Compliance relies on human intelligence. What is the role of machine intelligence and AI?
To analyse and define the role of machine intelligence and artificial intelligence (AI) in compliance, let's start by looking at the gaming industry. The ancient game of Go was invented in China more than 2,500 years ago. It is believed to be the oldest board game continuously played to the present day. While the rules of the game are relatively simple – two players compete to surround more territory than the opponent by capturing each other's stones or pieces, as they go along – Go is actually a very complex game. Comparing it to chess, Go has both a larger board with more scope for play and many more alternatives to consider per move (10^800th to be exact).
Two years ago, in London, technology changed this ancient game forever. Up until early 2016, Go was considered to be the last game with ‘complete information’ – where humans could perform better than machines. However, in March 2016, a team of data scientists from Google's DeepMind created a computer program/algorithm that became the first computer-based Go program to beat a human professional Go player. And it wasn’t just any professional player that DeepMind beat – it played against one of the world’s elite Go players. In a five-game match, the program beat the champion four to one. The victory was a major milestone in artificial intelligence research. Go had been regarded as one of the most complex challenges in machine learning, and it was expected to be out of reach for technology for at least another 10 years. Yet DeepMind’s program continued evolving and learning by playing against itself to constantly improve, and ultimately laid the foundation for AGI (artificial general intelligence).
That algorithm was called AlphaGo. It marked an end of an era when victories in popular board games could serve as major milestones for artificial intelligence. It also opened up a wider use and application of such technologies in other areas.
Compliance and the role of machine intelligence
Up until now, compliance functions have mainly relied on people. Over the last 10 years, we have seen an increased demand for hiring within all functions of compliance. Struggling with growing regulatory requirements, financial institutions have been constantly putting out fires. And the only way they knew how to do this was to throw more compliance staff at the growing burden.
Meanwhile, a number of pioneers and industry influencers have been speaking out and pushing for automation, especially within the back- and middle-office. There are a number of good reasons for this argument. For example, machines do not need to eat or sleep, and can work 24/7. They do not have feelings or emotions. They do not take sick days. They simply work as per the details fed into their mechanical brains. While humans behave as per their consciousness, machines perform as they are taught. Decisions made by machines can also be backtracked and analysed.
The situation we have now with compliance and technology resembles the time when the first vehicles were invented and manufactured in the 20th century. People on the streets who saw these vehicles thought that this was a new epoch and that infrastructure would change radically. According to the authors of the book ‘Integrating AI in Highly Regulated Industries’, we are now in the middle of the same kind of revolution – the AI revolution. Modern technology development is going to take this route. This is also something the authors of the book ‘The Amazing, Anti-Jargon, Insight-Filled, and Totally Free Handbook to Integrating AI in Highly Regulated Industries’ have discussed.
Advancements in AI
In a vast universe of technologies, artificial intelligence is the latest technology to play a part in the digital transformation of the financial services industry.
Andy Pardoe, director of AI at Accenture, defines AI as a broad term that covers a multitude of techniques, from simple rules-based methods through to natural language processing that uses deep learning. The main focus with AI at the moment is with a subset of techniques that fall into the machine-learning category.
Essentially, AI is a series of underlying technologies: natural language processing, computer vision, machine learning, neural network and others. They are all brought together within a cloud-based environment that can store and process gigantic quantities of data and allow for instantaneous AI interactions.
There is a wealth of research written and published about theoretical and practical implementation of AI algorithms. In fact, the field has been developing since the 1950s. And while some algorithms are very linear, the newer algorithms take advantage of larger volumes of data and hence begin to imitate human thought processes.
Right now, there are many very interesting machine learning/AI projects going on in the compliance field. Modern-day technology has powerful capabilities and analysts are looking at how AI can solve real-world compliance issues.
With the amazing recent advancements in technology comes hope that AI can solve all of our compliance problems. Yet at what cost? Throwing machine power at a problem could be justifiable. But is it always reasonable? And could other methods be used to improve things before throwing a ‘deep’ algorithm at it.
Besides advancements in technology, there is ground-breaking work going on in the industry around data management. For instance, in the field of symbology, one example relating to MiFID II is the extension of the use of ISINs for OTC derivatives. While the idea has received some criticism, there has been a lot of thought and effort put in to get us to this point.
We also see a strong case for businesses to focus on classifications and get their internal corporate taxonomy right. From recent developments, it is interesting to observe the evolution of taxonomies in the crypto assets space. There are a number of interesting projects in the ontology space too.
However, like other powerful tools in finance, there is the possibility of “sub-optimal outcomes” if AI is not accompanied by management judgement, said Eugene Ludwig, chief executive of the IBM-owned Promontory Financial Group, who recommended having a designated officer in charge of reviewing AI applications.
This leads the context of this article to a word which we have been hearing a lot in the last couple of years. That word is ‘semantics’, or ‘the meaning of things.’
Semantics and the state of data
Building a semantic web is a great idea in itself. We can state for sure, that semantic web tech adoption during the next 10 to 20 years will make banks far more transparent and manageable for the regulators.
Semantic technologies for smart data processing relate to ideas professor Tom Butler started developing back in 2002. The approach in itself is unique and definitely academically correct. However, since the beginning, semantic web paradigm has experienced some difficulties in getting set up and has had a lack of general-purpose usefulness. An earlier conference paper by professor Butler dated 2016 also suggested completing the semantic system with ‘knowledge engineers’, which adds the same risks which the industry has had up to now – that the growth of the amount of data is faster than the growth of employees' effectiveness who are expected to structure the data. Humans were expected to be involved in creating structured data – keywords and dictionaries – to build semantic models. It was the major weakness of this approach because it was manual, and hence quite expensive and not so easy to scale.
That's why we are happy to notice the shift in this paradigm towards an AI-based cognitive computing system. SmaRT Semantic web approach is a recognised technology. At ClauseMatch we see it as a helpful step in data structuring tasks. And it is an important argument in favour of AI-based RegTech solutions which will lead to the larger popularisation of these technologies.
Application of AI-based systems will significantly reduce the costs of semantic web systems implementation by eliminating the manual work weak point. The possibility of using structured data that applies AI-based algorithms could potentially increase the profits from implementing semantic web systems. The work regulators are carrying out in this sphere, with the FCA pioneering the way, is encouraging. With more data available through RDF-based endpoints, RegTech solutions such as ClauseMatch would become significantly more effective by being able to consume and analyse the structured data which is constantly coming from the regulators.
Our Experience
At ClauseMatch, we have started embracing new technologies, and have gained experience in leveraging semantic-based algorithms. This year, in collaboration with data scientists and machine-learning experts, we developed and tested a system that can identify and compare paragraphs across regulations and grade their relevance to each other based on their semantics.
The work of our algorithm was tested on the concept of ‘whistle-blowing.’ Results and output it produced were quite impressive – significantly outperforming classic statistical-based approaches. Relationships were detected even when paragraphs had absolutely no words in common but discussed the same things, since the machine learned how to represent text in the semantic multidimensional space, where phrases like ‘whistle blower’ and ‘anonymous report’ were close to each other. Even more impressively, adding the ‘UK’ word vector to the ‘money’ word vector in such a space gave us results around ‘pounds.’ Now, you can imagine all the possibilities that could arise from manipulating words and phrases like that.
We were able to write the algorithm well from the get-go because the approach is based on strong maths. It allowed us to take advantage of the recent NN advancements, and it was perfectly aligned with well-maintained structured, comprehensive datasets.
Who is better at performing regulatory compliance? Human or machines? Or humans with machines?
So, which side should we favour in today’s digital age? Humans or machines?
In some industries, the technology story is about replacing people. But not in compliance. The situation here is more about augmenting people.
In the battle of humans versus machines, we believe there is no clear winner. Because, ultimately it is all about the right balance. And when it comes to compliance, it is not about humans versus machines, but instead, humans WITH machines. Each has an important role to play. The strength comes from the two working well together.
AI is not a panacea or a magic bullet. It is a technology that can be leveraged to improve and stimulate good culture within an organisation on the inside, and create consumer protection and better service on the outside.
Compliance is about responsibility and ethics, which come naturally to humans. At our recent event ‘Compliance and Regulation 2030’, we ran a survey and asked the attendees whether we are likely to see robot compliance officers in 2030. With a 77% positive response, there was an interesting comment: “If so, compliance as such will not exist."
Compliance is evolving fast. And it needs to. In 10 years from now, what we are most likely to see is a three-level structure with a human at the top, AI at the bottom and an automated 'decision support system' in the middle, providing a 360-degree picture into the current state of compliance that prompts humans with the right decisions. At the lower level, artificial intelligence will be complementing the compliance function by offering to take various approaches to a problem and therefore helping humans, but not making decisions completely by itself.