CompatibL’s unique AI strategy pays dividends

In addition to its recent Buy-Side Technology Awards success, CompatibL won the Best AI technology provider category in the American Financial Technology Awards 2025, thanks to its CompatibL AI offering. Executive chairman Alexander Sokol discusses the firm’s unique approach to artificial intelligence and how its research around cognitive bias and behavioral psychology has helped significantly improve the reliability of its AI-based applications.

Can you explain what CompatibL AI is and how it was conceived? Did the impetus for its development come from within CompatibL or externally?

Alexander Sokol, CompatibL: CompatibL AI is our premier offering. That’s where all the AI functionality is delivered. Our initial AI delivery was three months after OpenAI released the first version of ChatGPT, although our software today is radically different from what it was then—we rebuilt it from the ground up based on findings from our research into the psychology of AI.

We talked about this research at various conferences during 2025 and we’re writing a paper on it. What we found is that the reliability of AI is driven neither by model shortcomings nor the shortcomings around how the prompts are built, but rather by how well you understand the psychology of AI—namely, the cognitive biases and psychological effects it shares with humans.

Our research shows that large language models (LLMs) display surprisingly strong cognitive biases and psychological effects. The ground-breaking books Thinking, fast and slow by Daniel Kahneman and Noise: A flaw in human judgment by Kahneman, Olivier Sibony and Cass Sunstein brought the concept of behavioral psychology into the mainstream, and introduced many of the psychological effects and cognitive biases that are important to human interaction.

With very few exceptions, AI is subject to the same cognitive biases as humans, for two reasons. First, some of these biases are directly learned from the training data. In other words, behavior exhibited by humans becomes part of the information learned by AI and then, unsurprisingly, copied to its responses.

What is surprising is that we also see some of this behavior is clearly the result of the functioning of the transformer architecture used by AI, yet it produces human-like psychological effects and biases. The striking similarity between how humans and AI process and respond to information indicates that perhaps human cognitive processes and transformer architecture have more in common than anyone realizes.

Thanks to our research in psychology over the past year, we have achieved dramatic increases in reliability. We were able to achieve 95%–97% accuracy before we started our research, but then we realized that what was holding us back from achieving near 100% accuracy was psychology, not engineering. We then rebuilt our new modules—Credit Advisor AI, Compliance Advisor AI and Legal Advisor AI—from the ground up, taking cognitive bias and psychology into account, and we’ve been able to achieve reliability on a totally different, human-like level.

What are the practical implications for capital markets firms using CompatibL’s AI solution and, specifically, which business processes are enhanced through the use of CompatibL AI?

Alexander Sokol: CompatibL has a somewhat unique approach to deploying AI in the financial services industry. Specifically, we don’t focus on generation—we focus on comprehension. This is where CompatibL AI can have the greatest impact.

There are many software applications that offer interactive assistance for coding, drafting, and so on, but there are very few that offer specialized comprehension of documents specific to finance. For example, determining compliance of a prospectus with a specific regulation, which involves reviewing a 500-page term sheet and extracting the relevant information from it based on an intricate set of legal rules and guidelines.

So, while we use AI for generation in a few projects, our focus is AI-assisted comprehension—effectively, converting free-form documents to data. We believe that, for comprehension, we have state-of-the-art, best-in-class capabilities, not only thanks to all the engineering that went into it, but also thanks to our recent research on psychology that we incorporated into our solution.

Our research shows that LLMs display surprisingly strong cognitive biases and psychological effects … The striking similarity between how humans and AI process and respond to information indicates that perhaps human cognitive processes and transformer architecture have more in common than anyone realizes.

Alexander Sokol, CompatibL

To what extent are the LLMs underpinning CompatibL AI able to learn from and improve the results they generate as they become more experienced? In other words, do they get better as they become more experienced?

Alexander Sokol: Yes, CompatibL AI absolutely gets better, and it improves not only through the evolution of the foundation models but, more importantly, through the evolution of the layers we build on top of them that takes into account psychology, and essentially guides the model through what we call a cognitive bias-free setup. We build value-added solutions on top of the foundation models that the model vendors offer us, and we deliver these complete, reliable and well-tested solutions to our clients.

One of the relatively recent developments in AI is the “thinking” model. These models are able to follow a chain of thought and come up with better, even if not immediate, answers.

Still, I would rather use CompatibL AI’s workflow with a year-old, non-thinking model than the latest thinking model without our workflow, as what we have found is that it’s not the model’s limitations that impact its reliability, it’s getting the model into a workflow where it can be free from cognitive biases and psychological effects.

We have a paper coming out shortly that demonstrates that you can use a mini or even a nano model. These are smaller, faster and cheaper non-thinking models that are not as capable as modern thinking models. They produce much better results with our layer around them than the most advanced and expensive models on their own, as measured by rigorous statistical tests.

The continuing improvements in CompatibL AI are driven by much more than the evolution of foundation models. When we enhance our software, we continuously add new use-cases or examples. Each time we see the model misfired in some way, we add a corrective few-shot example—providing detailed information about what went wrong and how to correct it.

We have also developed what we call a “reverse lookup”—a kind of database where you can find similar use-cases and see how the model behaved in related instances. We use reverse lookup to select the most relevant, corrective few-shot examples, which has a powerful effect on the model understanding where it went wrong and provides it with very specific, relevant guidance. And, each time we identify where the model misfired, we reproduce the same type of error with public, non-confidential data, and then we add it to the reverse lookup database. So, the next time the model encounters the same example, it has guidance to help it understand where it went wrong and it is able to correct it. That’s how CompatibL AI improves with each release, and this is an important part of our software’s value.

What has the uptake been like among CompatibL’s clients?

Alexander Sokol: The feedback has been phenomenal. In fact, more than half of our clients who started with other functionalities—trading and risk solutions, for example—are now also using CompatibL AI, and this will be probably be more than 80 or 90% by the end of 2026.

This rapid adoption is the most honest feedback we could receive; it confirms that CompatibL AI is delivering real value and helping clients use the technology more effectively. Unlike complex quant models, which take a long time to learn because you need very specific expertise, an AI-based tool can be built easily by any software developer, especially with the coding assistant tools available today. That’s an aspect of AI we welcome because it keeps vendors honest—we have to prove that we are delivering added value to the client and the high rate of adoption of CompatibL AI is a testament to that.

Sponsored content