2025-2026 financial year, many of which were low-value parcels filed by individual consignees rather than registered commercial entities.
The RMS was first developed as a static, rule-based system supplemented by manual intelligence inputs from field officers. Declarations that met predefined rule conditions were selected for examination and some selected for rule-based post clearance audit. The rule-based approach was able to deliver consistent and automated selection.
However, as courier and postal volumes grew, several limitations began to surface. First, the rules required periodic manual revision and could not adapt quickly to new forms of concealment. Second, each declaration was largely treated in isolation, so the rules did not identify previously unseen patterns, and cross-declaration signals such as the reuse of a single identity document across unrelated consignments went undetected. Third, entity-based profiling was weaker when declarations were filed by individuals rather than firms, limiting the value of conventional trader-level risk scoring. Fourth, exhaustive manual review of millions of low-value declarations was not feasible, leaving little room for risk-based prioritization beyond what static rules could capture.
These channels have also been repeatedly identified as conduits for narcotics, undervalued high-value goods, and mis-declared items, while separate analytical exercises had identified recurring patterns of systemic duty misclassification. When undetected, prohibited goods enter circulation and duty and tax revenue is lost. Moreover, broad rule-based selection subjects a larger share of legitimate consignments to scrutiny than required.
The Solution
NCTC-Cargo is undertaking a pilot programme to develop an AI-assisted analytical layer to extend the existing RMS's capabilities. The AI layer combines the established rule-based indicators with machine learning anomaly detection and pattern recognition, so that known risk logic is retained alongside the new methods. This capability will be deployed to support four related initiatives.
- The first gives a score for each express courier and postal declaration for risk at the time of filing, before assessment, flagging signals such as identity-document reuse across different recipients, valuation anomalies against benchmarks for similar goods, inconsistencies between goods descriptions and declared classification, and implausible or incomplete addresses.
- The second applies the same pattern-recognition approach to pre-arrival cargo manifest data, generating risk scores before goods physically arrive so that examination resources can be positioned in advance.
- The third embeds an image-classification engine within the X-ray scanners used at courier and postal terminals, scoring scanned images in real time and escalating only the subset flagged as high-risk for analyst review.
- The fourth, which remains at a conceptual and developmental stage, would extend similar image analytics to containerized sea cargo, consistent with the broader policy direction towards scanning of containerized imports at gateway ports.
As of July 2026, the programme was at the pilot and proof-of-concept stage. Image-based risk scoring at scanning points has progressed furthest towards operational use, pattern-based analytics for courier and postal declarations are being piloted against live operational data, and no component has yet reached nationwide deployment.
Across the first two initiatives, AI is used to produce a risk score for courier, postal and cargo declarations alongside an explanation of how the score was derived. In the image-screening component, AI flags whenever certain objects of interest are detected for officer review and further action. Candidate risk rules and anomaly thresholds are governed by a “propose-then-confirm” discipline: model-generated indicators are reviewed and approved by responsible officers before they are applied operationally, and periodic systemic-risk reports are signed off at the Additional Director General level before onward submission.
The AI system draws on declaration and manifest data already filed through Indian Customs electronic data interchange and the existing RMS. This is supplemented for address-verification analytics by publicly available post office location and addressing data held by the national postal administration, which feeds its outputs into the same operational workflow. The underlying data structures for courier, postal and manifest filings are aligned with internationally recognized customs data concepts, consistent with the direction of the World Customs Organization data model. The AI components currently operate domestically and are not yet integrated with cross-border data-exchange mechanisms, which NCTC-Cargo describes as a longer-term consideration as the programme matures.
The solution was designed and built in-house by customs officers with domain expertise working alongside an in-house technical team, without external commercial procurement, and funded from the department's existing administrative and technology budget. Because the underlying declaration data are confidential, the system was deliberately built to run within a controlled departmental environment on existing infrastructure, without external data transmission or cloud services, so the principal cost is generated by personnel time rather than recurring licensing.
Overview
| AI type | Machine learning, combination of rule-based, supervised, unsupervised anomaly detection, pattern recognition, computer vision |
|---|---|
| Training data | Customs declarations, cargo manifests, courier and postal data, X-ray images, identity and address reference data, and additional data (e.g. VAT, past offences) |
| Platform | AI layer over national RMS |
| Core functionalities | Piloted capabilities: declaration fraud and anomaly detection, pre-arrival cargo risk scoring, X-ray image analytics and systematic risk analysis on non-revenue risk, security risks and environmental risks from narcotic drug or psychotropic substances, weapons, hazardous materials, prohibited items, dual-use items |
| Primary users | NCTC-Cargo targeting officers and field customs officers; senior management and policy units |
Challenges
These challenges were managed through data normalization, an unsupervised analytical design, a human-in-the-loop model for all enforcement decisions, and incremental validation against real-world data.
Data quality and standardization
A large share of declaration fields is free text without standardized validation, producing inconsistent representations of the same information; a deliberate normalization step preceding any pattern detection is required.
Scarcity of labelled outcomes
Confirmed examination results were too sparse to support conventional supervised learning, leading to a combination of supervised, unsupervised and deterministic rule-based design that uses the limited confirmed cases for calibration rather than training.
Mindset shift
Officers accustomed to rule-based selection had to accept pattern-based scoring as a complement to, rather than a replacement for, experienced judgement.
Best practices
Hybrid risk model
Combining existing rules with anomaly detection preserved institutional knowledge while identifying new, previously unseen risks.
In-house, domain-led development
Officers with direct operational knowledge of courier, postal and cargo fraud shaped the risk logic, keeping expertise embedded and data under departmental control.
Capability built through hands-on involvement
Officers responsible for analysis and targeting developed familiarity with anomaly detection, pattern recognition, and the interpretation of model-generated explanations. This was enabled through direct involvement in designing and validating the system, rather than through a separate training curriculum. Formal training for field officers on acting upon system-generated alerts is identified as a necessary step ahead of wider deployment.
Phased, evidence-based piloting
Components are validated against historical and live operational data before wider rollout, allowing thresholds to be calibrated and confidence to build incrementally.
Confidentiality by design
Processing within a controlled departmental environment, without external transmission, addresses the sensitivity of trade and taxpayer data.
The impact
The programme is at pilot and proof-of-concept stage, so comprehensive before-and-after performance figures are not yet available, and internal seizure and revenue figures are not disclosed in this submission. The impact observed to date is qualitative, drawn from validation against historical and live operational data, with formal, organization-wide measurement of throughput, processing time and accuracy, planned as part of the transition to wider deployment.
Stronger risk management and enforcement
Validation exercises have shown that anomaly detection and pattern recognition find cases that the prior rule-based system would not have identified, in particular network-level signals such as the reuse of a single identity document across superficially unconnected declarations. Officers have reported that applying pattern-based analysis to previously unstructured, free-text declaration data made such relationships visible in a way that the manual review of individual declarations could not.
A related enforcement benefit emerged during the pilot beyond the original focus on individual consignments: the same data and methods can identify broad patterns of duty misclassification across large cohorts of historical declarations. Officers are now able to submit periodic reports to senior management on systemic revenue-risk patterns, a capability that fell outside the original scope of consignment-level risk scoring which will now support more targeted compliance action and stronger revenue protection.
Early indications of productivity gains in targeting and examination
The image-scoring component has demonstrated the feasibility of automatically pre-filtering X-ray images, reducing the proportion of total image volume that requires direct human review while preserving coverage of high-risk cases. Risk-based prioritization of declarations is similarly designed to concentrate officer time on the highest-risk cases rather than spreading it evenly across all traffic; it is expected to allow low-risk consignments to be facilitated more quickly, although this effect has not yet been measured. Targeting officers involved in the pilot have reported greater confidence and speed in prioritizing cases, since each warning flag is accompanied by an explained rationale that is easier to justify and act upon than an unexplained score.
Begin with a clear, honest assessment of the quality, structure and completeness of existing operational data, and of the realistic availability of confirmed outcome labels, before selecting an analytical method. The appropriate AI approach should be determined by what the available data genuinely support, not by replicating a method used elsewhere without first confirming that the underlying data conditions are comparable.
Glossary
| Computer vision | AI that enables computers to interpret and analyse images and videos. |
|---|---|
| Machine learning | The system learns from past data to make predictions and take decisions. |
Footnote
This case study is based on NCTC-Cargo's official written submission, supplemented by publicly available CBIC and DGARM materials.
← Back to Electronic Commerce and Digital Technologies
Created in September 2026