By: Dave Meynell
Artificial intelligence (AI) is having a rather less comfortable time in the headlines than it was a year or two ago. Questions are increasingly being asked about whether investment has moved ahead of realistic returns, whether the enormous expenditure on computing infrastructure can be justified and whether some of the expectations surrounding generative AI were simply too ambitious. Employment concerns have become more prominent as organisations move from experimentation towards deployment, whilst questions surrounding reliability, data security, intellectual property and governance have accompanied the technology almost from the outset.
Some of those concerns are entirely legitimate. The most recent research from the Stanford Digital Economy Lab, using payroll data covering millions of US workers through June 2026, found no evidence of widespread economy-wide job displacement, but it did identify a widening employment gap among younger workers in occupations particularly exposed to AI. The authors themselves are careful to describe these as early indicators rather than evidence of a simple causal relationship. That distinction is easily lost once research finds its way into headlines.
Against this background, it would be understandable if banks approached yet another conversation about AI in trade finance with a degree of scepticism. Yet much of the criticism currently directed towards AI has surprisingly little bearing on whether it can be useful within trade finance. Indeed, stripping away some of the exaggerated expectations may allow us to have a considerably more useful conversation about what the technology can actually do.
The problem begins when AI is treated as a single proposition. We seem to have moved from a period in which almost every problem was going to be solved by AI to one in which the failure to meet some of the more ambitious expectations is taken as evidence that the technology itself has been oversold. Trade finance offers a much narrower and more practical way of looking at the question because nobody needs to decide whether artificial intelligence will transform society before deciding whether it can help, for example, examine a documentary credit presentation.
What needs to be considered is the work actually being undertaken.
The document-dependent nature of trade finance
Trade finance remains heavily dependent upon documents, data and comparison. A documentary credit presentation can contain an invoice, transport document, insurance document, certificate of origin, packing list and additional certificates or statements. Information appearing in one place may need to be compared with information elsewhere, whilst the presentation as a whole has to be examined against the credit, UCP 600 and international standard banking practice.
Some of that activity requires considerable experience and judgement, but much of the preliminary work involves finding information, comparing it and identifying something that deserves closer attention. There is little benefit in requiring an experienced practitioner to spend time locating a date or checking an arithmetical calculation if technology can perform that task reliably and present the result for review.
This is where some of the discussion about AI in trade finance has perhaps gone wrong. Automated document checking has too often been presented as an attempt to reproduce the experienced document checker inside a computer, when the more credible objective is to use different technologies according to the nature of the task.
If a bill of lading shows shipment on 12 September and the credit requires shipment no later than 10 September, there is no particular reason to ask a generative AI model to reach an opinion. The relevant dates can be extracted and a deterministic rule applied. The same approach can be used for calculations, currency comparisons and many other objectively testable requirements. AI becomes more interesting where the information is less structured, where language has to be understood in context or where something within a document needs to be classified before the appropriate rule can be applied.
The distinction is important because it also changes the conversation about hallucination, one of the most frequently cited weaknesses of generative AI. A general-purpose model can generate an incorrect answer with considerable confidence, which would clearly be unacceptable if that answer were allowed, without review, to determine whether a bank honoured or refused a presentation. It does not follow that the technology has no place in the process. It means that the process should never have been designed around an unchecked answer in the first place.
A trade finance application can work within much tighter boundaries. Its analysis can be grounded in the credit and the documents actually presented, together with the applicable ICC rules, ISBP, relevant ICC Opinions and approved internal procedures. Matters capable of deterministic treatment can remain outside generative reasoning altogether, while issues involving interpretation can be brought to the attention of an experienced practitioner together with the material upon which the system has relied.
This approach is consistent with the wider direction of thinking about generative AI governance in financial institutions, where human expert oversight remains particularly important in higher-risk applications and controls need to reflect the particular characteristics of the technology.
That is very different from asking a general-purpose chatbot whether a presentation complies, and it also reflects the way experienced practitioners already work. A document checker does not approach every question in exactly the same way. Some matters are straightforward and capable of immediate determination, and others require reference to the rules, established practice or previous Opinions or Briefings. Occasionally the documents produce a situation in which the answer is not immediately apparent, and judgement becomes essential. Technology should support those distinctions rather than attempting to remove them.
Financial crime compliance through AI-driven contextual analysis
Financial crime compliance provides another useful example. Trade transactions can contain information concerning counterparties, vessels, ports, goods and jurisdictions, much of which may have significance beyond the individual document on which it appears.
ICC’s July 2026 guidance on vessel checking recognises that documentary trade transactions involving ocean shipments can give financial institutions visibility into the movement of goods which is not necessarily available in other banking products. That information can assist in identifying potential sanctions evasion, fraud and other financial crime risks, while the guidance also recognises the complexity of the task and the importance of applying proportionate, risk-based controls.
The opportunity for AI in such an environment is not to declare that a transaction is suspicious because it has detected an unusual feature. It is to help bring together information which might otherwise remain dispersed across documents and systems, allowing the practitioner to consider it in context. A vessel appearing on a bill of lading can be considered alongside external maritime information, while changes in routing or transaction behaviour may acquire greater significance when viewed against previous activity. Used carefully, technology can help direct attention towards transactions that warrant examination rather than simply generating another layer of alerts.
This becomes increasingly important as banks deal with enormous quantities of information. Automation that merely produces more exceptions is not necessarily an improvement. The value comes from making the available information easier to interpret and allowing human attention to be concentrated where it is most useful.
There is another area in which AI may prove particularly valuable, and it is one that receives rather less attention than document checking.
Knowledge access and expertise retention with AI
Trade finance has accumulated an extraordinary amount of knowledge. It exists in ICC rules, practices, Briefings and Opinions, internal procedures, previous transactions, discrepancy decisions and, perhaps most importantly, in the experience of practitioners who have spent decades dealing with unusual situations. Much of that knowledge has traditionally been difficult to retrieve at the moment it is needed. People remember that a similar issue arose several years ago but cannot necessarily remember where the answer was recorded, while experienced colleagues who once provided the answer may have moved elsewhere or retired.
AI-assisted retrieval offers a practical means of making that accumulated knowledge more accessible. A practitioner faced with an unfamiliar question can be directed towards relevant material without manually searching through years of guidance. The system need not provide the final interpretation. Simply finding the right information and placing it before somebody capable of evaluating it may already represent a substantial improvement over the present process.
This becomes more important as the industry faces the gradual loss of experienced practitioners. The answer cannot be to preserve every existing manual process simply because that was how previous generations acquired their knowledge. Equally, removing routine work without considering how future practitioners will develop expertise would create a different problem. AI can potentially assist here as well, particularly where systems explain why an issue has been identified and direct the user towards the relevant rule or practice rather than simply displaying an unexplained warning.
There is an interesting parallel with the continuing development of structured trade data. Much of today’s AI is being asked to interpret information created primarily for human consumption. Documents arrive as PDFs, scans and images, after which technology has to establish what they contain before it can begin applying any meaningful analysis.
As trade information becomes increasingly structured and machine-readable, the relationship changes. Reliable structured data reduces the amount of inference technology is required to make. Information does not need to be repeatedly discovered from the visual representation of a document if the underlying data can also be consumed directly by a system. Deterministic controls can operate upon reliable data, while AI can be reserved for those areas where its ability to work with language, context and less structured information genuinely adds something.
Seen in this way, the future of automation in trade finance is unlikely to depend upon one remarkable AI model doing everything. It is much more likely to involve structured data and established rules working alongside AI within a process in which people continue to exercise judgement. The technologies are complementary rather than competing, and understanding which is appropriate to a particular task becomes more important than attaching an AI label to the overall solution.
The current criticism of AI should therefore be useful to the industry because it forces a question that perhaps should have been asked more frequently during the period of greatest enthusiasm: what problem are we actually trying to solve?
That question becomes especially relevant when impressive demonstrations are converted into operational systems. A model may be perfectly capable of reading a document, answering a question or producing an apparently convincing analysis, yet none of those capabilities necessarily creates value unless it fits properly into the process surrounding it. The difficult part of enterprise AI is increasingly proving to be not whether a model can perform an isolated task, but whether organisations can integrate that capability into real workflows while maintaining appropriate controls.
Trade finance has seen versions of this problem before. Digitalising a poor process does not suddenly make it a good one, just as converting a paper document into a PDF does not necessarily constitute meaningful digitisation. The same discipline needs to be applied to AI. Before introducing the technology, the underlying process needs to be understood sufficiently well to determine where automation adds value, where deterministic logic is preferable and where human judgement must remain central.
This is also why governance should not be portrayed as an obstacle to innovation. Financial institutions already operate within environments built around controls, escalation and accountability. Generative AI undoubtedly introduces additional considerations because outputs can be incorrect or fabricated and because the technology can involve combinations of public and private data that create new legal, privacy and security questions. Current work on AI governance within financial services consequently places considerable emphasis upon oversight, validation, data provenance and explainability. Those requirements should make responsible deployment more credible rather than less attractive.
The changing role of human expertise in AI-augmented trade finance
Human involvement is sometimes presented as evidence that an AI system has somehow fallen short of genuine automation. In trade finance, that is an odd way of looking at it. Experienced practitioners themselves are not infallible. Documents are overlooked, information is missed, and two competent people can occasionally reach different conclusions when confronted with the same presentation. Technology can provide consistency in areas where consistency is possible, while people provide judgement where judgement is required.
The more interesting question is therefore not whether AI will replace the trade finance professional, but how the role of that professional changes when routine searching, extraction and comparison require progressively less manual effort.
There is every reason to believe that expertise becomes more valuable rather than less valuable in such an environment. Someone still has to understand whether the machine has asked the right question, whether the information upon which it has relied is appropriate and whether a conclusion makes sense in the context of the transaction. The difference is that expertise can be applied to the difficult part of the problem rather than consumed by the mechanics of finding information.
Employment impacts and the future of expertise development
The employment debate surrounding AI is relevant here because the distinction between replacement and augmentation is becoming increasingly important. Stanford’s latest work found that the weaker employment outcomes for younger workers were concentrated particularly in occupations where AI usage was more associated with substitution of human tasks, whereas employment was flat or rising where AI primarily complemented workers, especially among those with greater experience. Trade finance should take that finding seriously because the objective should not simply be to remove people from a process. It should be to reconsider which parts of that process require their expertise.
There is also a longer-term question about how that expertise is created. If junior practitioners no longer spend years carrying out the routine activities through which previous generations gradually acquired their knowledge, banks will need to think differently about professional development.
Technology can help here if it is designed to show the basis for an exception, identify the relevant rule and allow the practitioner to examine why a conclusion has been reached. A system that simply produces an answer may save time today while weakening expertise tomorrow. One that supports the practitioner in understanding the answer can contribute to both operational efficiency and education.
This is perhaps why the present change in mood around AI should not concern trade finance unduly. Some investment will undoubtedly prove excessive, and some projects will fail. Expectations will continue to be revised as organisations discover that introducing AI is considerably easier than redesigning a process around it. Legitimate questions about governance, employment and reliability will remain, as they should.
None of this requires us to abandon the technology. It requires us to become considerably more precise about why we are using it.
Trade finance does not require artificial intelligence capable of knowing everything. It requires technology capable of reading what we currently spend time reading, finding what we currently spend time searching for and carrying out comparisons which do not require years of professional experience, while allowing that experience to be concentrated on the questions where it genuinely makes a difference.
If we approach AI on that basis, the present wave of scepticism may ultimately do us a favour. It moves the conversation away from whether AI is going to revolutionise trade finance and towards the much more useful question of where it can improve trade finance in practice.
After all the predictions, demonstrations and increasingly dramatic headlines, that rather less ambitious proposition may turn out to be the one that delivers the greatest value.
