The case for GenAI in customs and logistics is still unclear

Artificial intelligence (AI) is replacing blockchain as the most in-demand technology solution in the logistics and customs sectors.

Unlike blockchain which is essentially a distributed ledger technology, AI covers a broad range of technologies and has been used widely in customs and logistics operations for over a decade.

Predictive AI systems, based on structured data and rules or deep learning techniques with unstructured data, have long been used to determine the best routes for delivery and to analyse x-ray scans of containers.

What is new here, and what people now refer to as AI, is generative AI (GenAI), most commonly large language models (LLMs) used to generate new content.

These models excel at processing human language and are able to provide a well-formulated answer to almost any questions, making them excellent for use in chatbots.

Current state of AI and GenAI integration

As discussed on the sidelines of the APEC Customs Business Dialogue in Incheon with service providers developing e-government solutions for trade facilitation, many agencies and officials now want AI solutions you can talk (write) to and get answers for everything they ask.

Risk management systems that would flag potentially high-risk shipments are not enough.

Users would like a risk management system that could explain in human language why a particular shipment has been flagged and be able to answer any follow-up queries they may have.

This was also the main request when we presented the first version of the Cambodia Trade Intelligence and Negotiation Advisor (TINA): Could we just type in the trade partner we have in mind and TINA would do all the required calculations and analysis and return us the answer, ideally in the form of a well-formatted report?

This would be a nice-to-have, indeed. But is this really a must-have?

How much of the project budget should be spent developing this human-like “intelligent” interface, as opposed to focusing on ensuring accuracy in the systems to flag the correct shipments for inspections or follow well-specified criteria?

These questions are particularly important for developing countries with limited resources and capacity to maintain these systems.

It is worth noting that at a recent webinar on AI for cross-border paperless trade, speakers from German Customs explained that while they were pilot testing GenAI for their online chatbot for interaction with stakeholders, they were pleased with the current “legacy” chatbot and had no immediate plans to replace it.

A key reason for this was that a rules-based chatbot was immune to the possibility of hallucinations, a state where GenAI models make up responses when faced with a paucity of data on a particular question.

It remains very difficult to fully control GenAI chatbots and solutions outputs, even when using retrieval augmented generation (RAG) and fine-tuning the underlying general purpose AI models.

To subscribe to the GovInsider bulletin, click here. 

Is GenAI necessary at all?

A recent IMF technical note provides an insightful and promising view of the current use of GenAI in tax and customs compliance applications.

At the same time, according to the MIT (Massachusetts Institute of Technology) Media Lab, 95 per cent of corporate GenAI initiatives showed zero returns and only five per cent made it to the final production with measurable value.

The MIT report reveals that real returns of AI initiatives so far have come from less glamorous areas: back-office automation, procurement, finance and operations.

This brings us back to “old AI”, machine learning and predictive models. Ensuring the models are trained on sound data and are fed high-quality data is most important and should not be compromised in any way.

The “garbage in, garbage out” principle in data analytics continues to apply and the lack of transparency on the datasets used to train and develop today’s GenAI engines means that what went into these models remains unknown.

Although unproven GenAI outputs may be more easily understood by untrained personnel, they should be approached with great deal caution at this stage.

Integrating GenAI and predictive systems may prove particularly beneficial but they come with challenges. Technical integration can be complex, with potentially different data pipelines to be established to feed both types of models and slower overall response time.

Feedback loops where GenAI responses guided by predictive outputs influence future predictions can amplify errors or biases.

Finally, GenAI is computationally expensive to train and run, adding significantly to cost and resource demands.

At present, the most feasible approach may be a superficial integration at the end-user interface level, using GenAI primarily to present the results of the underlying predictive model in natural language.

How governments should approach GenAI

The following recommendations can be made at this time. First, be very careful when considering general purpose AI engines as a foundation for critical e-government systems.

Second, prioritise the development of advanced predictive AI solutions and underlying datasets which can be validated and controlled.

Third, if there are sufficient resources and capacity, consider integrating predictive AI with GenAI, but focus initially on end-user interface-level integration, ensuring continuous supervision by skilled personnel to mitigate errors, biases, or misinterpretation.

Finally, given the great potential for AI in accelerating trade digitalisation, interested countries may consider setting up a dedicated intergovernmental working group to share experiences and develop common guidelines in this area under the Framework Agreement on Facilitation of Cross-Border Paperless Trade in Asia and the Pacific.

Yann Duval is Chief, Trade Policy and Facilitation, ESCAP, and

Rama Ha is Director, Korea Trade Network.