Using artificial intelligence in export controls

Export controls, contrary to what the term suggests, are coordinated among very different departments within a company. This includes in particular the areas sales, product development, and purchasing.

It is not uncommon for employees in such departments to only perform export control tasks in addition to their core duties. Time constraints also make it difficult to provide new employees with the in-depth training and ongoing professional development required here.

The result: Employees often lack the necessary in-depth expertise, and the regular changes to regulations make it difficult to always stay up to date. Similar questions therefore arise in all these various departments and often end up on the desk of the person responsible for export controls.

Custom GPT solutions can effectively support export controls under these conditions. The idea is to feed the questions into the custom GPT first, then consult the experts only for thornier issues. This saves time and frees up resources.

Multilingualism and documentation are further advantages: LLMs can handle a wide variety of languages, effectively removing language barriers. They also document all communications, providing verification of due diligence measures in the organization of export controls.

Effectively supporting such use cases was our motivation for developing our own customized LLM. For about a year now, we’ve been training a custom GPT specifically to answer questions relating to export controls. The results we have achieved are surprisingly good and, compared to what we know from practice, they eliminate many time-consuming and persistent errors.

How good the contextual responses of any LLM are depends on the quality of the stored data and what’s known as prompt engineering – the method of giving the LLM instructions to generate a specific output. Prompt engineering nudges the LLM to provide the expected answers, and the right prompting is critical to getting the desired answers.

One of the difficulties is learning how to formulate prompts to achieve the desired results. We noticed that in legal contexts such as export controls, it’s important to tell the LLM that it is dealing with a compliance check that:

• must strictly adhere to the legal text,
• must not deviate from the prescribed scheme,
• should be concise, and
• may entail liability risks.

In our own self-trained LLM, we have – in addition to these basics – also stored decision test trees and boilerplate text modules that our custom GPT must consult to provide good answers to specific queries.

Well done. This is the right approach and correct answer.

This case study clearly demonstrates the performance of our self-trained custom GPT in the direct comparison with ChatGPT: A check with the custom GPT yields the correct result, while ChatGPT in the example above produced an incorrect result. The custom GPT guides the user through the relevant steps, asks specific questions, and helps consider all aspects relevant to the check.

But despite the predominantly positive experience with our self-trained custom GPT, there are also challenges. Consistently accurate results are not yet guaranteed. Incorrect output still needs to be monitored and corrected to achieve steadily more reliable results.

It is important to note that specialized AI models cannot replace the subject-matter experts in the company. The final decisions must always be made by human experts who possess the necessary expertise and experience.

A custom GPT should act as a guide and support the specialists by accompanying them step by step through the process of running an export control check, pointing out any special circumstances and asking specific questions to produce a precise and complete solution. The idea is to make day-to-day work easier, but it’s always just a tool, not something that ultimately makes a decision.

The use of AI in export controls offers a promising opportunity to support companies in a complex and constantly evolving legal domain. Custom GPTs can enhance the quality and efficiency of export controls.

The continued development of our own specialized AI solution, presented in this case study, illustrates the potential of AI in this field. The integration of other specialized domains will make this custom GPT into a tool that helps companies master the challenges of export controls with greater speed and precision.

We’ll keep you posted on our progress. Sign up to our newsletter to stay in the loop.