AI adoption is growing throughout the manual-labour-intensive and detail-oriented field of customs and compliance. Faced with mounting complexities – like the UK’s new Border Target Operating Modelreleased in August and the upcoming EU Customs Reforms – businesses are understandably looking for new ways to manage growing workloads and the shifting web of regulatory challenges.
Supply chain and logistics companies are rushing to adopt natural language processing tools that extract, classify and process information from trade documents in an effort to accelerate clearance processes and reduce human error. While several organisations are investing heavily in AI to streamline customs clearance, if the widespread adoption of AI tools isn’t complemented with real human intelligence and expertise, it may end up creating more problems than it solves.
Trade compliance specialists are reporting an uptick in a new kind of service request: due diligence correction and, in some cases, damage control after AI-powered systems errors have led to delays and penalties at the border.
At Customs Support Group (CSG), our new survey of 200 European manufacturing firms found that 20per cent of firms were using some form of AI assistance to navigate customs compliance. While only one in 50 organisations relies on AI for classification and other critical customs activities, the data points shows a sharp drop off between use and trust. There is a significant confidence gap whereby firms are not saying ‘AI doesn’t work’; they are saying ‘AI isn’t yet safe enough to own the outcome.’
The problem lies in AI systems’ tendency to hallucinate, giving answers that look plausible but aren’t rooted in reality – and right now, these tools are struggling to keep pace with rapidly changing cross-border rules and regulations. Without human oversight – what we call Real Intelligence – an over-reliance on AI can be an undeniable risk.
The understandable appeal of AI customs tools
Generative AI tools built on large language models (LLMs) like OpenAI’s ChatGPT are often marketed as analytical tools. Feed a new customs and compliance framework into an AI model and it will not only be able to summarise key points and changes, but also complete compliance documentation tasks.
Cross-border regulatory restructuring is being driven by several factors, from Brexit and US tariffs to other shifts like the end of de minimis duty loopholes. Importers are understandably hungry for anything which means these new documentation requirements and tighter scrutiny on international shipments don’t translate into delays, higher labour costs and other potential threats to the bottom line.
We at CSG, along with other leaders in our field, recognise this potential and have begun introducing the technology to speed up the customs and compliance process, cutting through complexity and keeping pace with demand.
AI customs tools offer some compelling use cases, from data extraction and analysis to HS code assignment and preference validation. In house, we use Customs SmartAssist technology, aiding the scalability of complex declaration processing through machine learning and optical character recognition (OCR) to process documents accurately in various formats and qualities. This is already delivering significant improvements in processing efficiency.
AI’s promise of being able to accurately navigate regulatory frameworks to automate many of the traditional, manual steps in the customs declaration process has undeniable appeal. Leaving these tools to do so unsupervised, however, is something they just aren’t ready to do thanks to one key reason: hallucinations.
An AI hallucination is a “plausible but false” output generated by language models. For example, customs professionals at the International Meat Trade Association (IMTA) asked a large language model whether it was legal for the UK to export chicken feet to China. It’s not, but that didn’t stop the model from responding with an enthusiastic “yes”.
Why do AI models do this? There are multiple reasons why large language models hallucinate, but there are two primary and compounding culprits.
One, hallucinations occur when the data used to train a model contains contradictions, is incomplete or just flat out wrong. AI companies subscribe to a philosophy of AI development built on scaling laws. Basically, the bigger the model, the more data you feed it, and the more chips you run it on, the better the results.
Leaders in AI research believe that scaling laws are the key to unlocking AGI — an AI that is superior in meaningful ways to a human brain. It’s why AI companies can’t afford to be as discerning about where they find their data as they used to be. This appetite for massive datasets creates understandable issues when it faces the need for training data to be high quality, accurate and free from unsavoury material.
The second issue is that, fundamentally, a large language model is a prediction machine. Probability, not accuracy, is the end-goal of these deep learning models. When there’s insufficient consensus in the training data, an LLM’s goal is to fill in the blanks with something that looks plausible rather than accurate. This is why expert oversight is a non-negotiable.
AI errors and the importance of ‘Real Intelligence’
Our 2026 Strategic Radar Customer Survey revealed that businesses in Europe are taking a cautious approach to AI implementation in customs compliance relative to many other sectors. Adoption figures suggest gradual experimentation rather than rapid adoption, reinforcing the view that classification accuracy, accountability and regulatory defensibility continue to outweigh the allure of automation.
Nevertheless, there is a continued push from companies offering AI customs tools that risks misaligning industry perceptions of AI’s capabilities with reality.
While there is an undeniable appetite – even a need – for AI in customs and logistics to boost efficiencies and reduce manual work, the fact is that AI is no substitute for human expertise. That’s not to suggest AI has nothing to offer the customs and compliance space, or that organisations are irresponsible for embracing it. But engaging with AI needs to be done responsibly. Clear, direct and active phrasing, paired with sensible output checks, are key to capturing the benefits while mitigating the risks.
Hallucinations aren’t a deal-breaker, but rather something to be managed. As with any new technology, success will be rooted in a balanced, pragmatic approach. Right now, placing too much trust in AI tools to handle complex, high-risk compliance work risks fines, delays and disruption. AI tools may one day be ready to handle these tasks alone, but for the moment, human expertise and diligence – real intelligence, not the artificial kind – are still the foundation of a smooth, hassle-free customs strategy.
