Proceed with caution

Artificial Intelligence (AI) offers significant potential to support trade compliance, but it is far from infallible and must be used with caution. As everyone in this field knows, even a minor error on a trade document – whether a health certificate or a bill of lading – can have costly and sometimes catastrophic consequences for a supply chain. Such mistakes can lead to lost sales, damaged contracts, reputational harm and strained customer relationships.

Many companies are currently experimenting with AI to understand how it can support their operations. There are already promising applications, particularly in the digitalisation of trade documents. However, in the trade compliance world, it is essential that those testing AI tools also understand the risks – especially the phenomenon known as ‘AI hallucination’. An AI hallucination is where an AI tool gives you the wrong answer, but it does so very confidently, and it is not obvious that it is telling you a porky pie.

Over the past two years, we have been testing AI’s reliability by posing certain trade-related questions. One example is: “Can the UK export chicken feet to China?” We know the answer is no, as the UK does not yet have approval. Yet until very recently, Microsoft Copilot consistently responded with a confident “Yes you can” – a clear and unequivocal error.

It is important to remember that AI tools typically rely on publicly available information. Where public sources are incomplete or ambiguous, AI may draw incorrect conclusions. When we tested this question more recently, the tool referenced a Freedom of Information response from UK authorities, which provided a detailed and accurate explanation. Because this source is reliable and publicly accessible, the answer has now improved. This underlines the importance of checking the references AI tools provide to ensure you can trust the information. Relying on an incorrect answer could have serious and expensive consequences.

AI will undoubtedly improve over time, but users must still conduct their own due diligence. AI cannot access information held only by Government officials, and the nuance required to interpret legislation or compliance requirements often depends on direct clarification from authorities.

On numerous occasions, we have tested Copilot to locate specific legislative articles. At present, nine times out of ten it provides an article that has no relevance to the question – yet it presents the answer with complete confidence. AI systems also tend to favour black-and-white interpretations. Black-and-white answers are good because they are clear… but very bad when they are wrong, and unfortunately, this still seems the case too often when we test AI tools!

Where AI does show real promise is in the digitalisation of trade documents. IMTA is participating in an HMRC-led working group aimed at increasing the use of electronic documents and improving digital information flows within supply chains. However, because this work is led by HMRC, health certificates – which fall under Defra’s remit – are not currently a central focus. IMTA has been advocating for the digitisation of Export Health Certificates for more than a decade, and we had hoped the Border Target Operating Model would have taken this further. We already have QR codes and electronic PDF health certificates for EU imports, and this approach was expected to be extended to imports from the rest of the world, but this has yet to materialise.