# AI and the Future of Compliance in Financial Services - With Hani Hagras of Temenos

- **Show:** AI in Financial Services
- **Published:** 23 Mar 2020
- **Source:** https://aiandbanking.libsyn.com/ai-and-the-future-of-compliance-in-financial-services-with-hani-hagras-of-temenos
- **Transcript:** https://podcasts.opennash.com/shows/ai-in-financial-services/ai-and-the-future-of-compliance-in-financial-services-with-hani-hagras-of-temeno-b4d9df/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Hani Hagras of Temenos argues that banks adopting AI need explainable systems, meaning models that show their reasoning in plain language. Regulators, bank staff and customers all need to understand decisions such as a rejected loan. Accurate but opaque models are hard to defend in regulated areas like lending, so explainability is the condition for AI to be trusted and compliant in banking.

The central tension is accuracy versus transparency. Hagras says opaque neural networks are fine for speech recognition but not for credit decisions, and that bolting an explainer onto a black box falls short. His alternative is short if-then rules learned from large data sets and combined with human expertise. That lets banks explain a decision and tell customers what to change, which turns compliance into opportunity.

## Key insights

1. **Regulated lending cannot rely on black boxes** Hagras says accurate but uninterpretable models are hard to justify to regulators and stakeholders. Customers also have a right to know why they were rejected, so lending needs transparent models.
2. **Hundreds of inputs can become three-part rules** Hagras says Temenos can take 600 to 1,000 features about an applicant and produce rules with only three or four conditions. Anyone can read them, challenge them and show them to a regulator.
3. **Explanations let banks advise customers, not just refuse** If a bank knows why someone was rejected or will run short of cash, it can say what to change. Hagras describes suggesting faster collection of money owed, reduced spending, or an overdraft.
4. **Explainers layered on black boxes fall short** Hagras criticizes 'model induction', where a second model interprets an opaque one, because a black box still makes the decision. He wants models that are transparent from the inside out.
5. **Credit, cash flow and fraud are easiest starts** Hagras calls credit decisioning the low-hanging fruit because people want to know why a score is what it is. He also names cash flow prediction, fraud and conversational banking as strong early areas.

## Questions this episode answers

### What is explainable AI and why do banks need it?

Explainable AI is AI whose decisions can be understood and justified in plain terms. Hani Hagras says banks need it because regulators, bank staff and customers must be able to see why a decision, such as a loan rejection, was made.

### Why is a black box model a problem for loan decisions?

A black box model can be very accurate, but nobody can see how it reaches its answer. Hagras says that makes it hard to convince regulators it is safe and leaves customers unable to learn why they were turned down.

### How does Temenos make AI decisions understandable?

Hagras says Temenos generates a small number of if-then rules from large amounts of data, each only three or four conditions long. An example is that low income, a short time at the address and recent loan defaults make a customer likely to be a bad credit risk.

### How can explainable AI help a customer, not just the bank?

Hagras says that if the system knows why a loan was refused or why cash will run short in six months, the bank can suggest what to do. That could mean collecting money owed faster, cutting spending, or offering an overdraft or loan.

### Which banking uses of explainable AI will arrive first?

Hagras says credit decisioning is the low-hanging fruit, meaning the easiest early win, followed by cash flow prediction. He also names fraud and conversational banking, such as chatbots that give personal advice.

## Facts

- **Daniel Fagella** [0:00:33]: "We interview Hani Hagras, who's the chief science officer of Temenos."
- **Daniel Fagella** [0:00:33]: "Temenos is a financial services software company that's been servicing banks for decades now"
- **Hani Hagras** [0:02:46]: "Now it is very, very cheap to store huge amount of data. and also we have got very efficient databases which can easily organize such amount of data."
- **Hani Hagras** [0:05:40]: "as long as your Amazon dot is doing very well as to colonizing your speech, you don't really care what is happening in the background"
- **Hani Hagras** [0:06:25]: "if you have been rejected, for example, a certain loan or a certain financial product, they have got the right to understand why I have been rejected for this."
- **Hani Hagras** [0:06:55]: "They put it in black and white that if artificial intelligence is going to be accepted as a trusted entity in our society, it needs to be intelligible, needs to be explainable."
- **Hani Hagras** [0:09:13]: "We have been working in this area for about 20 years. And one breakthrough that we managed to do, and we have been granted several patents out of this"
- **Hani Hagras** [0:10:28]: "what's called model induction systems which basically still rely on a black box model running in the background and try to interpret it by another model"
- **Hani Hagras** [0:12:02]: "we are able actually to take 600 or 700 or 1 ,000 features about yourself, which come from any kind of credit referencing agency."
- **Hani Hagras** [0:12:31]: "generate from this kind of data models that are having few rules where the length of each rule is only three or four antecedents although that your models might be using 1000 inputs"
- **Hani Hagras** [0:13:31]: "If there is a model looking at your transaction and is able to predict that in the future, in the next six months, you'll be in deficit."
- **Hani Hagras** [0:15:13]: "if you speak about something like churn, there is no point of me just predicting that this kind of customer is going to churn by a neural network and that's it."
- **Hani Hagras** [0:16:00]: "And what is important as well to understand why this kind of prediction has been flagged as fraudulent so that basically we can reduce the false positive rates."
- **Daniel Fagella** [0:16:50]: "if you can't explain a certain amount of why you said no to these people and yes to these people, we can take away your bank license"
- **Hani Hagras** [0:18:29]: "the low-hanging fruit for this would be definitely credit decisioning, because everybody would like to understand why this kind of score is like this and what it would take them to improve it."
- **Hani Hagras** [0:18:58]: "People cannot afford actually to go to the bank for two or three hours and lose so many times of the day to see somebody on something that can be done on their mobile phone."
- **Daniel Fagella** [0:20:38]: "Yes, it's still going to be related to financial services, but there's a big name change ahead."

## Transcript

**[0:00:04] Daniel Fagella:** This is Daniel Faggella and you're listening to the AI and Banking podcast. Our work with financial services enterprises often focuses on two main goals. One is building or validating a powerful AI strategy with real market data and intelligence. And the other is finding high ROI AI projects, being able to understand what slices of the business are their AI opportunities that actually have proven results out there in the real world. Which of these sort of potential trends and nascent applications actually has evidence of results?

**[0:00:33] Daniel Fagella:** And while a lot of people are excited about customer-facing applications, a lot of those conversations about ROI come down to fraud and come down to compliance. There's a lot of interesting traction across the fraud and compliance function within banking and within financial services broadly. And that's exactly what this episode is about on the AI and Banking podcast. We interview Hani Hagras, who's the chief science officer of Temenos. Temenos is a financial services software company that's been servicing banks for decades now.

**[0:01:01] Daniel Fagella:** They're based in Europe. And Hani speaks with us about the present and the future of artificial intelligence for compliance purposes. So if you're interested in how banks can potentially reduce risk and where AI fits the bill for helping to reduce that risk, this is going to be an episode to tune into. If you're interested in more about how we track AI ROI across different sectors, if you're interested to see how we visualize that data and help clients make high ROI decisions about their AI initiatives, you can learn more at our AI Opportunity Landscape page.

**[0:01:30] Daniel Fagella:** You can find that at emerj.com slash fs1. That's F as in financial, S as in services, and then the number one. So emerj slash fs1, and you can see a short two-minute video about how we visually map the AI ROI space for financial services and learn more about our core service, the AI opportunity landscape. So without further ado, this is Hani with Temenos here on the AI and Banking Podcast. So, Hani, we'll start things off by just getting sort of high-level perspective on where you really see artificial intelligence playing the biggest role in compliance and banking today.

**[0:02:13] Daniel Fagella:** How you like to summarize its role in compliance to kind of a business audience. You can kind of take it away with where we stand today.

**[0:02:19] Hani Hagras:** We have just basically to step back a little bit and see the necessity for AI in general in a sector like banking. So we are dealing now with a completely new era, which is characterized by huge amounts of data, which can be easily generated from the different kinds of transactions, the different kinds of banking activities, the different kinds of sources that we can glue together about any kind of single transaction or any single customer in the bank.

**[0:02:46] Hani Hagras:** This is coupled by very huge advances in storage of data. Now it is very, very cheap to store huge amount of data. and also we have got very efficient databases which can easily organize such amount of data. This is all coupled with very cheap computing, which is getting more power for every day and even distributed in the cloud. We have got also the customer push for more interconnectivity. People are being used to have a lot of things interconnected in their life and you don't want banking to be missed out of this.

**[0:03:17] Hani Hagras:** So there is a need for a technology which is enabling us to source the relationship between thousands and thousands of features and try to get from them a meaningful model that can help us to personalize to the customer needs and enable us to maximize the banking self-interest in terms of better risk management, better fraud management, and maximization of profits, more sales, and also making sure that the customer is also very well taken care of in terms of financial well-being, better services, personalization, and so on.

**[0:03:53] Hani Hagras:** So the only technology they can sort this is artificial intelligence. The analogy is very simple. As a human being, you can actually sort and do different kinds of calculations on a piece of paper. But when things get complicated, you need a calculator or a computer to do this. This machine or this kind of technology will be artificial intelligence. The problem is with the growth of this, there is also at the back end the need for any kind of model to be compliant because we have that regulation.

**[0:04:19] Hani Hagras:** We have got different kinds of laws which we need to adhere to. And with this comes the need for an artificial intelligence system, which is compliant and which satisfies all the regulatory frameworks in the different countries.

**[0:04:34] Daniel Fagella:** Got it. So we're talking about a very broad scale sort of transformation of banking writ large. The audience that you're speaking with right now is that concept will not be going over their head. They're listening to the show explicitly for the reason that they're under the impression that transformation is at stake. So no major convincing there. But what you're mentioning, which is a really important point, is that this big transformative shift with artificial intelligence is going to require a big transformative shift in how all these new processes are now going to be compliant.

**[0:05:05] Daniel Fagella:** Because we don't want to enable a bunch of capabilities that get us, you know, oh, look at all this fancy new technology that's going to get our banking license taken away. Nobody's looking for that. They're looking to have this technology be able to support them and make sure that they stay compliant. So that's obviously a pretty complicated process to monitor all the new compliance laws and regulations. How do you like to sum up where AI plays a role there or how it does?

**[0:05:31] Daniel Fagella:** If you could paint the picture for really the value of AI for compliance, at least today, if we just start with where things are 2019, 2020.

**[0:05:40] Hani Hagras:** So different artificial intelligence systems can be classified into opaque box or black box models, which do very well. and they give us a very high accurate system and so on. However, there is no way to interpret them. And there is a huge difficulty convincing regulation bodies or stakeholders in the bank to say that this is a safe system to employ. So this can be accepted in systems or processes which does not include regulation or need human interference, Like, for example, in speech recognition or machine vision, as long as your Amazon dot is doing very well as to colonizing your speech, you don't really care what is happening in the background.

**[0:06:25] Hani Hagras:** But when it comes to a regulator authorizing a certain lending model to go through, it needs to satisfy different kinds of regulation points and different kinds of requirements. And it needs to be transparent. It needs to have the capability to be easily understood, analyzed, and maybe augmented with human expertise. And when it comes also to the end user, if you have been rejected, for example, a certain loan or a certain financial product, they have got the right to understand why I have been rejected for this.

**[0:06:55] Hani Hagras:** And also the ability for the system to begin having the so what scenario. OK, you have been rejecting me for this kind of loan or this kind of mortgage. What it would take me to be accepted for this? I believe that the way forward is explainable artificial intelligence. And there is a huge push from different kind of governments across the whole world. Here in the UK, the House of Lords select committee on artificial intelligence. They put it in black and white that if artificial intelligence is going to be accepted as a trusted entity in our society, it needs to be intelligible, needs to be explainable.

**[0:07:30] Hani Hagras:** Big consultancies all over the world are pushing for the need for explainable artificial intelligence. and when we speak about experimental artificial intelligence we speak about systems which can easily crunch the huge amount of data and inputs we have got and generate from them very simple systems which can be understood and alive and understood by the business stakeholders and also systems which can easily explain in very easy terminologies their decision and justify them in an easy to understand way and systems which does not only allow us to capture what the data did show us, but also enable us to integrate two sources of knowledge, data plus human expertise.

**[0:08:16] Daniel Fagella:** Yeah, to be able to have some structure outside of just a ton of neural networks and a ton of data pumped through the top, and then no reasons why being able to, like you said, have some of that structured knowledge baked in there in the first place. And that's the sweet spot that so many sectors are trying to get to. It's obviously not an easy problem to take everything. If you were to hold a pistol to the temple of whoever runs the recommendation engine at Amazon and ask them exactly why there are red rain boots on your homepage, they would probably have no satisfying answer for you.

**[0:08:48] Daniel Fagella:** Obviously, that doesn't make sense in finance. What does it take to bridge that gap? In other words, when we look at something like lending and we ask the question, why was I turned down? And what could I do to have this loan actually passed? What do we need to do beyond machine learning, just the pure data and neural nets to actually get that explainability in there? What does it look like to get past the black box a bit practically?

**[0:09:13] Hani Hagras:** Well, look, explainable AI is an emerging area. And it is very important that we'll begin going beyond just saying that artificial intelligence is neural networks and deep learning and random voice, opaque voice model. There is a huge area of research and explainable artificial intelligence. We have been working in this area for about 20 years. And one breakthrough that we managed to do, and we have been granted several patents out of this, is the ability to generate from huge amount of data to so many features, very small amount of if-then rules, which can be easily understood, analyzed by any human being.

**[0:09:52] Hani Hagras:** and by doing this you get the data to speak the human language because if i'm speaking to you in in terms of linguistic labels like okay if the income is low and the time at address is low the customer is likely to be bad you'll be able to understand this kind of rule you'll be able to exercise any kind of exposed opinion on it you'll be able to show to a regulator if you wanted to and if a decision has been taken you'll be able also to easily justify it in a very easy way so I think that there is a need to begin delving more and progressing more into the area of explainable artificial intelligence and not rely on

**[0:10:28] Hani Hagras:** what's called model induction systems which basically still rely on a black box model running in the background and try to interpret it by another model I think this can be relying at the end of the day on a black box model taking the decision and trying to justify it by a way or another this is not like basically making sure that your model is transparent from inside out. And I think from my personal point of view, if we are going to progress more with artificial intelligence in finance and in the banking sector in specific, we need to deal with models which are 100 % transparent, models which give very easy to understand decisions and models which are able to integrate

**[0:11:07] Hani Hagras:** expert opinions plus data-driven machine learning. So we need to get data to speak the human language so that basically both of them can work in tandem.

**[0:11:15] Daniel Fagella:** Yeah, and obviously we are not exactly there right now, practically speaking. As you had mentioned, there's certainly research being done. You've focused on this space for a while to be able to sort of get swaths of data to quote unquote speak human language. To say that that's common in finance today would obviously not be the case. but we've got to kind of, I guess, get there gradually. Is explainable AI sort of obviously in a burgeoning field here?

**[0:11:43] Daniel Fagella:** Is this starting to make its way into compliance in ways that are maybe tangible? What I'd like to try to do is maybe paint the picture of an example use case that maybe a business person can sort of imagine in their mind. Are there some places where this is sort of coming to life today, maybe in a workflow that we could explain or talk about briefly?

**[0:12:02] Hani Hagras:** Yeah, yeah. Look, we have done several advances within our organization, Explainable AI. So we have got a platform, which is actually an all-expermable AI system from inside out. So we have got this actually with customers taking decisions as we speak. If you'd like to have a very glaring example, it will be in credit decision. So we are able actually to take 600 or 700 or 1 ,000 features about yourself, which come from any kind of credit referencing agency.

**[0:12:31] Hani Hagras:** basically glue it with other sources of information that might be available about a single applicant and generate from this kind of data models that are having few rules where the length of each rule is only three or four antecedents although that your models might be using 1000 inputs and basically you are able to generate rules which anybody can understand like for example if the customer income is low, he has been living in his address for a very short time, and he's defaulting in his loans for the past three months is likely to be bad.

**[0:13:04] Hani Hagras:** Anybody will be able to understand this. And if a decision has been taken, like for example, I have rejected this customer, his credit score is 700. Any existing system will just say 700, and then that's it. No, we are able to say it is 700 because of so and so and so and so. And the main drivers are that this guy, for example, has been defaulting on his loan for so many times. He has got a lot of searches against him.

**[0:13:31] Hani Hagras:** He has got some kind of court judgments against him. Therefore, this score is like this. With explainability comes opportunity because if you know why this guy has been rejected, you might be able to feedback this to him. And you say to him, okay, you have not accepted you for this. But actually, if you do this and this and this, this can be very important for you in the future. The same for cash flow prediction. If there is a model looking at your transaction and is able to predict that in the future, in the next six months, you'll be in deficit.

**[0:14:03] Hani Hagras:** And identify why, for example, you have been spending so much in holidays or as an SME lender, you have not been active at collecting the money that you should be getting. In this case, we can give the customer advice and say, OK, we are going to be having negative cash flow. and then because the system is explainable we can begin answering the first question why because so and so and so and then answering the most important questions okay so what can I do in this case you can begin suggesting stuff to the customer like for example you need to be very active at collecting any kind of money that you owe in order to increase your cash flow or you

**[0:14:38] Hani Hagras:** need to reduce your spending here and there or if needed and there is no other way the bank can begin offering different kinds of services like overdrafts or loans or whatever. So this will be

**[0:14:50] Daniel Fagella:** only possible as explainability. Certainly. And it sounds like those are two nice discrete use cases of where the explainability within the rules that might otherwise be kind of black box would clearly have actionable steps. Like you said, with explainability comes opportunity. If we know what's going on, we can actually do something about it. That's really powerful. And In any sector, that's tremendously powerful.

**[0:15:13] Hani Hagras:** The use cases are endless. So basically, if you speak about something like churn, there is no point of me just predicting that this kind of customer is going to churn by a neural network and that's it. What is more important to understand why? And what is even more important to understand what it would take me to keep this kind of customer from leaving? Without explaining that, you cannot do all of this. The same applies for marketing.

**[0:15:36] Hani Hagras:** There is no point of you going and sell a product for a certain customer and say, this is not efficient. What is more important to understand? Why is it not going to take it and what it would take the customer to take this kind of given product? The same applies for fraud. There is no, OK, a lot of black box models are very efficient in fraud in banking in general. But from my point of view, yes, it is good to predict that this kind of customer is going to be fraudulent.

**[0:16:00] Hani Hagras:** But what is more important to understand why fraud is happening? What is the most common pattern of frauds so that we can basically kill it from source? And what is important as well to understand why this kind of prediction has been flagged as fraudulent so that basically we can reduce the false positive rates. So you can see that any kind of use case in banking that you can think about, it needs explainability so that you can do it in a more efficient way.

**[0:16:27] Daniel Fagella:** 100%. And I think obviously there's some use cases where it's, and I have one little very short final question for you, but to layer onto what you're saying, I think in any situation where we could either have no idea why the system is making a decision or have a very clear idea, clearly the latter is a better scenario for business leaders and people trying to actually do something with that. There's some instances where it's absolutely necessary.

**[0:16:50] Daniel Fagella:** So for lending, it's possible that if you can't explain a certain amount of why you said no to these people and yes to these people, we can take away your bank license and then you have to fire all of your employees. That's a pretty big consequence. If someone asks Amazon, why are there red rain boots on my homepage? Amazon actually doesn't have to fire anybody. And so clearly the degree of urgency around what we need explainability around varies.

**[0:17:12] Daniel Fagella:** But as you're pointing out, the more the better if we can find it because there's an opportunity there. Final question is for you, what's required to get to explainability? So you mentioned cash flow, You mentioned the lending instance. You know, it might not be the case that every use case is just as easy to get to explainability as every other use case. What are the commonalities of a use case where you feel like in the next two or three years we'll see real traction in terms of explainability?

**[0:17:41] Daniel Fagella:** What do those processes have in common?

**[0:17:43] Hani Hagras:** Well, look, from my point of view, any kind of process in the back end of banking or the front end needs explainability. and I can challenge any kind of use case in the banks that actually does not need explainability. You need explainability at different kinds of levels. Sorry, sorry.

**[0:18:01] Daniel Fagella:** I'll reframe myself, honey. I didn't mean to say it the wrong way here. Surely everything I think could require it. I think everything maybe should have it in some ideal world. What I was asking was, are there some use cases where it will be easier to get there than others? Like in other words, will there be some use cases that are lower hanging fruit? You're a researcher, so you actually understand the science here. Are there some use cases that are technically more simple to arrive at explainability than others?

**[0:18:24] Daniel Fagella:** Because I think for business folks, that's a little bit over our head, but I'd love to know your thoughts.

**[0:18:29] Hani Hagras:** Well, the low-hanging fruit for this would be definitely credit decisioning, because everybody would like to understand why this kind of score is like this and what it would take them to improve it. Cash flow prediction, all of us would like to know how we are going to do in the next few months, and all of us would like to have some kind of help. And this is very important as a banking message, which is looking at the ethical point of view, that banks should be actually helping their own customers for their kind of financial well-being.

**[0:18:58] Hani Hagras:** Of course, fraud is a very important and low-hanging fruit as well, given the amount of data involved. And the new trend in banking, which is conversational banking. So we are at an era where everybody is basically using their mobile phones, do the vast majority of their own stuff. and people would like to see banking giving them the 24-7 service as if they have got a relationship manager in front of them. People cannot afford actually to go to the bank for two or three hours and lose so many times of the day to see somebody on something that can be done on their mobile phone.

**[0:19:36] Hani Hagras:** In this case, it is very important to have interfaces like chatbots or text messages, which provide an interface for an AI system in the background that understands who you are, what you want, and basically try to give you advice, which is going to be very bespoke to your own given circumstances and make sure that it improves your financial well-being. Again, explain it to be a very important role here. And this is very important for the digital transformation and a very important 200 foot as well.

**[0:20:12] Daniel Fagella:** Yeah, and we're going to be a ways away from having that really robust conversational machine experience, but hopefully in time, you know, the vision that you're painting, I think a lot of folks are really excited to see come to life. And obviously, we need researchers to get there. So I have my fingers crossed about the progress that you folks make in the years ahead. I know that that's all that we have for time, Hani, but I sincerely appreciate you breaking down some of your insights and painting a bit of the future for the audience here on AI and Banking.

**[0:20:38] Daniel Fagella:** So thanks again. You're welcome. So that's all for this episode of the podcast. Would you believe it if I told you that the podcast name is going to be changing? Yes, it's still going to be related to financial services, but there's a big name change ahead. We're going to be doing a little bit of an update episode, and that's coming out this Wednesday. So in two days, this episode goes live on a Monday. So So in two days, we're going to be kicking out a team update episode.

**[0:21:05] Daniel Fagella:** We'll be talking about why the podcast name is changing, what new content you can come to expect here on the show, and also what's new at Emerj, what new research findings have we found. You get to look under the hood a little bit in terms of some of what we've discovered research-wise and some of what we're working on here in the back office. So I hope to kind of pull you behind the curtain a little bit and talk to you about what's going on on the show.

**[0:21:24] Daniel Fagella:** The normal episodes will kick back in next Monday, but I invite you to join us for Wednesday and learn about why the title of the show is changing and what's going to be coming up. So be sure to stay tuned in then and I will catch you for the next episode of the show.
