Much of the discussion about artificial intelligence (AI) in trade finance has started with what appears to be a perfectly reasonable question – what AI can do for us. Document examination has inevitably featured prominently, alongside transaction screening and the possibility that increasingly capable systems might determine whether a presentation complies with the terms of a trade transaction.
However, there is a problem with starting there because it places the technology before the process. A more useful approach may be to examine what actually happens during a trade finance transaction and then consider which technology is most appropriate for each part. Once the question is approached in that way, the role of AI becomes rather more specific and, arguably, considerably more useful.
Documentary examination illustrates the point particularly well. We tend to describe examination as though it were a single activity, yet anyone who has actually examined documents knows that it consists of numerous individual activities requiring very different levels of judgement.
Matching technology to process: Deterministic rules vs contextual AI
When a credit requires shipment no later than 10 September, and the bill of lading shows an on-board date of 12 September, there is little reason to involve artificial intelligence in concluding. Once the dates have been identified, a deterministic rule can compare them perfectly well. Much the same applies to calculations and tolerances, as well as to many other requirements for which an objectively correct result can be established.
There is no particular advantage in applying AI simply because AI is available. Yes, doing so can introduce uncertainty into something which was not uncertain in the first place.
The position changes as soon as language and context become significant. A goods description may be expressed differently across documents without necessarily creating a conflict, while a document may perform the required function despite carrying an unfamiliar title. Something that appears perfectly acceptable on its own may warrant another look because of information elsewhere in the presentation.
Problems of that nature are less easily reduced to comparisons between fields, which is where AI begins to offer something that deterministic technology cannot provide quite so readily. The two should not be regarded as competing approaches. There is no reason why one technology should be expected to examine everything when a combination can reflect much more closely the different ways in which documentary examination actually works. This also changes what we should expect from AI.
Redefining AI’s role beyond autonomous decision-making
Much of the wider debate has concentrated upon whether systems will eventually become sufficiently accurate to make autonomous decisions, but autonomy is not necessarily the most useful measure of intelligence in trade finance. Experienced practitioners know that some questions have an obvious answer while others require further consideration, and occasionally the available information simply does not justify an immediate conclusion.
Recognising those differences is itself part of expertise, meaning an effective AI-enabled system should operate within clearly understood boundaries. When an issue can be determined objectively, it can be dealt with accordingly. Where contextual analysis can assist the practitioner, technology can provide it, while questions involving genuine uncertainty can move into a process in which professional judgement becomes decisive.
This is more meaningful than simply describing the arrangement as “human-in-the-loop”. That phrase has become sufficiently broad to encompass almost any process in which someone eventually approves what a machine has done. Meaningful human involvement requires more, as the practitioner needs to understand how the conclusion was reached before deciding whether to accept it.
If a system identifies a possible discrepancy, the information underlying that conclusion should therefore remain visible. Where the analysis depends upon UCP 600 or international standard banking practice, the practitioner should be able to see the relevant material. An ICC Opinion identified as potentially relevant should be available for examination, as factual similarities do not necessarily mean the Opinion answers the transaction currently under consideration.
This moves us towards a rather different way of thinking about confidence. A system that simply returns “compliant” or “discrepant” conceals much of what an experienced practitioner would want to understand, whereas one that identifies the information it considered and explains why a particular requirement may be relevant gives the practitioner a basis upon which judgement can be exercised.
AI as a tool: Turning trade knowledge into an accessible asset
The opportunity becomes even more interesting when we move beyond individual transactions. Trade finance has accumulated an extraordinary body of knowledge. ICC rules and ISBP provide much of the foundation, supplemented by Opinions, Briefings, and other guidance. Banks have their own procedures and previous decisions, and experienced practitioners carry knowledge accumulated through years of dealing with circumstances that do not always fit neatly into written guidance.
Accessing the right part of that knowledge at the moment it is required has never been particularly easy. Anyone who has dealt with an unusual documentary question will recognise the experience of remembering that something similar arose years ago without being able immediately to recall the Opinion, guidance or previous decision in which it was addressed.
AI-assisted retrieval may eventually prove particularly valuable here. Rather than expecting AI to provide the answer, it can help the practitioner find the material from which an informed answer can be developed.
A difficult presentation could bring forward potentially relevant UCP provisions and ISBP paragraphs, together with ICC Opinions and Briefings that appear to address comparable circumstances. Subject to appropriate controls, previous internal decisions might also form part of that available knowledge. This is already being considered as part of the preliminary work on an ongoing ICC project.
Responsibility for determining whether any of it actually applies would remain with the practitioner, because retrieving information and interpreting it are not the same thing. An ICC Opinion may appear strikingly similar until a small factual distinction changes its relevance, whilst an ISBP paragraph cannot necessarily be lifted from its surrounding context and applied mechanically. Reducing the effort required to locate the material gives the practitioner more opportunity to concentrate upon understanding it.
How machine-readable data changes the equation
There is a connection here with the continuing development of structured trade data. Much of today’s AI is being asked to interpret information that was created primarily for human consumption. Documents arrive as PDFs, scans or images, and the technology first has to establish what information they contain before meaningful analysis can begin.
As trade information becomes increasingly structured and machine-readable, some activities currently considered potential applications of AI may instead become straightforward data processing. The ICC Digital Standards Initiative has highlighted the inefficiencies created by repeated manual handling of trade data and the benefits that greater standardisation can bring.\ Where reliable structured information is already available, asking an AI model to infer the same information from the visual representation of a document adds little.
This points towards an architecture in which the technology used depends upon the problem being addressed. Structured information can remove unnecessary extraction, deterministic logic can deal with requirements capable of objective comparison, whereas AI can assist where language or context makes the problem less predictable. Professional judgement remains relevant across that environment because somebody still needs to understand the transaction and determine whether the outcome makes sense.
Such an approach is less spectacular than the idea of an AI engine examining a presentation from beginning to end, but that may be an advantage. It allows automation to be introduced based on what genuinely improves the process rather than on what produces the most impressive demonstration.
It may also help address automation bias. Once a system produces convincing answers sufficiently often, there is an understandable tendency for users to gradually challenge those answers less frequently. Generative AI makes that particularly relevant because an incorrect conclusion can be expressed with precisely the same fluency as a correct one.
For trade finance, the objective should therefore be to give practitioners sufficient visibility to verify what the technology has done, rather than simply encouraging them to trust it. Sources and underlying information need to remain accessible, particularly when AI-generated analysis, rather than deterministic logic, has contributed to the outcome. Practitioners must also remain able to disagree with the system, with those disagreements providing useful information for improving future performance.
This sits comfortably with the culture that already exists within documentary examination. Practitioners have spent decades learning not merely to read documents but to question them, and there would be little sense in introducing technology in a manner which gradually discourages them from questioning its output.
Research into AI agents and knowledge work provides an interesting indication of where this relationship may eventually lead. Work at the Stanford Digital Economy Lab has examined how delegation to AI agents can shift human activity away from manual implementation towards supervisory and higher-order work. Trade finance provides an obvious environment in which such a change could occur because so much operational effort is currently consumed before professional judgement can even begin.
Circling back to the main question
Asking whether AI can check a documentary credit assumes that examination is an activity that can be handed off from a person to a machine. In practice, examination contains activities that can be calculated, information that needs to be understood in context and questions for which professional judgement remains essential. Once those elements are separated, insisting that the same technology should perform all of them becomes difficult to justify.
The more useful question, therefore, is not whether AI can automate documentary examination, but how technology can be designed around the way documentary examination actually works. Some activities can disappear into the background because there is no benefit in continuing to perform them manually, while others can be supported by AI without surrendering the judgement upon which the final decision depends.
This is where the next stage of AI in trade finance will differ from some of the early expectations. We may hear rather less about machines examining presentations autonomously and rather more about technology removing unnecessary effort while making relevant knowledge easier to find. We may also better recognise that knowing when a question requires experienced judgement is itself a valuable capability.
If that sounds rather less revolutionary than some of what has been promised, it may also be considerably closer to what trade finance actually needs.
