Alioune Ciss, Chief Executive Officer
Anyone who has spent time inside a Customs administration knows the paradox. The volume and quality of data have never been better. The systems have never been more connected. And yet the gap between what officers know and what they need to know keeps widening, because the world on the other side of the border is moving faster than the systems designed to manage it. Trade routes shift overnight. Sanctions and tariff changes land without warning. Fraud schemes adapt the moment enforcement catches up with them. The volume of declarations keeps climbing while the time available to assess each one keeps shrinking.
This is not a technology problem in the conventional sense. To understand where Customs administrations now stand, and what the next step actually requires, it helps to trace the progression of how technology has transformed border management over the past three decades. Three distinct generations of capability have emerged. Each one solved a problem the previous generation could not. Each one also created new ceilings that only the next generation could break through.
The first transformation was digitization. Electronic declarations replaced paper files. Single Windows replaced the labyrinthine relay race of stamps and counters that once defined cross-border trade. Risk engines replaced manual spot-checks. Customs Management Systems created, for the first time, a single operational environment in which declaration data, trader history, and clearance status could be held together and acted upon.
The gains were substantial and real. Processing times fell. Revenue leakage from purely administrative failures, such as lost paperwork, missed deadlines, and manual transcription errors, declined. The foundation of modern Customs practice was laid.
But purely digital systems follow instructions. You tell a system: if the risk score exceeds a threshold, flag the declaration. If the HS code is among watch list, generate an alert. The system executes. It does not think, it does not learn in real time, and it certainly does not go looking for a problem you did not know to ask about. Every rule requires a software engineer to translate policy language into system code. When a tariff schedule is amended, someone has to write the update. When an enforcement priority shifts, the process of embedding that intelligence into the system takes weeks, sometimes months.
This latency between policy intent and operational execution is what practitioners have come to call the “translation gap”. In a stable regulatory environment, it is manageable. In the environment that now exists, where sanctions list update without warning, tariff schedules shift in response to political decisions taken thousands of kilometres from the border, and fraud schemes evolve continuously, it is not.
The second generation of Customs technology went further. AI-assisted tools can process large volumes of data, identify anomalies, and surface insights that a human analyst might have missed in a stack of ten thousand declarations. Machine learning models, trained on historical transaction data, can detect statistical outliers in declared values, flag mismatches between commodity descriptions and pricing patterns, and generate structured risk indicators for officer review. These tools still operate within frames that humans define, but they operate within those frames with a precision and scale that no manual process can match.
Customs valuation offers a clear illustration: Detecting undervaluation requires an officer to compare declared values against reference prices, assess the plausibility of the transaction, and make a judgement call that carries legal and revenue consequences. In a busy Customs environment, doing this rigorously across thousands of daily declarations requires structured support. Modern valuation and classification solutions make use of machine learning to analyse declared values against historical transaction data and external price databases, flagging statistical outliers and generating structured risk indicators for officer review. The ML classification tools can achieve automation rates to as high as over 95%, depending on data quality. Officers spend their time on cases that warrant professional attention, not on manually verifying routine shipments that fit established patterns.
Transit tracking offers another illustration: Using GPS and RFID technology to provide continuous real-time monitoring of transit vehicles, with automated alerts generated the moment a shipment deviates from an authorized route or stops in an unapproved location, allows border officers to be far more effective than physical escorts or intermittent manual seal checks could ever be. Such a system integrates directly with transit declarations, so the officer at the control platform is working within the same operational environment as the declaration, the electronic tracking note, and the alert simultaneously.
The results of second-generation deployments are documented and concrete. In Benin, an integrated AI-enabled environment is successfully linking Customs operations, the National Single Window, the Port Community System, valuation and classification tools, and electronic cargo tracking through a unified interface. The deployment proved the operational feasibility of real-time multi-system coordination.
Yet for all its gains, the second generation still has a fundamental ceiling. AI-assisted tools improve targeting. They cannot transform operational logic. They cannot propagate a regulatory change through connected workflows. They cannot coordinate actions across systems in response to an emerging pattern. The translation gap narrows but does not close. A risk engine grafted onto a Customs Management System, however capable, is still an add-on to fixed infrastructure. It learns from patterns it has been exposed to. It is less effective for emerging fraud typologies, where the pattern has not yet been codified into a rule.
A new threshold is now being crossed, and the transformation it implies is more fundamental than anything that came before it. Agentic AI systems do something qualitatively different from their predecessors. They pursue objectives rather than executing instructions. They correlate signals across multiple data sources without being told to. They adapt their operational logic as new patterns emerge. And crucially, they can orchestrate workflows across systems, thus pulling a transit declaration, cross-referencing it against cargo tracking data, comparing it to historical pricing from the valuation database, and routing the result to the right officer with a structured recommendation, all without a human triggering each step, and at machine speed across thousands of simultaneous transactions.
In the domain of risk management, the shift is equally significant. Traditional risk engines are calibrated to known threats: origin country, HS code, declared value relative to benchmark, trader compliance history. This works for established fraud typologies. Agentic risk systems operate differently. Rather than waiting for a human analyst to identify a new pattern and translate it into an instruction, a continuously learning system can detect anomalies in real time, behaviours that deviate from established baselines across multiple dimensions simultaneously, and surface them for investigation before they become systemic. Such a system processes transaction history, travel patterns, behavioural indicators, and cross-agency data not as isolated datasets but as correlated signals.
The WCO and the IMF have both noted something worth stating directly: reducing unnecessary human interaction at routine decision points also reduces exposure to corruption. Automation protects revenue not only by detecting fraud but by reducing the conditions under which improper facilitation occurs in the first place.
This is also where the architecture question becomes decisive. The distinction between an AI-native system and an AI-adjacent one is not semantic — and it is visible at the level of daily operations.
In an AI-adjacent environment, an officer opens a declaration, consults a separate risk scoring tool, checks a valuation module in another interface, and then makes a routing decision manually. The intelligence exists; the coordination does not. Each system has learned something. None of them talk to each other without a human in between.
In an AI-native environment, the same officer receives a pre-assembled case file: declaration data, cargo tracking status, valuation benchmark comparison, trader compliance history, and a structured recommendation — correlated across all systems before the officer touches it. The difference is not the quality of any single tool. It is whether the operational environment was built around intelligence from the start, or had intelligence added to it afterward.
Large language models embedded across every layer of operations make this possible. Customs officers and system administrators can interact with the system in natural language. Regulatory updates can be operationalized through agent-assisted workflows rather than manual recoding. The system learns continuously from live transactions and adjusts accordingly. In this model there are no developers standing between a policy decision and its execution. The translation gap, which defined the structural limitation of first-generation systems and was only partially addressed by the second generation, can finally close.
A Customs administration is not a technical process with a human element attached. It is a sovereign function involving legal interpretation, enforcement discretion, institutional accountability, and consequences that affect individual traders, entire sectors, and national revenues. The officer at the gate carries authority that cannot be delegated to a model, however capable.
What is emerging in the most thoughtful deployments is what might be called delegated operational autonomy within governed parameters. The system executes the mechanical: processing routine declarations, correlating data, routing cases, generating alerts, validating document consistency, preparing recommendations. Where ambiguity arises, where a legal question is engaged, where the circumstances are exceptional, escalation is automatic and immediate, relegated to a human. Officers become more effective, not redundant. They spend less time on tasks that do not require their judgement and more time on the ones that do. The system supports. The officer decides.
There is also the matter of explainability, which is not a secondary technical concern but an institutional necessity. Customs administrations cannot operate legitimate enforcement environments using systems with vague reasoning. A trader whose shipment is delayed must be able to understand why. An officer escalating a case must be able to articulate the basis for the flag. An administration defending a valuation adjustment must be able to demonstrate the analytical chain. Explainability is a prerequisite for institutional legitimacy, and for the private sector confidence that trade facilitation ultimately depends on.
Administrations evaluating technology investment should also ask a question that goes beyond capability: whose intelligence is it? The analytical patterns a system develops from live transaction data represent an institutional asset of considerable strategic value, and that asset should belong to the administration outright, not to the vendor that built the platform. Data ownership and the intelligence derived from it are not negotiable extras to be traded away for a lower price or a faster deployment; they are sovereign assets that must remain fully and solely in the hands of Customs. This is a governance question that deserves resolution before contracts are signed, not after.
The divergence between administrations investing in adaptive operational capability and those that are not is already happening, and it will compound over time. Most administrations are still at the stage of AI-assisted chatbots and isolated analytical tools. The administrations thinking clearly about this transition are not asking whether to adopt AI. That question is settled. They are asking how to build governance frameworks that allow intelligent systems to operate within institutional boundaries that remain human-defined; how to ensure that explainability, audit trails, and escalation mechanisms are built into the architecture rather than retrofitted; how to sequence the transition so that officers are partners in the change rather than its casualties.
The progression from purely digital systems to AI-enabled tools to agentic architectures is not simply a story of technology getting more powerful. It is a story of the relationship between policy intent and operational execution getting progressively tighter. The first generation automated the process. The second generation made the process smarter. The third generation makes the process adaptive, capable of closing, in near real time, the gap between what the administration knows and what it needs to act on.
The analogy that resonates with those who have worked in Customs long enough is straightforward. For decades, the officer standing at the gate represented the state’s entire capacity to make a decision at the border. A good officer, with experience, instinct, and knowledge of the traders, could do an extraordinary amount with limited resources. Agentic AI provides the infrastructure that ensures no officer ever has to make consequential decision without the best available intelligence behind them.
The gate is still staffed. It will be. But what happens before the container arrives, and what the officer knows when it does, is changing fundamentally.
Published on WCO News Magazine on June 30, 2026.
Useful resources
🔗 Digital addict: Leading the digital revolution in Africa
About Webb Fontaine
Established in 2002 and headquartered in Dubai, UAE, Webb Fontaine is a leading technology company specialising in Artificial Intelligence-driven solutions for global trade. With offices spanning Europe, the Middle East, Asia, and Africa, the company leverages its extensive expertise to provide governments and communities with innovative solutions that streamline trade processes and enhance efficiency.
Webb Fontaine is renowned for its pioneering technologies that help reduce trade fraud, improve Customs revenue, and expedite clearance times, supporting smoother and more profitable trading ecosystems. The company prides itself on a diverse workforce of over 700 professionals from 41 nationalities, emphasising a culture of excellence, innovation, and integrity.
The firm’s commitment to research and development is unmatched, owning the largest R&D centres in the trade sector, which are pivotal in advancing trade technology and practices. Webb Fontaine’s accolades include numerous international awards and certifications, underscoring its dedication to quality and leadership in trade facilitation.