What you'll Learn:
- Marketing with AI: Achieving a 736% ROI for an international bank through predictive customer targeting.
- Digital Bank Creation: Engaging over 10,000 new users with a 90% adoption rate, reducing operational costs by 40%, and increasing transaction speeds by 50% for freelancers and small businesses.
- ID Verification Simplification: Helping a large bank streamline secure document management, enhancing customer experience and security.
- Smarter Mobile Banking: Developing a credit improvement app for a credit company, reaching 100,000 users within 9 months.
- Intelligent Call Routing: Enhancing call center efficiency with AI for a large global bank, achieving a 95% prioritization accuracy.
- Customer Targeting in Turkish Bank: Improving SME product response rates for a prominent Turkish bank using advanced data analysis.
Transcription:
Chris Brown (00:10):
All right. Good Afternoon everybody. Welcome to the session. Thank you for being here. It's 4:00 PM on day one, you are not AIed out, which says to me this is where the people who are serious about doing something with AI are here today right now. So the people that are tired of talking the talk and want to walk the walk and actually implement some AI and get some return of investment for it are in this room right now. So you can have the first round of applause. My name's Chris Brown, I'm the President of Intelligence. I'm here with my esteemed colleagues today. My esteemed colleague and technical crutch I will have is Jonas Da Cruz is here as well ahead of marketing and I'm going to talk, the main discussion today is going through real use cases, use cases that we have implemented in the real world, in the finance industry that have genuinely driven ai, return of investment.
(01:07):
And that's going to be where I'm going to spend most of my time today because I think you can get wrapped up around the axle of all the different layers and pieces of AI, and I'll go into that in a second. But really today's about how do you drive return of investment from AI in the real world. So a little bit about intelligence. We are 22 years old. We have been doing deep tech and automation in the finance industry for 22 years and 10 years ago, I sat in a room with HOAs and we said we should do AI. And that was 10 years ago and it was super early and we have done the whole journey of being the weird kids selling witchcraft and black boxes 10 years ago to suddenly becoming mainstream and trendy and no one's ever called me mainstream and trendy before, but that's where we are today and we've been through that whole journey and that's put us in good stead because we've learned a significant amount along the way over those 10 years, and there's very few people in the world right now that has got 10 years experience of actually walking the walk and implementing AI in production in real life.
(02:17):
Underpinning key processes for an organization, especially in the finance industry. I want to talk about applied ai and what do I mean by applied ai? It's super simple. I mean AI that underpins any real world challenge and solves that problem in the real world. We don't care if it's LLMs, ltms, CNNs, whether it's supervised learning, unsupervised learning, we really don't care. What we really care about is does it create return of investment? Does it give you operational efficiency? Does it create new markets of growth? Does it change the way you can engage with your clients any way that return of investment can be clearly measured from an implemented technology? That's the AI we're interested in talking about to our clients.
(03:10):
All of this, right? And this is just eight elements. There are thousands, not thousands, but there's lots, many, many different layers of AI that you can talk about. Maybe there are thousands actually, and you can quickly get lost in your whole conversation around what model am I using? Is it model as a service? Is it open source? Am I buying it off the shelf? Is it this architecture? What concepts of streaming am I using? What data? There's a whole load, a whole load of layers that you can find overwhelming whether you are in technology or whether you are not in technology. And I want to put the argument forward that either you're a data scientist or an ML ops engineer and all of this stuff already, or you're not a data scientist, you're not an ML engineer or have any ambition to be one, and therefore you shouldn't be worrying about this and leave it to the people that know what they're talking about and help them understand where they can apply their skills within an organization, within an industry, within a business to really drive some value from the implementations and the skills that they have from all of their training and expertise because this is still true.
(04:33):
This was still republished. Again, we've used this number. It hasn't grown over the last three years. This has Gartner numbers. 85% of AI projects still today fail to bridge the chasm from innovation lab or from proof of concept into production, which means 85% of AI projects, despite the investment that's going in, are not generating any returns for the business. And to be inside that 15%, the good thing is there's 15%, which means some people are doing it right. It is happening, it's real, because if that was 0%, then we are really in a hype cycle, right? And we've got a major, major problem, but there are 15% of projects that are making it into production, and I'm going to give you some clear examples of that today. I'm going to give you three examples. That's all we've got time for, but I can sit and talk to you about 10, 20.
(05:30):
Many, many examples of production AI. It is happening. You can get into successful production with ai. It's happening today, it's happening in your industry and it's entirely possible as long as you focus on the right things that will allow you to get your business comfortable and understand the change management as well as the technology that goes with the deployment of artificial intelligence because it's very, very different to any other kind of software project that you'll ever take on within your organization. I said we'd spend 20, I'll just give you a quick anecdote. I said we spent 22 years doing deep tech and automation projects pre ai, and we still do a lot of non-AI projects today, quite honestly in the industry. And we know at the start of a project, if it's not AI, we're going to deliver that project. We might be a week early, we might be a week late, but we're going to deliver that.
(06:20):
It's binary. And I think that was said in the keynote this morning, right? In terms of the predictability of software, non-AI software projects, it's pretty predictable that you're going to deliver with AI at the start of every project. The answer to nearly every question that you ask, the answer is, I don't know. Do I have enough data? I don't know. I haven't seen your data. Do I have quality in the data? I don't know. Do I have correlation within my data? I don't know. And so on. What accuracy levels can I get? I don't know. What accuracy levels do you need? I don't know. So there's a lot of unknowns that you have to remove from an AI project and how you engage and how you go through that rapid experimentation to risk and to mitigate risk of time in both time and money spent is really, really important.
(07:06):
And that rapid experimentation and how you engage through that project, that's how you get into these 15% pieces because the only AI solutions that matter in our eyes are solutions that drive return of investment. I've spoken to many, many, many financial institutions CEOs over the time that we've been doing this, and I can tell you more or less, none of them care what model you used, what training you used. Did you use use unsupervised or supervised learn? Did you use LSA? All the pieces I talked about before, they just don't care. They just care about how is it going to change my engagement with my customers? How is it going to change my operational efficiency? How is it going to give me an opportunity to grow in a market and create some space and some differentiation? That's really the only thing that they care about.
(07:51):
And I would argue if you are a non-technical person that's leading your business, if you're a CIO, I have a different conversation for you because you've got six things you need to worry about. But if you're not a CIO, you're and you are working out, how can I help my organization absorb artificial intelligence into my business and really drive ROI? These are the three things. These three things are the things I would urge you and argue that are the most critical three things that you can concentrate on in order to get, in order to get AI into production. First one is where to apply it. I'll talk a little, this is a whole conversation in its own right. I definitely don't have time for this and I'm going to fast forward in a second to get to the case studies where to apply it. We've got a quick map and a second that I stole from an amazing map that somebody put on LinkedIn.
(08:38):
I'm not ashamed. It is a fantastic financial institution map of where you're going to apply AI and I'll share it with you and where we got it from. Consumption models the most under talked about subject and so critically important when you're engaging on any AI project, am I going to go off the shelf and just buy it so it looks like software as a service? Am I going to use model as a service? Most friends example is GPT or am I going to go full open source model? And all of those have massively, I know they sound technical, they have massive non-technical impacts on your project that you have to consider from the outset. A lot of them being can my data leave the firewall or not? If you can't leave the firewall, then stop talking about GPT and stop talking about model of a service because you just can't use it, right?
(09:22):
So make sure we're having that conversation first and today's not the day for that conversation, but it's the most undertook conversation in my opinion. And then how to deploy. What are you going to do? Are you going to buy it if it's off the shelf? That's a pretty straightforward conversation you're going to have. It's pretty straightforward decision. It's the only option available. It looks like software as a service. You're going to go buy it. Are you going to build the capabilities of some really large banks within the US or spend billions of dollars to build massive teams in order to go and they're putting a full scale bet on AI as their strategy of the future? Or are you going to partner with organizations like ours and others? Obviously we can't partner with everyone, but how are you going to get that AI off the ground?
(10:02):
Because a conversation to build a billion dollar team, which is what one, you guys have all seen it, right? It's JPMorgan Chase. I'll spend a billion dollars on building an AI team within their organization. That's a hard business case to sign off when you don't know how your organization's going to adopt it. Are they going to understand the change management that goes with it, not just the technology. So they're the three things that I would argue, but I spend too much time on that. But I have to show you this map because I found it on LinkedIn, I stole it. It's amazing. I love it. And it's from a guy, let me just find out his name because I want to give him the credit. It is fantastic. And I love the map. It is from a guy called Sam Bobbe and he put this map up on LinkedIn and I thought, wow, that's amazing.
(10:50):
He spent some time on doing that. We've all got it in our brains. We spent so much time. But he put this up and he broke it down into he has five places. And I'm not saying it's completely exhaustive, but it's a brilliant map, right? He has five places that you can implement AI within your business, whether it's in product management, customer acquisition, you can read them. And then he even went as far as saying, and by the way, here's all the things that I can think about that we can do and I'm going to talk about and what's great for us is we're going to talk about some use cases that sit on this map today that we have put into production for large global banks. And I promise you, they're not in innovation, they're not in laboratories, they're not POCs. These are underpinning critical processes of a bank today or a FinTech in some cases.
(11:35):
I'll let someone take you a picture of that. I'll also somehow, we have to get you the link to this because it's amazing and I love it and fair plate one for posting that. Alright, let's go straight into what you actually came here for, which is the case studies. And I spent too much time on that introduction by the way, you can interrupt me. It's a nice intimate room if you want to interrupt me and ask me questions at any time on what I've said on any of the case studies we're going to talk about, honestly anytime, just interrupt me. It's not a problem at all. I'd much rather have a conversation than my monologue tones coming at you. I'm going to start with this one. This is quite an interesting story. We do a lot of these examples. I picked a particular one out because the numbers are just extraordinary, absolutely extraordinary.
(12:24):
But I'll tell you the story behind this. This was a bank that had a set of products and they had a long list of customers from the CRM. I'm going to try not to bore you on this, but it's really interesting. It gives you an underpinning of how the change management comes around. A long list of clients. They had a million. So they picked a million from the CRM and said, I want to take a short list of this. I want to create a list of 300,000. We're not involved at this stage. I want to take a short list of 300,000. I want to do an outbound campaign on this suite of products and I want to know what I sell. So they did the whole campaign. They came back and they made about 8,900 sales from 300,000. From 300,000 outbound campaigns about two to about 3%, something like that.
(13:07):
It was just under 3% and they were okay with that. Wasn't amazing, wasn't great. And they asked, well, can we have a look at our data to understand if we can accelerate what we can sell from a campaign point of view and can we use AI? And I said, I don't know because it's the answer to every question when you stop, but you've got a hell of a set of training data for us to use because you just ran the campaign. So why don't we take a look at the data, select another and they said select another 300,000 and we'll go after them. So yeah, we'll put a budget together, 300,000 end users and can you give us the, sorry, we'll give you a million. Give me 300,000 budget for 300,000. I'll give you another million. And our clients are really big clients and we're not very big and we don't like giving them bad news, but we trust in the data.
(13:59):
And we went back to them and said, I can't find 300,000 people within this list of a million that I think have a high confidence that we're going to go sell to or that you are going to go sell to. We can find 156,000 but I can't find 300,000. And we gave 'em the news and we stuck by the data and they went out and said, right, well we'll take the 156,000. And they, they did the same campaign on the same products but with the recommendations that we gave them on the products and they sold just over 30,000 products from their 156,000. So it wasn't quite four times the sales with half the amount. So it wasn't 800% increase, but it was an enormous increase by just looking at patents in profile data to understand who should I be targeting? Who should I be looking at?
(14:50):
How will I go? And I just had a long conversation with a guy called Eric from First National Bank and he was talking about how do I get my bank comfortable and safe in if I do this in operations, if I do it in fraud? And I'm going to talk about some operations in fraud pieces where if you're not super comfortable with ai, it's difficult, right? Because they're really conservative areas that are going to take a lot of change management and confidence management in order to get it in. But in here you can make small differences that make massive amounts of impact and not necessarily going to cripple your bank if you put the right risks and the right risk mitigations in place for this and they're really, really good places to start because you can have high impact really quickly. So that's the first use case I wanted to start with.
(15:34):
It's super impactful. It's not the only one. We've done a lot of product recommendation, a lot of lead scoring, but this one was in particular a very good set of numbers. They're not all at 736%, but they're all up there in the multiple hundreds. You are getting multipliers from just looking at patterns of data and what it tells you is there's strong correlation in your historic sales patterns within finance institutions that if you apply the right techniques, you can uncover those and you can really make massive impact on what your sales teams can achieve with the same resources. Any questions on that one? Alright, I can just move on. None. Okay, I'll go on to the next one. This is a big topic. Everybody talks about it and the reason everybody talks about it is because it just so happens that non rule based fraud detection, looking at patents and data is order of magnitudes better than rules.
(16:32):
It just is. And it doesn't matter if you're using us or somebody else. The fact of the matter is that underpinning technology that can detect this is just much better. We're obviously much, much, much better, but the technology itself is much better for this. We always really start in a similar place, which is organization. This is so to be really clear, this is about transaction fraud. There's many different types of fraud within financial institutions. This is not AML. We do AML projects. This is about pure transaction fraud and detecting that fraud. And it always starts in a similar position, which is we have a rule-based engine in place. It can be IBM, saver pavements or something else that's quite industry standard. And we want to look at either augmenting that rule, so using artificial intelligence to import some knowledge into that rule set and then continuing to use the rule set. And then others that are super brave are, I'm just going to take it out and I'm going to go for it and let's just steady, let's see how we turn the dial. We can talk about switching switches and turning dials on AI in a second.
(17:47):
The important part about this one, and it's going to play its part in the next use case as well, is data. The effect of data. So if I look at the previous one, if I go back to the previous one, this one that was model heavy, this was about the model, this is about training the model. This is about having great, of course you've got to have good data, but this is really about how do I train a really good model. These next two that I'm going to talk about this and I'll spoil the next one, are really about, it's not a hugely difficult model problem, Hans is going to argue with me in a second, but relatively speaking, but the pre-processing and the creating of features and having rapid feature sets. So creating feature stores that you can rapidly update and make sure that they're streaming.
(18:38):
Where I was just at a lunch earlier where they were talking about quality of data and AI streaming. If I was talking about that, that other three side of the CIO coin when I said these are the three things for non-technical, if I was going to talk about technical, it would be data-driven architectures and AI streaming and having feature stores. So taking raw features, then adapting those raw features that you can only do in inference and then doing features that are historical patterns as well. In combining all of those, it is that work that is the real key aspect of ensuring that fraud management can take place to these levels. And when you have a rule-based system, you're always battling, and you guys probably know this way better than I do, we learn a lot of things along the way, but you guys are the experts.
(19:24):
You're always battling this fixed axle of true positives go up as your false positives come down. And it's really quite a fixed axle like this. And what AI does is it breaks the axle, it allows you to get away from the rule base, it allows you to break the axle. And what happens when you break the axle is you can either increase your true fraud detection while reducing your false positives or you can, depending on what your economics look like, and this is again goes back to that change management. How am I going to deploy the technology? I've got choices to make, right? Am I going to increase my true positive detection or am I going to just hold my true positive detection where it's at and reduce my false positive all the way down? And the reason we talk about that is in this case, this is a bank in the us, a large bank in the us, it's in production and the fraud detection rate we increased by 56%.
(20:23):
That translated to a 34% increase in I should say true positive value, right? Because what happens is as you go up the stack, the value of those transactions starts to reduce. So even though you're getting 56% increase, you are only getting a 34% increase in that true positive value because the big transactions down here were already being captured anyway. That was worth in itself tens of millions of dollars. It was great, right? It's awesome. It's a lot of money, tens of millions of dollars and there's no way they invested that kind of money in this solution. The false positives is in the multiple hundreds of millions of dollars. When you reduce it by 72%, it is worth so much revenue that flows through the system and less customer irritation because they're not getting declined by all of the transactions. And that's why I say now you've got a choice because your false positive reduction for every percent is worth 10 times the worth, 10 times the true value up here.
(21:24):
So where are you going to go? What does safety look like? What is the risk profile that you want to take? And now you've got to tune your engine and you've got that massive fine tuning capability that you can start to draw confidence levels up and down within your AI solution in order to get this in the shape that you wanted for your organization that fits your revenue and risk profile. That's really much more difficult on a rule-based engine where the axle is fixed. And so this is why AI fraud detection, especially in the transaction world, I don't think you'll go to any conference in finance for the next five years where this is not talked about because it's just a massive revenue driver and it's just a fundamental order of magnitude change in capability with the technology.
(22:14):
I don't assume. No questions. Any questions I can see? No. Alright. I am going to skip that one. Oh no, I'm not. I'm going to keep that one. We'll keep this one. This one's interesting. I've only got seven minutes, so I can't do them all totally different. Absolutely. Massively different concept. And I want to use this one because it's away from transactions, it's away from patent data. This is an image problem and you don't get many image problems to solve within banking that we find, right? It's not a huge amount of image challenges. And this one was really interesting because you think it's an image problem, but it's not. It's a lack of data problem. Because what this solution is is about how do I scan and read automatically official documents? So whether they're passports, whether they're driving licenses, how do I read those? So if you've ever used an application where you're holding the document in front of you and you're scanning the document actually works pretty well, OCR works pretty well there because it's like no, no, no, move it to the left, move it to the right, move it up and down, more light, less light.
(23:22):
So it's guiding the user until OCR can actually actually function. But where that solution is not the solution that's implemented where it's about static images being sent in, it's not just finance institutions, rental cars, send it to insurance companies. There's a whole place where you want to read national documents, passports, official IDs, driving licenses, et cetera. Then OCR fails because the image is a little bit to the left or to the right. It's overexposed, it's underexposed, it's blurry, it's not working properly, and you get a big failure rate in OCR. And this is a big process on onboarding and banking and other places with loans and all different kinds of, again, the processes better than I do in terms of where you're using this with AI. This problem is really easy for the model. It's not a difficult problem. The problem is you can't train it, you can't train it because you can't store a million IDs, you can't take a million Florida or Californian driving licenses or Texas or any driving licenses or passports.
(24:30):
You're just not allowed to have them. So the interesting thing is the documents themselves, the format is public data. So the format of those documents and how they're structured is public data. But to train an ai, we need the data and you can't have it. So what we did with this one is it's all about data. So we started to create an entire solution where our data scientists created first of all, tabular data, tabular data that fits within the real world capabilities or the real world look and feel of what should be in those fields. So a driving license number looks like a driving license number and an address looks like an address. It might not exist, but it looks like an address pictures. So you randomly start generating images and pictures of people and you start doing dates of birth and every field on there we create in the tabular data and then we run a complex and ask Hona about this.
(25:27):
Don't ask me because it's super complex synthesization process to go from tabular data to actual real world but not real world driving licenses or passports. And then we've got these perfect data pieces, but that's not enough because I just said before when people take a picture of it, it's underexposed, it's overexposed, and now we just generate noise across the whole of the data set. So we create over exposure, under exposure, rotations, blurring, just general noise, poor quality, camera, low light, all of that good stuff. And then we start to train the ai. And once you've got all of that data and you've solved the hard problem of the lack of data, now you've got an AI that's very, very accurate in being able to uncover all of the data within these pieces. And this is a project where we did unusually for us to be totally honest, we've got it as an off the shelf, an off the shelf piece of capability where we can sell it on a use basis.
(26:33):
Nobody buys it so much, I'll be totally honest with you. But what tends to happen is banks and financial institutions want to solve their problem themselves, but what they can't do do is what they haven't got is the skill to synthesize the data. So they ask us to synthesize data for them so they can then go and solve the problem themselves. So just a completely different look and feel going back to that map that Sam put up there of thinking about where can I apply artificial intelligence within my business? And obviously the possibilities are pretty broad and wide. I wouldn't say they're endless and they're not, but they're broad and they're wide and that is where business leaders have to really be thinking about, if I'm going to really stop talking on the talk and walk the walk, where in my organization can I drive genuine change from a return of investment perspective in order to apply artificial intelligence in the real world? And that's what we've been doing for, like I say for 10 years, we stole a march. We went through the whole process of being the geeky kids in the corner, selling witchcraft and black boxes and through to the position that we're in now of, look, we've been doing this for 10 years. We understand what it takes and we know what the failure rates look like and how to avoid the 85% chasm of doom as you go through that entire process.
(27:55):
I'm going to stop there today. I'm going to stop there because I've got two minutes left and I'm hoping that we might get at least a couple of questions. I could do that one as well, but I won't bore you with another one. I'd love to hear some questions or some challenges to anything I said. Any technical questions upon us is in the room. I brought 'em for a reason, so feel free to give him some questions as well. But the go on and I'll stop.
Audience Member (28:24):
85% what you define, is it something that still in production today or that something product
Chris Brown (28:33):
Still in production? That's a good question actually. They're all in production. I was fast to answer that. I would say if it worked once, it depends how much money it made really, isn't it? If it made billions of dollars and it worked once, it was pretty good in production, I don't understand. I don't. I was faster. I think longevity is the real answer. I think there's probably cases where you might want to hit a problem once and it's a one-off problem. You hit it. It was worth a lot of money in that that you can class that as success. That's repeatable though. That's absolutely repeatable that it's a one hit marketing campaign, but the process is completely repeatable. So I'm going to go repeatable still in production today, but I think there's probably some cases where one-off hits are probably good enough value for you to consider them as successful. I wouldn't want to do one-off fraud detection. No, that feels like you might need some longevity in it. Any other questions? Well, oh, one more here. Sorry. Guidance
Audience Member 2 (29:41):
To give with build versus buy versus partner.
Chris Brown (29:45):
Yes. I've got lots of guidance to give about buy versus build versus partner. So the question, just to repeat because I asked was if nobody heard was what's your guidance on buy versus build versus partner? Start with the consumption model. How are you going to consume AI? There's a big map of artificial intelligence that you can buy off the shelf today. I think the answer to that question is really straightforward. It's the only answer. You've got to go buy it if it's suitable for the challenge that you're trying to solve. We use off the shelf AI today. The example I always use because I love it, is 11 labs.io. We used to spend weeks and weeks and weeks on actors voiceovers on marketing videos or training videos in 11 labs.io. I type in my script, I can choose a voice. It's extremely authentic and it's really cheap.
(30:33):
It's really fast. It's off the shelf. You don't need, you don't need customer ai, you don't need organizations like us to help you. And there's a map of amazing AI out there that you can just consume like software as a service that's buy from a partner and a build. It depends where you are on your maturity scale. Because to build a team, we've been doing this for 10 years and we turn up on day zero knowing exactly what it's going to take. I would argue it would take the best organizations, the very best organizations two years from scratch to even get yourself in a position where I think you've got enough skill and capability to put AI into production. I honestly believe that. Obviously I'm a little bit biased because I can't say I'm not, but putting my least bias head on, I think it's going to take you two years.
(31:25):
It's going to take that length of time to understand what your processes are, how you're going to embed that in your organization, what frameworks you're going to use, what orchestration you're going to use. Just there's a whole plethora of capabilities that is going to take you a long, long time to do. I think if you are AI forward as a strategy, you should start thinking about build Again, I'm saying I'm totally unbiased. You should start thinking about how do I go about persuading my organization to build capabilities within my organization to make that happen. While you're on that journey, I think you have to partner because I think the cost and the time and getting the confidence within your organization and all sorts of stay up with the competition. This is happening. There are 15% of organizations that are succeeding right now, and most of them, a significant amount of them are in the financial services industry. So you can't afford to wait two years. So maybe you want to think about a dual strategy, but for specific cases that you want to go and do now, I think you need a partner. Unless you started two years ago to build your team. If you haven't started now, you're too late. You need to start partnering right now and thinking about building for the future. If you've got an AI forward strategy would be my least bias, two p worth.
(32:49):
Any other questions? Wow, you did amazing to be here. 4:00 PM I can't ask for any more than that. I am out of type. Oh, I think red means I'm two and a half minutes over time, so I apologize about that. And you might have somewhere else to go. Thank you for coming to join us. We're here for the next two days. HOAs is here if you've got technical questions. I'm here from change management implementation, questions about partner buy, build consumption methodologies. I'm here. It's a whole conversation in its own right. But once again, thank you for joining us. I really appreciate it. My contact details are up there and enjoy the rest of the conference and I will see you around in the bar maybe.