Introduction

The world of finance is evolving. We pay for goods and services with watches. New banks and financial companies appear seemingly daily. We regularly make cross-border payments to global marketplaces. Many of us barely use cash. Crypto is relatively commonplace. In our first article in our Future of Finance series we spoke with Ricardo Correia about what is happening inside the global Banking sector with regards to digital currencies and how major change is literally round the corner.

Where there is money to be made, there are criminals. Turning the proceeds of crime into legitimate money is the process of money laundering – and it can be a sophisticated process. Criminals don’t want to be traced so tend to use cash to pay for illegal goods and services. Cash is difficult to spend, especially in an increasingly cashless society. Whilst the world of finance is evolving the criminals are evolving to exploit holes in the system.

This is the second article in our Future of Finance series, in which the amazing Dr Janet Bastiman talks about how “intelligence driven” anti-money laundering and compliance technology can rise to the challenges of different payment devices, microtransactions, and digital currencies. There are also some juicy Artificial Intelligence and Machine Learning (AI/ML) topics to sink your teeth into, like ethics, hallucinations, and data quality.

The situation today

Have you ever had to do mandatory training on financial crime at work? If you have then I am sure you will have a good sense of what money laundering is. But, just in case you weren’t paying attention, I asked Janet to explain what anti-money laundering is and how it fits into the world of finance.

“Anytime there’s criminal activity and funds are generated, there’s a need for that money to look like it’s legitimate. That’s the basis of money laundering. It’s sort of like hiding the proceeds of crime so that someone can then spend that money on cars, houses, whatever they want to spend it on. So, the people that do this are trying to hide that trail.”

“Before we all went very, very digital, money laundering involved cash transactions”, continued Janet, “changing the currency that money is in, and moving it around different bank accounts can make it very hard to trace. But now, the majority of transactions, particularly in the UK, are online”. You can see the trend of payments in this Bank of Internation Settlements paper. “It makes money laundering much harder because you have a digital trail. But not impossible. There are lots of gaps where information gets lost, where it’s very hard to trace the true origins of money, particularly with all of the data privacy laws that we have in the UK and worldwide. Really getting to the bottom of where money has come from and if it is legitimate is very difficult”.

You would think it is easy to detect money laundering transactions in a digital world. Everything is electronically recorded. But criminals make it as difficult as possible – after all they don’t want to get caught. They have developed all sorts of innovative tricks to launder money.

Janet began to elaborate what her team does at Napier, “so one of the things that me and my team do is look at different typologies. These are the different methods that money launderers use in order to hide and obfuscate criminal proceeds. Transactions typically follow patterns. We detect unusual activity so that the banks and the financial institutions can really look into those transactions and then make that final determination as to whether it’s something legitimate or not”.

I know AI is already used in aspects of the anti-money laundering space. I also know very few organisations don’t use AI to make final decisions as the technology is in its infancy and fraught with challenges. I wanted to be clear, so asked Janet if it is a human or machine that is making the final determination on a transaction that is suspected of money laundering.

“At the moment that very final step is human”. Janet continued to explain, “they will have loads of algorithms in place, so there’ll be a whole set of standard deterministic rules-based heuristics to look for odd transactions or behaviours. There will be a layer of, sort of machine learning processes, that will adapt based on those behaviours. There will be some overarching workflows, some of which may be automated, some of which may be fully human. That will end up at a point where someone will make a decision as to whether the account, and certain transactions within it, need to be raised as something that the authorities then need to take further action on. And the finance institutions themselves can also make a determination to maybe block accounts if they think that’s appropriate as well”.

So, the process today is each transaction is screened by automation, but a human makes a determination as to whether they believe it to be money laundering. That is reassuring. There has been so much bad press about AI making incorrect decisions, having data biases, non-repeatability (meaning if you run the same analysis twice you can get different results), and of course hallucinations, that relying solely on technology to decide if a transaction is legitimate feels wrong. Janet went on to provide a really good example of why having a human in the decision making process is so important:

“A lot of the typologies connected with money laundering really, really overlap with the patterns that you see in the accounts of some of the most marginalised in society. So, if you think of people with like five or six, zero hours contracts, people who are very cash based and their money is quickly in and then out as they just don’t have the savings. You see in their accounts this rapid exchange of money, you see odd amounts coming in at odd times, and it is those sorts of patterns that can flag up automatically”. This completely aligns with what Janet said earlier about money launderers moving money between accounts. “The last thing a financial institution wants to do is block accounts of people like that because it has a big material impact on them. So, as a result, you tend to have that human oversight, rather than automatically stopping everything”.

The very language that’s being used around this, hallucination, implies that it’s this higher functioning thinking thing that’s imagining amazing things, when it’s an output from statistical solutions

“I know a lot of banks really try and protect their users as much as possible, but they’re balancing that against that level of frustration. [But], if something went out of your account that you didn’t want, I mean this is more the fraud side than money laundering, you’d still want to be reimbursed. And one of the great problems with a lot of financial crime is that if you lose money through fraud and you’re reimbursed, that’s not your money coming back. The criminal still has your money. It’s someone else’s money that you’re getting back. So, we all have a duty to try and stop this as much as we can”.

WATCH THE VIDEO FOR THIS SECTION

Intelligent anti-money laudinging with AI/ML

From an anti-money laundering point of view, we’ve still got a lot of humans making decisions in whether transactions are legitimate or not. Good. We also know that some transactions are incorrectly blocked. The financial institutions can’t/don’t want to hire an army of people to investigate transactions, it just doesn’t make sense. But there are so many false positives to investigate, where normal transactions are flagged for investigation but found to be fine, that they need different, reliable approaches. With such a seemingly recent rise in AI capabilities, as people external to the world of technology may perceive them at least, (see AI in Reality), I rather flippantly asked Janet if ChatGPT could solve all these problems 😊

“Well, maybe at some point in the future”, smiled Janet. “the problem that we have now is large language models like ChatGPT are very, very good at taking vast amounts of data, whether that’s text or video or images, and spitting out something that looks legitimate. You know, an image that looks real. The problem is that while it can look compelling, it’s not necessarily precise. And while there’s plenty of news items like ChatGPT can pass this test, if that information is freely available on the Internet, then yes, it can regurgitate it. But when you’re looking for something that’s outside its scope of knowledge, it will fail. But it won’t tell you that it doesn’t know how to do it. It will come up with something convincing”. The industry has coined a name for this type of response, which we will discuss later.

“When you point it at financial transactions and say: give me a summary of these, is there any evidence of money laundering, it will come back with something. It might say yes, it might say no, but it will come back with something. And even when you try asking it to explain, it won’t necessarily go back to the original transactions. You don’t have that connection and that detail, and this is what it’s missing. So, until these models have the level of explainability that more historical models have … you’ve not got the confidence that it’s accurate”.

Janet started to explain another challenge of LLMs/AI in the anti-money laundering realm, “the money laundering sphere is not static. We’re not looking for the same thing. The criminals have their own teams and they’re constantly changing what they’re doing because they’re taking advantage of different markets and opportunities”. Where we’ve seen some great applications of AI in things like cancer detection or drug innovation, this data is static. The models are looking for the same thing each time. When the patterns are established, they barely change. In the realm of anti-money laundering, the criminals are always trying to change, to innovate, to evolve – to beat the system.

The next thing Janet said gave me an idea of what Janet and her team are up against, “You know, during COVID there was a huge influx of money laundering and taking advantage of the increase funeral homes and the higher death rates. [The criminals] will adapt based on where they think they can move the money and get the quickest turn around. If you trained a model based on historical data, it won’t pick up those new patterns”.

Whilst there are still a lot of challenges around LLMs and AI, despite the hype that surround them, one particular phrase has Janet troubled: hallucinations.

Every role I’ve worked in has required precision. If somebody transferred money to you, and you got most of it, would you be happy? If you order online and you typically get what you ordered, would you be happy? If you press the brakes on your car, and more often than not they slow you down, would you be happy? Don’t even get me started on plane travel! Can technology that always gives you a response, but “hallucinates” be used in production systems? But the hype around the industry is compelling people to develop solutions using faulty technology. This is one of the key reasons why “copilot” style adoption is so popular. In a copilot deployment, the AI analyses vast amounts of data, and a human can use its output to inform their decision.

“We’ve already, you know, always had accuracy and precision and really measuring our false positive rate, what’s our false negative rate, you know, how far are we deviating from truth in our predictions? And we need to start applying this to these models before they’re allowed to be used in production, because once we’ve got that measure of error, and we need to call it what it is, it’s error, then we can say, well, what is our acceptable risk for using these in production given that the error rate is X percent”.

Risk can be a terrible word in the world of finance. I say can be because quantifiable risk is a different beast to indefinable risk. As per Janets explanation, if I know there is a 10% chance I won’t receive all of the money that is transferred to me, that is very different to 0.001%. And both of those are very different to shrug, dunno. For anti-money laundering in financial institutions, the balancing act is between risk and the cost of false positives and manual intervention. I asked Janet if the risk of introducing AI made the financial institutions really cautious.

“It does vary, so depending on jurisdiction and how supported the financial institutions feel by the regulators in their area, that does affect how far they’ll go with AI. But you’re absolutely right, they’re very, very risk averse. And quite a lot of the conversations I have are about identifying that risk, because everyone wants to reduce false positives. But they think of it at that level: I want to reduce my false positives because then my team will be less overloaded, and we can focus on the true positives and doing thorough research, which is a great aspiration. But when you say what really what is your risk level, what are those false positives that you always want to see and investigate … that’s quite a difficult question to answer … sometimes it can be hard to encapsulate in words, because it’s gut feel of experienced analysts”.

Janets words go right back to another of AI’s challenges – not being human. AI can do some amazing, seemingly human things far more easily than humans. It can hold incomprehensible amounts of data in its “working memory” when compared to humans. It can access this information at lightspeed, compared to humans. It can combine and create content far faster, and often better than humans. But it doesn’t have some of our more human qualities at all: intuition, empathy, sarcasm, emotional intelligence, moral and ethical judgement, spiritualism, compassion, genuine creativity. AI definitely doesn’t have gut feelings; the type of feelings a seasoned anti-money laundering investigator will have.

“Too often, people will jump straight to: we’re overwhelmed, so let’s throw AI at the problem. But then, because they’re pushing in poorly labelled and poorly classified data up front, they’re not going to end up with a successful implementation”. This is so true, and something both Marcus Desai (Gaming Anti-Cheat Through Data Science as a Service) and Nayur Khan (AI in Reality) said when we spoke with them.

“I think it’s definitely the case that these tools can help us spot things faster, because even with a team of the best human analysts, you’re limited by the data that you can retain in memory at any one time. It doesn’t matter how many spreadsheets or BI tools you look at, you need to know what you’re looking for in order for things to pop out for you. If you want to step back and see everything and look for patterns in very large multi-dimensional data sets, you need tools for that, and that’s what AI is very, very good at”. You know there is always a but, right?!

“That said”, I told you, “we are more restricted because we have to work within the law, and we have to work within an ethical framework, and as such we’re limited on what data we can look at, where we can look at it, and we have to keep data sets separate. We have to make sure that we’re doing all of our testing, so that we’re not unfairly biassing, which is all good stuff and absolutely needs to be done. But the people that we’re working up against, who are trying to make money and trying not to be caught, are not subscribing to all of those principles. So, while [AI helps us to] detect a lot faster, the only way we can stay ahead is having that sort of Red Hat thinking of, if I was trying to move money, how could I potentially do it”.

WATCH THE VIDEO FOR THIS SECTION

Anti-money laundering and crypto

Having spoken with Ricardo Correia a few months ago about the changes to the world of finance coming through digital currencies and crypto, I was keen to get Janet’s view on how anti-money laundering could be applied in this new world – especially off the back of all of the negative press around financial crime in the crypto space. Remember, there is a question of whether financial crime is more widespread within crypto, or it is just more heavily scrutinized because it can be. Decentralisation and typically public, cryptographically signed and distributed ledgers have a permanent record on the blockchain that can be viewed by all.

“The problem is the difference in the data protection aspect, because with standard financial institutions, you know exactly who the person is. You know, they’ve been fully verified and we can see the to and from person [for the transaction]. In crypto, unless you have access to the crypto exchange because you’re providing regulatory services for them, it’s a completely opaque system. However, law enforcement agencies, if they need to trace through that, they can do”.

I mentioned above about typically public ledgers. There is a huge trust issue with blockchains that are not public; why wouldn’t you be public, what are you trying to hide. In our first Future of Finance article with Ricardo I referenced Facebook Libra. From the article: Libra was going to reside on a private blockchain where members performed reconciliation behind closed doors. This felt like an attempt to prevent anyone but the members of the Libra Association from being able to analyse the distributed ledger, negating many of the promises of blockchain based finance. Why? Immediately you think there is an underhand motive.

Janet continued to talk about anti-money laundering in crypto, “You know what? We’re actually seeing less use of [crypto] because, you know, the whole nature of blockchain is that it can be fully traced. So when you want to move money around, it’s not actually as obvious a source as maybe it was when it first got set up. At the end of the day, the criminals still need to take the money out in order to pay for things, so there will be an in point and there will be an out point, and all of the wiggles in the middle. If we need trace, it can be traced by the relevant people, but you know if large amounts of funds are coming into an account from crypto, that’s just as much of a red flag as if it’s coming in from any other source”.

“From my point of view, we’re looking at companies and individuals and their interactions with the wider network. We’re looking at movement and the typologies around those. So regardless of the origin or the beneficiary, whether it’s crypto, cash, a deposit, a loan, all the different types of sources that those things could come from, you can apply very similar techniques. I mean they’re not identical, but you can apply very similar ones and you know you might need to tweak how you’re looking at things depending on the source”.

This brings a lot of hope for the future digital banking system that Ricardo described. If it is harder for criminals to move money around, and when they do they are more likely to be traced, then hopefully we will all be better protected. Maybe the negative press around crypto, and criminal activity, shouldn’t be quite so negative?

WATCH THE VIDEO FOR THIS SECTION

What does 10 years’ time look like to you?

We’re talking about the Future of Finance. With her AI/ML and anti-money laundering expertise I asked Janet how she sees the future, especially as everything feels like it is moving at lightspeed.

“If I think back to 2014 and the thought of me paying for something using my smart watch, it’s just crazy. I can go back a further 10 years to the thought of even having banking on my phone. I had a colleague who was [working on a] very early banking app on a, I hesitate to call it a smartphone, but it had a colour screen. I remember thinking, why would anyone do banking on their phone? I remember my parents telling me when they got married in the 60s, in order to get money out at a bank where they were on honeymoon, they had to meet their bank manager and set up an arrangement between bank managers of the different branches. In maybe 2 generations we’ve gone from this very cash based society to, I literally can’t remember the last time I took cash out of the machine to pay for anything”.

I love the way Janet thinks. There are so few people that I speak with, particularly in the world of technology, that when you ask about the future they first reflect on the past. As Confucious said, “study the past if you would define the future”. History repeats itself and there are so many patterns in human behaviour and evolution, and mistakes for us to build upon, that reflecting on the past is something I wholeheartedly recommend.

“We are very, very digital now and with the Open Banking changes that have come in, you don’t even need to have five or six different apps on your phone, you can have a banking app and all of your different accounts are in there. I know some people don’t feel comfortable about that, but maybe in a few years’ time that’s going to be the norm”. I have to admit, this isn’t something I have latched onto, yet. I quite like the idea of one great banking app, with an amazing User Experience (UX), but I personally still feel there is a Darwinian evolutionary race to define that great UX.

“There’s also a sci-fi stage where you just have a retinal scan and it takes money from your account. I know some countries have been looking at the idea of paying for things just through facial recognition, and I can see all of that becoming the norm very early in the next 10 years or so. What seems silly now will just be normal because it’s more convenient, and if it’s more convenient, we’ll do it. If it becomes less convenient, it won’t happen”. This is so true. I literally sigh when I have to put my pin in nowadays, or try to remember my CVV when it isn’t auto-populated during online payment. And yet I can still remember how convenient it felt to pay using an imprinter or ‘click-clack/knuckle-buster’ that created an imprint of a credit card on carbon paper. These paper slips were then sent away to be processed manually by the merchant. “But then when you think of tracing, you know, from my point of view, and you look at the financial crime around that, you’ve got that verification step at the beginning. Is this person who we think they are. Once that’s in place, you then need to do all of the same things we currently do”.

“You’re still fundamentally looking about the movement of an amount of funds from one point to another, and the patterns that go with that. But it might start to make it easier to identify potential problems because as soon as you start getting any sort of biometric recognition, you might have extra information. Similarly with geolocation. If your eyeball has been scanned in Leicester Square and then 30 seconds later, your face is scanned in Swansea, you know something’s amiss, and they’re not well! Well, until teleportation is invented anyway! Who knows, we might get to the Star Trek stage where money isn’t even a thing anymore”.

From what Janet has said, I feel like the next 10 years could give the anti-money laundering companies and financial institutions a fighting chance to work on their data classifications and AI processing to reduce false positives. It also gives the AI industry some time to move towards production ready solutions – although I feel there are some fundamental challenges that need to be overcome yet. There are also performance considerations of running all the algorithms in real-time to potentially block transactions. I wondered if Janet felt that the next 10 years would indeed provide an opportunity to consolidate, and maybe take an upper hand in the fight against the criminals.

“The data will always be messy. Over my entire career, I’ve been dealing with messy data, you know for the past 25-30 thirty years. Data has always been awful because of the nature of, when you collect it, you don’t know upfront everything it is going to ever be used for. You make decisions based on what you know at that time. In five years or so, when someone says: oh, but you’re not been collecting this field, everything becomes messy because people try and retrofit it. So, all of that is going to continue to be a huge problem”. This is something we’ve heard from both Marcus and Nayur. Understanding your data when you collect it, and understanding why you collect it are so important.

“The number of transactions is still going to increase, as you have more people, and they [digitally] pay for more things. Then add all the different transaction types. If you think of all the microtransactions that people make nowadays for online games, you know, 99 pence here and there, that volume of transactions just swells everything. If you think of that exponential growth, and other areas of our life where microtransactions come in, that information is just going to balloon”. This is so true. If you think about how our spending patterns have changed, we know pay little and often so there are far more transactions to screen.

“10 years ago, we all went and we bought software. You had a licence. If you wanted the upgrade, you would go and buy it again. Now everything’s a subscription. We don’t seem to go out and buy CD’s anymore, it’s all just an online streaming service. The music and the TV and the games and you know, even our operating systems, everything is just so much a month. And then you add our phones and other devices. It might start free, but then 99 pence here or there, it just exploding. So those sorts of things have the possibility to overwhelm. Going back to the money laundering, you’re looking for patterns and you’re looking for ways in which people can hide large amounts of money in transactions. The more transactions there are, the easier it is to hide. You know it’s the needle in the haystack when the haystack’s the size of the entire United Kingdom”

Final thoughts

There is always going to be a game of catch-up, with the criminals changing direction to remain undetected, and the anti-money laundering companies trying to detect their new approaches. With the increasing number of transactions, through microtransactions and different types of payments, this challenge is going to increase in difficulty. But many of the same underlying principles for anti-money laundering hold true.

But Artificial Intelligence and Machine Learning (AI’ML) are showing promising signs in this space, particularly in an augmented, copilot role. AI/MLs ability to scan vast amounts of data goes beyond the capabilities of humans, whilst human traits such as gut feel, intuition, and even empathy cannot be replaced. However, before AI/ML can be brought to fruit, some of major technical challenges need to be resolved.

Checkout the podcasts for this article