#52 Artificial Intelligence Part 2 - AI in Financial Service Provision

How to think of AI in the bank and some predictions on how AI will impact the banking industry.

Artwork by Mary Mogoi - Website

Hi all - This is the 52nd edition of Frontier Fintech. A big thanks to my regular readers and subscribers. To those who are yet to subscribe, hit the subscribe button below and share with your colleagues and friends. 🚀

Introduction

We spent quite a bit of time and effort setting the foundation of what AI is and where it’s likely headed. It was important to set that background so as to convey the importance of AI to our generation and to understand whether it’s something that requires your attention or just a passing fad. The summary from the first article is below;

1. AI is not a new thing, it has been in the works for over 70 years since Turing asked the question “Can machines think” and later proposed the imitation game;

2. For the last 70 years there have been incredible advancements in the supporting technology that has led to the current AI boom. These include;

   1. Advancements in Transistor technology;

   2. Advancements in Graphics Processing Units;

   3. The Internet and its role in creating a data explosion;

   4. Advancements in algorithmic efficiency from rule-based systems to the current deep-learning algorithms;

3. The latest models such as ChatGPT4o, Claude 3.5 Sonnet and Google Gemini can be thought of as smart high schoolers. They have developed the following core capabilities;

   1. They are multi-modal i.e. can take inputs in different ways such as audio, visual and text;

   2. They are very good at most standardised tests;

   3. They have their limitations in terms of simple logic tests but they are improving at a rapid pace;

   4. Over the next 3-4 years, there should be a 10,000x improvement in AI Capabilities driven by;

      1. Algorithmic improvements;

      2. Increased computing - US$ 300b cluster;

      3. Unhobbling;

4. Future models will be like smart pHd student and additionally they’ll have the following capabilities. The most recent release by OpenAI of ChatGPTo1 shows us that we’re getting there faster than expected.

   1. It’s anticipated that the next generation of models will have 1 million word context windows i.e. you will be able to give it a prompt that has 1 million words. Imagine the most complex project brief you’ve ever come across;

   2. We should have 1,000 step chain of thought reasoning meaning that the next generation of models should be able to execute complex tasks that have a longer time horizon and must be implemented sequentially;

   3. The next generation of models will be able to use tools. Already current models can do simple web search and can be integrated via API to other apps. The next generation of models will significantly expand this capability. Combine with expanded chain of thought reasoning, this will enhance the ability of AIs to execute complex tasks;

   4. Multi-AI collaboration will become a thing. AIs will be able to work together to accomplish a task. Think of inter-departmental or inter-organisational AIs.

The Competitive and Operational Landscape

Within the financial services industry, developments in AI have to be seen in the context of the latest developments in the global banking industry that will shape how AI will shape the industry. Some of these include;

1. A global shift towards open banking or open finance - The idea behind open banking is explained here. The previous consensus around open banking was that it would unleash innovation in consumer finance and weaken the stronghold that banks have on retail and SME consumers as regards to inertia. Simply, people don’t change their primary bank accounts. This has not played out as expected. Nonetheless, open banking will have different ramifications in an AI driven world. The decision on who owns your bank data and what constitutes your “data” could be very critical in determining where value lies in financial services provision. It could mean the difference between being a dumb ledger and deepening the existing customer relationships to the point of extreme stickiness.

2. A growing trend of government led digitisation - Governments are moving towards digitisation of services. India has led the way with the India Stack and other countries such as Brazil are following closely. Countries such as Estonia, South Korea and China have been making progress for many years with most government services and databases being almost fully digital. Africa is slowly catching up albeit at a glacial pace. In Kenya, E-citizen has been expanding its suite of services and citizens can execute a number of services online. Nigeria also has an E-government initiative that enables citizens to get basic services such as passport applications, getting basic tax information and getting drivers licenses. Most countries in Africa are within a spectrum of initial efforts towards digitising government services and advanced implementation of government services. Rwanda for instance is on the upper side of the spectrum with an advanced digital government initiative. This matters because in an AI world, the more that can be accomplished digitally, the more useful the AI will be;

3. The growth of Neobanks like Nubank and Revolut - In the last few months, both Revolut and Nubank have reported impressive financial performance validating the Neobank business model. Nubanks latest set of results has been especially impressive. The key outtakes were,

   1. A 21% yoy growth in customer numbers;

   2. 88% yoy growth in gross profit combined with a 600 basis point improvement in its gross margin;

   3. A 134% yoy growth in net income combined with a 1,400 basis point improvement in Return on Equity

   4. Impressively, an efficiency ratio of 32% down from 92% just three years ago. I don’t think there has ever been such an aggressive improvement in efficiency by any bank in history. This is proof that digitally native banks have an operating leverage that has yet to be seen in traditional banking;

4. Why does this matter? Well, banks like Nubank and Revolut are digitally native. In a former post, I wrote about how they built their entire tech stack in house including their core banking system. What this means is that their journey to fully integrating AI will be much faster both from a technological perspective but also from an internal skill-set perspective. It will be like turning on a switch. Already Klarna has made great strides towards this;

5. Basel III and the rise of Private Credit - Ever since the global financial crisis, banks have reduced their share of lending to middle market companies i.e. SME’s and mid-sized businesses. This has largely been a result of increasing capital requirements driven by Basel 3. Simply, lending to mid-size markets is very capital intensive given the amount of provisioning that is required by Basel. This has led to a situation where banks have retreated from these markets and the private credit industry has taken over. The IMF estimates that the global private credit market is worth US$ 2 trillion having grown from approximately US$ 500 billion as of only 10 years ago. Private credit markets have taken different shape in different markets. In the West i.e. Europe and the US, it has been in the form of PE funds that have set themselves up to lend to mid-size businesses. In China, it has been in the form of “shadow lending”. In Africa, it has mostly taken the form of alternative lenders such as MFIs and Development Finance Institutions. This matters because if the trend continues, most mid-sized businesses will not be turning to banks for their financing needs and this will change the nature of the relationship between SMEs and banks;
   
   

How to think about AI in your Company

The two cover images try to tell a story about how we should think of AI. In the first image from last week’s article, scientists struggle over decades to build an AI agent. In this week’s image, the AI agent shows up at your company and is onboarded by HR. Just like a colleague, AI is now part of the company and is a factor of production. Think of a new batch of analysts or Management Trainees who have joined the organisation. Nonetheless, these are not just your typical management trainees. These trainees have the following characteristics;

1. They know everything about your bank as they’ve been trained on all the bank’s data from its inception. Think of a trainee going to the filing room and reading every file that has ever existed. All the past credit files, all the KYC documents, all regulatory filings ever made, all client correspondences, all internal memos and all debits and credits ever recorded on your core banking ledgers. Basically, every bit of information that has ever been created by your bank;

2. Not only do they know everything about the bank, the also know everything there is to know about banking. They have read industry reports, academic literature on banking, all the news within the industry and have a good grasp through LinkedIn of all the key players in the market together with their backgrounds;

3. They are polymaths who are excellent in multiple fields;

   1. They are the best programmer you’ll ever have;

   2. They have exceptional product management skills;

   3. They are brilliant at credit risk analysis;

   4. They make for world class data analysts;

   5. They are the most detail oriented compliance people you’ll ever have;

   6. They can scrutinise trade documents at a level that is unheard of;

   7. They have emergent properties such as the ability to think of new ways of doing things in a more efficient manner;

   8. They are great at strategy - No need to hire McKinsey;

4. They can work throughout and are on call at anytime. If you ever wake up in the middle of the night and want to run a few things through them, you can just text them or give them a call;

5. They don’t take leave, maternity or holidays off;

6. They don’t need benefits and incentives such as bonuses to produce great work. They simply work their hardest for you at all times;

These new trainees are digital species that exist within your systems. They are integrated to your core systems, your email service, your digital platforms and all other back office and front office touch points. You interact with them through a multi-modal chatbot. Unlike the current ChatGPT, these AI trainees can manage multi-step tasks, work with tools including your core systems and coordinate with other AIs. Some of the tasks that you could hand over to these AIs may include;

Reviewing my Tech Stack

💡Hi Bank AI, my technology costs are going through the roof. The costs to maintain my tech stack and keep it running are growing exponentially and I’m unable to invest sufficient funds in actually improving overall customer experience. Ultimately, I’m falling behind in the market. I need a comprehensive review of my entire infrastructure with the key result of coming up with proposals and an action plan that would enable me to reduce my maintenance costs by 50% whilst freeing up resources and talent that would help me invest in growth. The exercise should involve;

1. Review all current vendors with an emphasis on any overlap of services with other vendors i.e. redundancies and costs with a benchmarking analysis on market costs of similar services. The feedback from this should be either a decision to cancel the contract, a decision on whether to renegotiate the contracts with detailed negotiating points or a decision to continue with the contracts;

2. Review my technology team;

   1. Review their LinkedIn profiles and give feedback to whether better talent can be obtained at a lower or similar cost;

   2. Set up personalised tests for their competence, share with them and provide me with their scores and a training plan for each based on their performance;

3. Review my bank technology stack with an emphasis on the following;

   1. Cloud infrastructure review - specifically review my existing Datacentre infrastructure with an emphasis on how I can move to a fully cloud based system. This should include insights into what should go on the private cloud and what should go on the public cloud. Additionally, look at my existing talent and recommend the best way to structure my infrastructure team including potential new hires;

   2. Review my investments in core banking and specifically digital banking providers - Given my existing resources, advise on their suitability with input on how best to replace them if the need arises;

The output should be a succinct 5 page document together with a detailed project plan that I can share with my Executive Committee and Board for sign-off. You can work with other AIs but remember not to share any sensitive bank data that would go against any internal or external policy documents.

3.2 Corporate Credit Risk Review

💡Hi BankAI, my credit losses in the trade and manufacturing sectors are going through the roof. My NPLs in these sectors are 25% and I’d want to bring them to a maximum of 5%. The work should entail the following;

1. Review the key drivers for credit losses in the sector not only in my bank but across the industry. For internal data, utilise all past credit memos, client correspondences, industry news and insights and any other data sources that you think may be relevant;

2. Define the top drivers of credit loss in this portfolio and advise on new ways of monitoring these drivers. For instance, if inventory overhang is a major driver of credit losses in trade, propose a mechanism for monitoring this across the portfolio;

3. Create synthetic data and build a data simulation using these drivers i.e. simulate the entire portfolio and the underlying commercial performance for our trade clients;

4. Re-run this model by removing the variables that don’t matter until we get the results we’re looking for. That is a 5% NPL and the key performance drivers and their associated monitoring mechanisms.

In a nutshell, we’ll be giving the bank AI complex tasks that will enable banks to be run way more efficiently and insights to be gathered much faster. Of course, front-facing AI applications will be useful, but the point of the above exercise is to show that internal facing AI will be just as important if not more important than customer facing applications.

Likely Outcomes Long-term Outcomes of AI in Banking;

1. Superstar Dynamics

The growth of the internet era has led to the concept of super-star dynamics. Simply, this is a situation in which the returns to an industry are accruing more and more to the top 1% or even top 0.1% of the talent pool in that industry. This has been witnessed through platforms such as Tiktok, Youtube and even Substack. In the corporate world, the same dynamic is playing out. The top 10 companies on the S&P 500 account for 32% of the entire market cap of the S&P 500 whereas 60 years ago, the top 10 accounted for only 10%.

AI and specifically AI agents will likely replace middle-management and white collar administrative roles across the banking sector. The key skills that will be required in each department are deep domain knowledge, interpersonal skills and importantly, deep domain based intuition. Departmental leaders and talent across the industry will be supported by an army of smart pHd like AI agents. Given that banking revenue is likely to grow further due to AI, then returns will accrue to these select bankers and their value will rise significantly. It will make a lot of sense for banks to start thinking about talent in a post AI world. Specifically this means having techno-bankers who have deep banking and industry knowledge coupled with a sharp understanding of technology. This is already happening, in Kenya for instances, the number of deposit holders has grown by a factor of 14 between 2008 and 2023 whereas the number of staff has grown by a factor of 1.4x only. This will be accelerated even further but just within a shorter time frame, maybe three years.

2. The rise of Tier 3 and 4 banks

In my experience, besides larger balance sheets and wider presence through branch networks, one of the major advantages that the top banks have over their smaller competitors is talent. The top talent always goes to the larger banks since these banks can pay bigger salaries, provide better benefits and are generally a better look on your resume. Smaller banks are therefore left with generally lower level talent. Traditionally, the best way for a small banks to compete is through equity heavy incentive structures that within Africa’s context are not really useful given the fickleness of property rights within the continent. Simply, people don’t really value equity as a form of compensation.

Large banks therefore benefit from the best relationship managers, compliance people, credit analysts and tech talent. With AI agents smaller banks will only need to focus on getting superstar leaders through the door. This can be achieved through appealing to a sense of mission and personal agency that sometimes is lacking within the larger organisations. An equalisation of talent across the industry can then give smaller banks a chance to compete with their larger peers. I genuinely think that talent can be a great equaliser;

This will not only apply to talent but even through the entire vendor stack. Smaller banks will have access to better legal advice, world class tech platforms as the cost of building tech approaches zero and even audit and tax advisory. This should be a boon to the industry and ultimately lead to greater outcomes for consumers.

3. Increased Competition From New Entrants

In the same theme of talent, an equalisation of talent would enable new entrants to compete on a more equal footing with incumbents. Neobanks would access industry expertise in terms of credit risk data, compliance know-how and many other capabilities that incumbent banks benefit from through their workforce. An additional factor to consider is that Neobanks would benefit from being “AI native” from the start. Whereas incumbent banks will have to deal with a very long and drawn out HR restructuring process, Neobanks are much more AI ready. Already we’re seeing this with Klarna which has quickly executed on AI for customer service. According to Klarna, the new customer service AI does the work of 700 customer service agents and has led to a US$ 40 million profit up-lift for Klarna. AI could be the disruptor that enables new banks to have a true differentiation in cost structure and customer service.

4. Improved Efficiency

In my experience of managing a bank, one of the big challenges that leaders face is ‘hidden information’ across the organisation. These are simply decisions that are taken that are largely for the benefit of individuals within the organisation but don’t make sense for the organisation. An example is technology leaders making vendor decisions either due to past mistakes (path dependence) or personal interests. These decisions are justified through obfuscation and technology jargon that is meant to confuse a non-tech leadership team. The outcome of this is usually acquiescence since you don’t want to unearth problems that you are not equipped to deal with. This leads to a lot of inefficiency across the organisation. This similarly applies to other domains such as procurement and legal. Leaders have limited bandwidth to take their time and scrutinise all decisions that are made within the organisation, this leads to a cloud of inefficiency that is managed through more bodies and more committees. AI Agents will be able to unearth such inefficiency leading to organisations running leaner and better decisions being made across the entire organisation. You will simply ask your AI agent to review a specific decision made in the context of internal and external constraints. In fact, if I were a bank board member, I’d be pushing to have a specific board AI that enables me to countercheck the information that management is feeding me in real time. It will force people to confront their own bad decisions that are unanimously swept under the rug.

My experience in leadership always led me to believe that you can achieve a 10-15% improvement in net margins just by cutting out bad decision making.

5. AI Brokers and Lower Margins

Open banking could yet have a meaningful impact on the financial sector. People don’t really pay much attention to their financial lives. We live to pursue our dreams and ambitions and banks are simply a tool towards our goals. This apathy towards our financial lives leads to extreme stickiness in our primary bank account. This has been witnessed across the world particularly in retail and SME banking. There is little to be gained in switching and the effort to reward ratio simply doesn’t make sense. This stickiness is a big factor as to why retail and SME banking is very profitable for traditional banks. Given the customer lock-in and high switching costs, consumers of retail banking services are price inelastic in the sense that you can increase tariffs and still retain them. Open Banking was meant to increase competition given that data was now portable and you could now move to a new financial institution with your data. Nonetheless, the principle that people have apathy towards their financial lives rang true and the work required to actually switch was not worth it. We don’t live to bank. We live to do other things.

In an AI world though, this could change. Personal AI agents could lead to increased account switching. For instance, you could ask your personal AI to help you plan a vacation in a year's time. Your AI could develop a budget and discover a credit card that can enable you to get sufficient cash back so as to cover some of your travelling expenses. Given that this AI has agency, you would just instruct it to do everything possible to achieve your goal and your AI will open your new bank account, switch all your standing orders to your new account and cancel your cards from the previous bank including all the subscriptions that went along with those cards. This “schlep” is part of the reason why switching is so difficult. AI agents will make this effort as simple as clicking a button. This hyper intermediation by your AI agent would then drive increased price competition between banks leading to lower margins across the industry. This would have its own ramifications in terms of strategy. For instance, large banks could retreat into becoming infrastructure players by providing customer facing smaller banks with access to payment rails, KYC and other services via APIs whilst focusing on corporate banking. This is already happening and is partly being accelerated by Basel 3 which was detailed earlier.

6. Bank Tech’s Cottage Industry - Core Banking and Digital Platforms

There exists an entire industry built to support bank technology requirements. The total market size for core banking platforms (the ledgers) and digital banking platforms (which build customer experiences) are roughly worth US$ 40 billion per annum in terms of revenue and are expected to grow to US$ 160 billion by 2031. Banks have traditionally focused on building out expertise in terms of traditional banking i.e. credit risk analysis, relationship management and compliance. They have outsourced the build-out and maintenance of supporting technological software to players such as Temenos, Backbase and Finastra. The underlying framework is that technology and software development are not core competencies for banks and therefore this should be outsourced. It’s a simple and logical conclusion.

Smart pHd agents could confer the following competencies to banks;

1. World class software development capabilities;

2. World class product management capabilities;

3. World class project management capabilities

With these capabilities, the dynamics between a bank and software vendors could alter dramatically. Coase Theorem suggests that firms will organize activities internally rather than outsourcing them if the transaction costs associated with using the market are higher than the costs of organizing the same activities within the firm. Transaction costs in this case could include the cost of talent, the cost of information and the cost of setting up complex projects. The Core Banking industry could suffer as a result of AI enabling organisations to bring in more work in-house. In simple terms, banks will be able to do everything that a CBS vendor does in-house including R&D, software development, tech support and deployment. Gpto1 and Replit are showing us that this doesn’t sound crazy. Traditional bank technology vendors will need to think long and hard about their unique value proposition in a post AI world.

7. Agents as a Representation of your Bank;

What’s becoming evident is that AI agents will be widespread. In fact, AI agents will be added to every organisation’s list of digital assets such as websites and social media pages. There will be at the very least two AI agents per bank. An internal facing one to help people with their productivity and an external facing one to clients. There could be further branching of these agents into corporate banking agents, SME agents, wealth management agents and personal banking agents. The ability to train and improve the agent will be a source of differentiation for banks. Whereas social media chatter is often focused on how bad a bank’s app is, it will soon include how “stupid” a bank’s agent is. I think experimentation around this needs to start happening rapidly because this will be a very interesting aspect of customer experience.

Here’s a video in which Klarna’s CEO gives his views on AI

How to be AI Ready

The path to being AI ready is complex and fraught with risk. Move too early and you could end up creating structures that won’t make sense in a couple of years time, move too late and you could lag behind for years. It’s a nightmarish scenario. Nonetheless, there are some steps that would make sense to make;

1. Internal Competency - Depending on a bank’s resources, it would be useful to get in a senior level AI person at the Executive Committee level. JP Morgan for instance has done this through recruiting a Chief Data and Analytics Officer who reports to Jamie Dimon directly. The main role is to create the capabilities and competencies across data and tech that would allow the bank to go fully AI. Where resources and bandwidth may not allow, getting experts at board advisory level may work. Moving towards building an institution that is relevant in the world of AI seems like it will be a series of many clever decisions and this will mean having the right people in the room when the decisions are being made. A Business as Usual mindset may be detrimental in the long-run.

2. Internal Hack Team - It would be useful to get a small core of smart executives that is tasked with creating the foundations for full AI deployment. This should include members from different departments but should be a maximum of four people so as to avoid the tendency for big teams to align towards the least divisive answer rather than the best answer. That means having compliance, credit, business and tech minded people make up the team. The key questions this team should answer will include;

   1. What will regulation look like in the world of AI. Which regulators should we engage, when should we engage them and how should we engage them? A number of issues arise from this such as;

      1. What are the legal implications of AI - will AIs need to be licensed and attached to a legal entity so that any decisions that are made by the AI can be traced back to a legal entity?

      2. What are the permissible boundaries in which an AI can operate? Should they be allowed to sign documents on behalf of a client for instance? The questions are endless but will need some clear direction;
         

   2. What are the operational consequences of AI? What will consumer journeys look like? How will payment reconciliation look like? What will trade document handling look like in the age of AI. What are the likely implications in terms of organisational structures and competencies?

   3. How do we define AI maturity? Simply, what parameters are we tracking across all foundational models and AI vendors that will enable us to get comfortable with full AI deployment within the bank? It may be that we want specific advancements in the current AI models e.g. it must attain x on this logic test for us to be comfortable in deploying it. Moreover, what is “good enough” for our internal models? This discussion between Rex Salisbury and James Dyett of OpenAI gives a good overview of the steps Morgan Stanley took to have 100% of their wealth management team using their internal AI.

   4. What should our tech and data stack look like in an AI world? This will specifically look at issues such as cloud strategy, API strategy and how data flows across the organisation;

   5. Security and governance - What are the key principles that will guide AI and Data security and governance. This will be more of a “Geneva Convention” type document that lists principles as opposed to tactical information given the ever evolving nature of AI.

3. Getting your data right - I see this as the biggest cottage industry that will develop over the next coming years in the world of AI. AI is very much a garbage in garbage out product. To make the most use of Generative AI, it will be critical that you’re feeding the models with the right data at the right time and in the right sequence. Additionally, given that models just want to learn, then the more the data the better. In my view this will require a complete rethink of how people work with data capture being the primary goal of every task or function that happens within a bank. The aim will be to develop world class multi-modal data gathering capabilities that power a multi-modal AI agent. Getting your data right will be split into two tasks;

   1. Historical Data - How do we collate and organise all the bank’s historical data so as to be valuable to our models? This will mean uploading past credit memos, internal and external correspondences, internal memos, regulatory circulars and all other relevant data/documents into databases that feed into the AI. This will be quite a task across the industry;

   2. Current and Future Data - On an ongoing basis, workflows will need to be redesigned so as to enable the model to be fed with real-time data throughout. This could mean moving work from Excel to say Google Sheets so that all this information is captured by the AI. It could mean creating tools that capture notes across client and internal meetings. Simply, its creating a data-first approach to every workflow that ensures relevant data is being captured all the time;

   3. This article by a16z clarified a good chunk of my thinking. If I were thinking of starting a tech start up, I’d work on enabling institutions embed AI into their workflows. Hebbia has succeeded using this model and SanifuAI in Kenya is doing some interesting work in the space.

4. The outcome will include updated policies across the bank from procurement to compliance to ensure that all workflows are built with a data first approach.

5. Choice of AI Partner - It’s hard to say how the AI model industry will develop particularly as regards to B2B engagement. Currently there’s a race for which Frontier model will become the most capable. The leaders include Open AI, Google, Anthropic and Meta. Different models will pursue different areas for instance Meta will likely focus its AI on consumer applications such as social media and e-commerce. Google and OpenAI (Microsoft) are likely to focus on commercial applications. There’s a choice to be made here that will include frameworks such as;

   1. Cloud competencies - How much work will it take to onboard into this AI model. Microsoft could bundle AI with Azure;

   2. Application Competencies - How much work will it take to switch over applications such as CRM, Document management and other internal productivity tools. Microsoft has an advantage given that everyone basically uses Office365;

   3. Partnerships with Core Banking providers - I think this will become a selling point in the near future. CBS providers will rush to certify themselves as being compatible either with Gemini, ChatGPT or Claude with a view of enabling data from core ledgers to flow seamlessly into the AI models;

   4. Most importantly, I think skill and focus in AI will really matter. OpenAI is winning a number of contracts in the enterprise space due to their ability to focus on AI. The challenge incumbents like AWS and Google are facing is that their sales staff are incentivised to close deals whereas OpenAI are incentivised to ensure AI is deployed as best as it can.

Wrapping Up

If there’s anything to be taken out of the last two articles, it’s the following;

1. AI is real and it will get better at an accelerating rate therefore every organisation in the world will be impacted by this new digital species;

2. With the advent of the internet, we had an initial bubble followed by a more gradual change in the world in which the internet has indeed changed everything. The successful incumbents that managed to embrace the internet and thrive such as Walgreens and Walmart took their time to really understand the implications of the internet. Importantly, they took bold decisions to adjust their business models. AI will require this “Good to Great” mindset, take time to understand (not too long) but act courageously when it’s time to act;

3. It’s important that AI experimentation starts immediately;

As always thanks for reading and drop the comments below and let’s drive this conversation.

If you want a more detailed conversation on the above, kindly get in touch on samora@frontierfintech.io