Droit and Artificial Intelligence: Selecting the right tool for the problem at hand

By: Leith Dennis

As artificial intelligence becomes an increasingly hot topic and a popular tool to solve certain business problems efficiently, we are frequently asked about our usage of this technology and how it compares to our Adept platform.

The discussion of artificial intelligence (AI) today most frequently refers to one of the following three types of algorithms: unsupervised learning, supervised learning, and large language models (or generative AI).

• Unsupervised learning takes a collection of information and splits it into categories or groups. This approach can be great for inferring outliers or exceptions, though it can often be hard to tell why something is an exception.
• Supervised learning aims to find an answer to a specific question based on a training set. As an example within a space familiar to Droit, one could imagine leveraging supervised learning to infer if a transaction should be reported to a regulator based on a training set of previously reported transactions. Here as well, the explanation of why something should or should not be reported is many times a “black box” or statistical in nature.
• Lastly, LLMs (or Generative AI) predict the text output that fits a scenario or answers a question. This approach is great at composing a reasonable sounding answer as to why, e.g., something complies or does not comply with a regulation. However, the challenge of “hallucinations”, or invented responses not based in fact, makes relying on those answers tricky at times. Even in instances when answers are more reliable, the tool does not provide any transparency into its decision making.

These three technologies are powerful tools with myriad applications. But none of them is a great fit for the automation of decision processes around regulatory compliance. Within this space, the ability to audit a decision to understand precisely why an action was taken, without ambiguity, is paramount. And these types of AI lack the required level of transparency. That lack of transparency may impede an institution’s – and particularly legal and compliance teams’ – ability to develop a level of comfort and confidence in the behavior of a decision process. Furthermore, any wrong decisions taken by such decision processes due to statistical outliers or unexplainable logic may play poorly during regulatory audit.

In comparison, Droit Adept is a knowledge-based approach to decision making, providing formal answers using logical reasoning. Droit’s Adept Decision Service evaluates provided input facts against a defined knowledge model and thus mathematically guarantees decision making which is explainable and repeatable. The construction of these models is driven by a process we refer to as Knowledge Engineering.

The decision space of regulations governing capital markets and financial services is incredibly complex. A knowledge model compiled from various rule sources will not have a precise success definition. Take, for example, the definition of a “swap” under U.S. law. While this concept is defined within Title 7 of the U.S. Code, the full richness of the rules for determining whether a given product is a “swap” extends across hundreds of pages of law, regulations, rules, and determinations. So how would we come up with a knowledge model? What conditions or rules should be modeled as formal logic? Which definitions and concepts should correspond to input facts to that decision? The answer to these questions requires thoughtfulness and a level of understanding regarding the context in which the definition is operating. Determining the best approach – and understanding where to draw these lines – is something, currently, only a human can do.

This responsibility sits with Droit’s Knowledge Engineering team – a group of specially trained individuals responsible for the creation of these knowledge models. Our team constructs formal logic, which operationalizes regulatory requirements, and defines an API for how calling systems will interact with that decision process. In construction of these logical models, our Knowledge Engineers capture explicit attribution to a digitized version of the text – establishing an unbreakable link between that version of the text (viewable within our Digital Library) and the corresponding version of the logic which implements it (viewable within our Logic Viewer). That logic is then deployed within our Decision Service, allowing institutions to make real-time compliance decisions with respect to complex global regulations. The process of reading & analyzing regulation, defining an appropriate data model, constructing formal logic, and annotating that logic to source text is relatively time and resource intensive. But this approach is necessary to produce an accurate and transparent model that can be maintained in lockstep with constantly evolving regulation.

These carefully constructed models thus enable a decision process which is repeatable and explainable. For the same set of inputs, and for a given version of the rules, Adept will deterministically produce the same output decision. The decision process operates over formal logic annotated to digital text, meaning the evaluation path taken can be visualized as a logic tree, and each point within this logic tree can be traced through to the appropriate version of underlying source text. This approach to decision making enables the level of explainability and transparency required to evidence regulatory compliance.

Modern AI is an incredibly useful tool in certain applications. Here at Droit, we view it as an effective co-pilot or assistant – in the creation of code, written content, and other data – which can be leveraged to improve efficiency within our own production process. However, for us it is an aid and not a replacement for Knowledge Engineering in the creation of a robust decision process providing complete auditability for regulatory compliance.