Estimated reading time: 6 minutes

If its advocates are to be believed, the technology known as Generative Artificial Intelligence (GenAI) will in due course transform nearly everything we do in the modern world today. And to an extent, since the advent of ChatGPT by OpenAI in November 2022 it already has: entire industries are now managed differently including education, coding and research.

There are as many, if not more examples of where the hype has exceeded the reality of the (lack of) value added. And of course, there have been dangers released in the latest technical revolution, such as breaches of personal privacy, generation of misinformation and malign use by bad actors and autocratic governments (not to mention the threat of the robot apocalypse).

But what about the impact on trade finance?

It’s trade finance, but not as we know it

There are many experiments and implementations of GenAI in trade finance in its very young life that have failed to deliver on the hype. And many more that have actually created unintended and impactful consequences.

The potential of large language models (LLMs) and the capacity for GenAI to reference dramatic amounts of data in milliseconds – as well as learning new ideas and self-improving via supervised or unsupervised machine learning – provides for a significant amount of upside in the trade finance space. 

However, it will be a journey of discovery where caution and baby steps will be required, rather than massive leaps forward. 

Time to clarify: What is an LLM?

Let’s start with some basics. 

While LLMs and GenAI are closely related, they are not exactly the same. GenAI encompasses any AI technique that can create new content, not just language. This includes generating images, music, code, or other creative outputs. It has a focus on creation and prediction: Its primary goal is to produce original content based on existing patterns and data.

LLMs generate and understand human language, based on massive text datasets to process, comprehend, and respond to language in various forms. Under the hood of an LLM lies a complex system, not unlike the human brain. 

At their core, LLMs are a massive neural network trained on vast amounts of text data. Think of this network as a web of interconnected nodes, each representing a concept or piece of information. 

When you feed a prompt or question to the LLM, it activates relevant nodes in this network, forming pathways based on statistical relationships learned from the data.

Data extraction from PDF documents

The industry is rightly focused on eliminating paper from trade finance operations. In reality, this has largely been achieved by banks, accelerated by COVID-19, as the majority of the industry processes transactions based on images of paper documents. 

The effort to replace the underlying paper with data points is gathering speed thanks to Herculean efforts by multilaterals, associations and advocates in the private sector leading to legislation supporting standards creation and the use of electronic documents.

Today’s reality is that most banks still process PDF images of documents. And today’s reality is that GenAI can assist existing solutions and achieve 95-100% success rates in data extraction.

Yes, including unusual documents, complex tables, handwritten information, stamps, images – and more. And yes, including invoices, possibly the least standardised trade finance and factoring document. And any shortfall in performance is quickly closed by model uptraining, as more documents are processed.

Learn, baby, learn

One of the most interesting use cases for LLMs in trade finance is education. This spans several levels. 

Commercial activities like assessing potential investors/participants in a trade distribution deal or onboarding suppliers to a payables finance facility can be sped up significantly. This is not much more complicated than inputting relatively basic prompts to a GenAI frontend today. 

Furthermore, reactive glossary-type services, when a transaction is progressing or a deal is closing, will become the norm. No more use of the “help” or phone-a-friend, information relevant to the process will be intuitively available to be consumed or ignored.

Finally, as trade finance operators’ activities are elevated to decision-making only, rather than partially processing and data entry, the ability to learn from the tech will increase. 

Automated narrative generation for discrepancies under Letters of Credit (LC) or potential financial crime risks will assist users to become better document checkers and compliance professionals – potentially the perfect on-the-job learning and the new version of classroom learning. 

Interestingly, this is where LLMs are potentially more useful: Gen AI does not necessarily care whether its output is “correct” or not. It simply reflects the majority view after assessing a large amount of data. LLM, however, is more specific to learning.

Compliance: “No-one likes it but it has to be done” … so let the machines do it!

Here is where it gets really interesting. As an example, excessive Sanctions hits in trade finance processing have long been a major issue for trade finance lenders. The number of hits generated versus the actual risk of a bank entertaining a transaction with a Sanctioned entity is a worsening problem, not one which is becoming easier to handle. 

Until GenAI. 

Counterparty risk assessment, understanding ultimate beneficial ownership, entity resolution, assessment of true dual-use goods risk and many other practical use cases will be – and are being – revolutionised by GenAI, due to the ability to search and distil information from vast amounts of sources. 

Difficult processes like KYCC/KYCS at a client level, or AML red flag assessment at a transaction level will become significantly easier and much more well-informed thanks to the application of these new techniques. 

Document examination under LC … err, not yet

As promised, it’s time for the reality check. There are certainly very real risks that GenAI impinges on privacy and ethics. Oversight of AI models that utilise GenAI is important to mitigate the risk of model bias and unsupervised learning drift.

One less worrying example of where GenAI has been proven to be less than useful is in the automation of LC document checking under the Uniform Customs and Practice for Documentary Credits (UCP) 600

When assessing a check for a potential discrepancy the large language model cannot distinguish between an established rule and what participants in discussions and debates might regard as customary practice or decisions made according to a rule.

In fact, GenAI is so far off getting this right, that there is a level of validation to the theory that machines may never learn “common sense” and thereafter self-awareness.

The time has come, but to walk, not run

Interestingly document checking against LC terms can be achieved relatively accurately by non-GenAI technology. Maybe this is the harbinger of an all-machine smackdown battle in future, as GenAI gets better at distinguishing content across billions of data points. 

But for the time being it is safe to say that, while there are plenty of constraints, errors and risks, there are many and growing numbers of use cases for GenAI and LLM in trade finance. And the transformative power is significant. 

In addition, the ability of technology to empower and assist human activity is exciting. A future where the mundane aspects of managing trade and supply chain finance are managed by bots with minimal error rates, while people learn from the AI and make decisions is closer than we think.