Panel: Data Detectives: AI, machine learning, and data analytics in practice
Napier’s Chief Data Scientist Dr Janet Bastiman appeared on a panel of experts for a Regulation Asia webinar.
Also appearing on the panel were:
- Rana Datta (Moderator) - Managing Director at Protiviti
- Jee Meng Chen – Head of Financial Crime and Regulatory Compliance at AIA Insurance
- Luke Raven - Senior Manager, AML Advisory and SME Financial Crime at Westpac
Janet and her fellow panellists discussed the use of data analytics, AI, and machine learning for AML/CFT, challenges to adoption, and how to build an AI/machine learning roadmap.
Here are some of the key observations made by the panel:
1) AI can’t be successfully implemented without basic data hygiene
The talk kicked off with Janet observing that time and time again, companies are looking to expand their systems and implement AI technologies but are missing the basic level of data sanitisation that’s needed for an AI system to access the data it needs to make predictions effectively.
AI relies on a broad selection of high-quality data, but many financial organisations still store their data in very siloed ways. It’s not just the storage of the data that presents issues, but the data itself, which should be captured in manner that lends itself to being ‘clean.’
Where data is collected in free text fields, for example, it’s hard to split the collected details accurately for later use, so it’s important to get the basics right before introducing AI technology to existing systems.
Jee Meng echoed the sentiment noting that in the APAC region data is a very big issue. He highlighted that another key prerequisite to implementing AI technologies lies with the IT architecture of the company. Companies that have been operating for many years will at some point upgrade their technologies over time, but the issue remains with legacy technology.
Implementing AI for AML into legacy systems
In legacy systems, certain data parameters for today’s requirements are simply not there. Clearly, the full scope of the data’s purpose (customer experience, crime prevention) was simply not considered.
Jee Meng also highlighted that customer risk profiles are not necessarily all treated equally, as it is more likely that low-risk profiles might be updated less frequently than high-risk customer profiles.
Luke raised the matter of building for a minimum viable product, where businesses - especially start-ups - are product-focused and have not prioritised compliance beyond fulfilling the minimum requirements to operate thus far. He also echoed Janet’s point that data quality and hygiene are key, and that a single customer view is important prior to implementing AI into AML systems.
Finally, digitalisation of historically collected information – and building to collect data in a digitally savvy manner in future - is crucial to the success of an intelligent AI solution.
“With AI and AML, before you can get that in order, you really need to make sure that the well it’s drawing from isn’t poisoned.” - Luke Raven
2) The most effective applications of AI in AML
Janet observed that AI is great at speeding up complex but well-defined processes.
“When looking for a needle in a haystack, you either want to remove the hay or make the needle bigger” - Dr Janet Bastiman
She elaborated that AI can remove the white noise (or ‘hay’) in processes like screening where high volumes of false positives are distracting and time-consuming to sort through manually.
There is also the option to make the metaphorical needle bigger. What is meant by this is that instead of looking for a hit in a sea of false positives, the user can refocus their efforts on evaluating a smaller group of results that have been identified by AI to be the most suspicious.
Client and transaction screening are perfect for automation with AI, as the process throws up a huge number of false positives. It’s also an area that is likely to already have a good quality of data.
Luke added that screening is a good candidate for automation as it’s a process that is binary in terms of outcome, so it’s easy for the regulators and analysts who review the results to see in simple terms what needs further investigation. Even though the process of reaching a decision is complex, reviewing the results of the AI – the human part - is made simple.
Jee Meng said the beauty of AI is that it can compile information “at the push of a few buttons” from several systems to provide a snapshot view that helps analysts make quicker decisions around screening and monitoring without rushing.
Conceptually, this application of AI is a highly appealing and convenient addition to any business’ AML process, but the practicalities - the compatibility of the data sets, data hygiene, and the cohesion between systems – are major considerations that need to be addressed before an organisation can implement AI effectively and reap these benefits.
3) How the COVID-19 pandemic has made AI a must-have for AML compliance
Jee Meng explored the impact of the COVID-19 pandemic and noted that face-to-face interaction used to be the norm, and solely digital interactions the exception. This balance tipped and it’s ordinary now for interactions like customer onboarding to happen entirely online. He also pointed to a potential application of AI here to verify certain documents submitted digitally.
4) Why the AML industry is still hesitant to adopt the latest RegTech
The key factors discussed were:
- A desire for continuity; to continue doing what has historically worked.
- A need for explainability. Presenting regulators with the final product from a technical solution alone isn’t enough; the user needs to know the product well enough to explain the outcome and the process to satisfy the regulator.
- Regulators and regulatory guidance bodies – like the FATF – could do more to inspire confidence in AI, by providing more resources or by implementing it in their own systems.
- While government institutions are encouraging the adoption AI and machine learning for AML processes, they are yet to provide fully detailed legislation that explains exactly how organisations can follow this recommendation.
- Reluctancy can come from cost. Where cost and resources have already been sunk into implementing current AML systems, organisations are understandably less eager to cast these legacy systems aside or invest in further updating their systems. This is exacerbated by the fact that it takes time for new RegTechs to become widely accepted as being worth the cost and effort of implementation.
5) The operational challenges of implementing AI in AML
Luke highlighted a key fear surrounding the implementation of AI across all industries: human redundancy. To that, he suggested the narrative be reframed to view AI as a means to improve the effectiveness of a team; maximising resources rather than displacing valued team members.
Another key challenge is in training up the AI machine to recognise what’s ‘good’ or ‘bad’ in customer behaviours:
Fraud is comparatively binary in that a machine can recognise something as fraudulent or genuine, and therefore fraud detection is far more easily taught to a machine. In AML, behaviours are judged by the far less clear-cut definition of whether or not they are ‘suspicious’ - a concept which is influenced by a risk appetites and policies and are unlikely to be the same at any two organisations.
- The future of AML lies in the successful implementation of artificial intelligence and machine learning.
- Criminals currently adopt new technologies at a faster rate than the AML industry does, so it’s crucial that we are able to keep up to fight financial crime effectively.
- Implementation isn’t always straightforward, so data hygiene and sanity must happen before an AI system can start delivering tangible value.
Find out how to improve your AML processes with award-winning technology.