How to use AI to promote financial inclusion (S4, Ep1)

Hello and welcome to the Innovation and Financial Inclusion podcast focusing on innovative and emerging practice in financial inclusion. I'm Paul Vic and I'm from Community Finance Solutions at the University of Sulphur which is a research unit specializing in financial inclusion and community finance. Now in this season we'll be exploring how we can use artificial intelligence to promote greater financial inclusion. We'll be looking at how different organizations are using AI to improve financial access and outcomes for financial excluded households. In this episode I'm doing by Munga Chinika to discuss the potential of AI in promoting financial inclusion. Munga is a policy analyst in the capital markets and financial institutions division at the organization of economic cooperation and development and he's currently on synchronement from the Bank of Sanvia. OECD is an intergovernmental organization that promotes economic and social well-being worldwide. So Munga, perhaps we can start by talking a bit about the background for artificial intelligence. So what is artificial intelligence? How would you define it? You know what are some of the main forms or types of AI that exists? The OECD has a formal definition for AI and this is described as a system as machine-based and even explicit or implicit objectives. It can make determinations or viruses from data that's important to the system and then in container it outputs this can include predictions, content. I think in our daily lives we see a lot of this maybe just that we never used to apply the term artificial intelligence but it is there. So even those algorithms that Google uses to print search without the recommendations, I should like cooking or baking for instance, you're looking for a recipe. There's AI driving that so the input is what you're requesting from the system and then the output will be what the system gives you back. And what also because obviously you refer to Google and I guess a lot of people will know about I guess the most obvious example from most people will be chat GPT because that's a large language model so are there other types of AI applications? A lot of different types, two NIME estimation but the main ones are you've got things like regression that's used in econometrics and economic modeling although the magic happen is under deep in the machine usually don't see the actual formulas or mathematics based on on which the AI is based. You've got machine learning as well in HR recommendations and then you've got things that you've mentioned very well, large language models so dealing with the language and interactions and so on. You also got generated AI which can craft completely new content based on the input data and how it's been trained but those are just the main different types and exist. And you've been recently been doing some work around AI and financial inclusion so there's a recent report called harnessing AI in finance for financial inclusion in Africa so can you tell me a little bit about about this work? That's also the OECD published it's in inaugural African capital market report and I contributed the chapter in the reports discussing financial inclusion and AI across African markets so we're exploring first of all how capital markets the capital market is in Africa and then we focused on how AI applied to capital markets and their understanding in the Africa context. I hope AI can be applied to deepen the capital markets as well as expand or broaden financial inclusion. The chapter that does a good financial inclusion in a situation where thin the broader market development goes and then also emphasizing what gains can be obtained from artificial intelligence applied to digital finance. On a practical use case what I can give an example is there is mobile money widely used in Africa this would be equivalent to mobile banking perhaps in the United States or the UK or France and the United EU so it's mobile money this is mean you also due to legacy technology and infrastructure and also ease of access and also when you look at it's just what is suitable for the environment so most of these mobile money service providers do provide let's say small loans micro loans to people so you look at the context of Africa it's challenges with language they're challenges with digital literacy and so on so imagine someone in a rural area trying to access some credits to finance their business or their personal needs because they're not directly in the system so it'd be difficult to score them and grant them the credits and this is where the magic of AI financial inclusion comes in so based on how they use their data so they're using things completely linked away from from money but those that data is a profile of them and they are able to access credit so their live models being used for credit scoring using AI the more you purchase airtime the more it builds up your risk profile and also the amount you borrow and watch you hear how long it takes all these are fed into this AI model that can fit in 10 grades the customers accordingly so if a customer is consistent over time let's say for three months four months the system itself would even recommend to say hey your limit is only a thousand the joint go up to 1,500 more favorite potatoes so that's one area another area is in also as I mentioned digital literacy another area is combating fraud there's a lot of social engineering taking place and I can say prior to the role that I've got it being two roles back I was leading the compliance function in empty and mobile money was empty and fintech as it is called now we have saw a lot of crime using digital financial services lots of fishing vision and so on and we were able to apply AI for fraud prevention and it worked very well let's say from year by more than 70 to 80 percent of instance where it used AI is actually working positively forward within the digital finance piece you mentioned there are a few positive applications of AI some opportunities associated with that that the assessing creditworthiness and that using new forms of data to assess risk and so on and then forward prevention and detection what are some of the risk potential risks associated with AI from talking about preventing fraud and reducing incidents on the other hand as well AI is also being used by malicious actors to find new loopholes of new ways over undermining safe guns or controls as mentioned there's generative AI now it's very easy for a malicious actor to build a model that even mimics the familiar presence of voice so I'm so on this can lead to enormous losses also in terms of KYC AML and so on so AI models can be used to create fake profiles to circumvent controls in place and then there's also the digital literacy mentioned earlier it can undermine public confidence in certain systems and even lead to system instability so when you look at I'll give the mobile money example again if you look at mobile money or digital finance I'll give an example of Zambia which is home if an AMI model is deployed with malicious intent and spread the message throughout for all users that your your custodian of your funds or the service provider is crashing get your money out it's not systemic risk because it's individual accounts yeah assisted risk comes in now from the consumer perspective in that if they run on any institution immediately and it's digital if you have a million wallets or two million wallets at the same time trying to withdraw money it does become systematic because the agency network would shut down they wouldn't have enough liquidity in terms of cash for people withdrawing so those are the interesting interactions something that is non-systemic risk and becomes systemic because of the amplification of the risk profile to AI do these risk make it harder for financial institutions to to adopt the AI because obviously there are as early as you said that there are lots of opportunities to say it's a bit AI but what is holding financial institutions back from adopting AI to greater extent what are some of the challenges they face? So, majority of institutions when we talk about strictly from the financial perspective they are wary to adopt a by mainly because of the risks mentioned. And then also when you look at it from an economic or financial perspective, the cost of developing these systems is not low. Even if you're getting something off the shelf, there will still be a cost student. And this cost includes time, there's a learning curve that will need to be applied to you amongst the considerations also be when you look at cyber security and data privacy. So for financial institutions, data is highly priced. So we're using an AI system also could undermine your own models is the model explainable. Is it interpretable? And then that's the cost element. There's also the regulatory aspect. Or more financial service providers come from a heavily regulated industry or sector where the supervisor authority wonder, cannot explain, cannot understand or interpret the model. And the AI app also cannot explain the model. Maybe they don't even understand it themselves. Feeling with consumer data and finances, it presents a real risk. Also look at finance as a sector is becoming more interoperable. You've got FMI's, you've got banks, you've got insurers, you've got even governments linked to the same network. So where AI is poorly deployed in one instance, it could have a ripple effect to the entire system. And as mentioned earlier, even in the case of mobile money, there is that risk of contagion just to add on to that. Privacy as well, data privacy. So what an AI system even if you are it developed in house, there'll be some repository online or there'll be some linkages to other systems. So if you don't have either good controls, that's the data privacy risk. Also, privacy linked also to security, someone accesses your system and unauthorized access to your records. It could bring the whole business cashing down, not to mention also the fragmentation and international regulations and laws. You've got AI, you're using a certain system, you're based in France, you're providing services to users in Canada or South Africa. This jurisdiction takes precedence because what can be allowable in one jurisdiction might not be permissible in another jurisdiction. And then the end use are also they've been adequately informed. Your data is residing in a server in Croatia, or a service provided in France for someone domiciled in South Africa. We create some interesting considerations. But those are the key risks, the keep the areas. And we can't run away from also not just the economic cost, but even the human capital cost. Because you need ID skilled people, this is a very fast and changing field. The technical expertise and know how is also required and it doesn't come cheap. I think people were being lured by meta form hundreds of millions of dollars, just to quit their jobs and doing them. I guess it's a different way of doing, like you talk about the mobile money aspect, which quite often involves new actors that would not been involved in financial service provision. So the mobile phone networks might be directly involved or involved in partnership. So yeah, it's sort of, it's potentially changes the makeup of the financial service providers. So that's interesting. Of course it's got potential or capacity to process vast amounts of data. And also to use data, great responses at the rate that the human being cannot possibly be expected to. So this also leads to bias. The system could develop bias based on the data is being trained. So to give an example, for consistency in the same kind of scoring system, if you took that model and applied it to a user base in Europe, what's been trained on will be very different from the profile of this new set of users, even the use cases for why do they need this liquidity within our many months or weeks, will they repay it and so on? The model will actually get confused on its own and start giving the TLS start hallucinating at the end. To talk about some of the risks and challenges facing fast tuition. So what can be done from a policy perspective to help in terms of the adoption of AI? What does most then recommendations perhaps that have come out of your work or other other areas? Risk and implications. The OECD has got some high level principles on AI. These were previously published and updated in 2024. There's also what's called the OECD recommendations of the cons of artificial intelligence. And that's why both the place of bias because systems can amplify bias, they're trained on wrong data and then they can reflect discrimination. So this is covered adequately in the AI principles. And then their privacy and security concerns as well. The OECD also has a framework for digital security. Which covers general risk management as well as artificial intelligence. You've got the EU AI Act which talks and gives explicit guidance on the classification of AI systems, I-risk, generic and so on. There's also the EU's Digital Operational Resiliency Act. She talks about resiliency, digital spaces. And why it doesn't address AI directly within finance. It does provide a very good legal framework of what to adopt and what to look out for as well as giving guidance from best practice. So speaking of other policy actions, I had mentioned the AI principles and the OECD actually provides recommendations on implementing the principle of through national policies, particularly focusing on investment in research and development, building skills and also infrastructure. Of course, the systems must run in a certain environment. And also in developing these models to ensure at least the explainability and interpretability. Because AI is also bringing convergence of different industries. So there will be more interoperability and cross-ready later sector oversight we needed. Other practical implementations are looking at privacy enhancing technologies with the focus on data, how it's shared, how it's stored and also how it's applied. Consumer protection, definitely. And they have consumer protection guidelines also from the policy. And then international initiatives. I don't know if you've come across the Hiroshima process. It's a process reporting framework for AI, which can complement the injunisticions that already have their policies. It will complement the domestic policies and also encourage best practice. And another good thing about this type of framework is that it's standardized. So if it's applied in one jurisdiction, it's applied the same way in another jurisdiction. And this also creates more favorable networks of the site and provision of AI in finance. Maybe my last point would be also incident reporting is key. Right. So it's not just about having AI in place, but having robust frameworks as well to report incidents. And looking at how other frameworks have applied. Different jurisdictions are taking different approaches. And the rules based should be principle based. But having their framework in place creates a good environment and is good for the governance of AI in finance. That's a good point to end on because I guess also that's as financial institutions and financial service providers use AI. There is no obligation on them as well to understand how AI informs the decision making on a certain level of transparency for consumers and end-users in terms of how they use AI. Because that's one of the great potential risks for consumers, I guess, that it's a bit of a black box. And you don't really understand how it's used or how it influences the services you receive. It's definitely under which also brings to mind the need for also an AI catalog or reference. It would standardize definitions and use cases, the explanations of models because if everyone is on the same page, it's easier to understand, it's easier to scale up, it's easier to provide oversight and to regulate this system. Now I was a fascinating and insightful conversation with Glinger and I feel I learnt a lot about it. AI adoption and the work that OECD is doing around AI, but perhaps there are full things that stand out for me. The first one is that AI adoption brings opportunities to promote greater financial inclusion. This includes enhanced underwriting by using non-traditional data such as airtime and data bundle purchases, fraud prevention and detection and enhancing financial education. The second takeaway for me is that there are clearly risks associated with AI. This includes low digital literacy of users, it includes malicious actors using it to undermine safeguards and trust them for financial systems, but also the limitation that exists in terms of adequate governance frameworks for AI technologies and limitations in terms of the explainability of the models deployed. The third point for me is that financial service providers, including community finance providers, face several barriers in AI adoption. This includes development costs, cybersecurity concerns as well as inadequate AI skills or inadequate inadequacy of skilled staff. The final and potentially most important takeaway for me is that explainability and interpretability is critical for appropriate AI adoption. Everyone, including consumers, financial service providers and supervised retool authorities, need to understand how the technology is used so it can't be a black box. And that's it for the Innovation and Flange Inclusion podcast with me Paul. In the next episode of this series I'll be exploring the role of AI in insurance with Jake Adfield from Fairfrile Finance. If you want to find out more or contribute to the discussion, go to hub.solford.ac.uk/cfs or follow us on Twitter @cfs_sbs. Bye for now.