How the AI revolution gives financial services the chance to work for the average person | George Dunning , Co-Founder and COO at Bud

From branch manager to robo-adviser: the rise of the ‘fintech’

‘Fintech’ is part of almost everybody’s life, in one form or another, today. In my view, its driving force has always been: “How can technology make financial services work for the average person?”

From the shift away from personal banking undertaken in high street branches to the rise of online solutions, fintechs emerged to address the specific customer needs that were once met by named bank managers.

Despite significant advancements in the past decade or so, empowering consumers remains a complex task. While the likes of neo banks, budgeting apps and robo-advisers have improved accessibility, there’s still a sense of consumers working far too hard to navigate the financial landscape.

Many products and services only address one customer problem. “How can I invest my money more easily?” or “How can I stop getting hammered with fees while spending money abroad?” This means that the savvy consumer might engage a dozen different providers to cover their needs for everyday banking, budgeting, savings, investments, credit, travel and so on. And more often than not, even the most motivated of consumers are still left doing the donkey work to establish what the right product, decision or action should be.

Behind the scenes, some might say it has been a golden era for technological revolutions in financial services: the dawn of Open Banking, the great cloud migration, the rise of the challenger banks, the acceptance that new core infrastructure for banks is a good thing.

But can we honestly say that financial services really work for the average person?

My answer: Not yet. But the groundwork has been laid for a complete transformation: AI.

A clear picture of the consumer to unlock personalised value
From the early days of creating Bud, my co-founder, Ed, and I shared the same vision. We realised that if you had your finances all in one place, connected other financial products via API and linked them together, you could have a single system that would know everything about you and connect you to the right new financial product at the right time.

This turned out to be a highly labour intensive journey with countless technical hurdles to overcome. But we realised that the thing that would underpin all of this was data quality. It depended on accurately enriching transaction data to give a clear picture of the customer in order to unlock any personalised value to a business or a consumer.

That’s how our obsession with making best-in-class enrichment models started. What we could not and did not solve for is how to generically interpret that data and link it into any new system seamlessly. Happily, the recent wave of AI provides an opportunity to solve exactly that. Now there is a technology that can understand a limitless list of queries about your finances, or those of your business or customer base. It can give the data back in a readable manner and pass it on to another place. It will create a world of hyper-personalised finance.

The revolutionary opportunity opened up by AI
Most announcements of banks using AI over the past 12 months have been for operational use cases, often to process documents faster. There’s nothing wrong with streamlining business operations, but it’s not revolutionary. To truly transform the financial sector, we need AI on proper financial data: transactions, investments, loans and so on. Without this we’re just tiptoeing around the potential that the new technology has to offer. We shouldn’t run at it blindly, but we certainly shouldn’t avoid it.

What about hallucinations? What about explainability? What about data integrity?

Let’s tackle the first two at the same time.

The problem with financial data is it’s extremely messy and, in its raw state, doesn’t mean a lot to the average user let alone the average computer system. Layering a generic LLM (large language model) over the top of a few billion raw transactions to give you answers on what’s happening with your user base, is a slow, expensive and ill-advised project.

Generic LLMs are not built for processing, understanding and enriching transactions a billion times over. You need to use the right tools for the right tasks. Streamlined financial models can process an unlimited amount of transactions and generate accurate insights that a generic LLM can interpret and play back for any given situation with confidence. The generic LLM does the ‘finding’ bit rather than the ‘calculating’ bit of the task. By constraining the model to specified tasks and enforcing it to play back the information, it’s discerning the insights it gives, meaning you can eliminate hallucinations and give solid explainability.

An example:

Input: How many of my customers are gamblers?

Model response: There are 245,769 customers that are gamblers in your user base. This is determined by all of the customers listed having consistent spending (at least 5% of their income) in the ‘gambling’ category over the past six months. Would you like to get in touch with this cohort of customers?

Data integrity is everything
The time, cost and effort saved by being able to confidently run a query like this and countless others is monumental versus what businesses have today. But it only works when you have reliable data. Even though it seems simple enough, if you were to try and do this query over a raw data set without enrichment, it would be costly, unreliable and inconsistent. Once you have data that is easy for the model to understand, with key insights to quickly interpret, you can confidently string these queries together again and again creating limitless possibilities.

Data integrity is a key constraint that anyone dealing with financial data should have at the front of their mind. Knowing where the data is going, locking it down to a contained environment and only pushing it out from there given the users’ consent. This is where enterprise models like Google’s PaLM 2 become more favourable over something like GPT4– knowing you can lock down the sensitive data, perform analytics on it and interpret the finding with an LLM in one single instance.

A glimpse into the future…
So what does financial services nirvana look like? Well I’m confident that my view will be outdated extremely quickly, but, for the time being:

On your way to the airport your financial app notifies you that, having seen the flight purchases among your transactions, it’s confident you’re going to be abroad for the next week. Given your daily spending habits over the last few months it recommends transferring £1,000 onto a travel card to cover your spending for the trip.

This transfer may put at risk some outgoing obligation while you’re away given the rise in utilities in recent weeks. It recommends to accommodate for this that you draw down on your instant saver account into your current account £200. So that you don’t impact your savings too heavily, though, it can initiate a round-up service so that you can passively increase your savings account back to where it was over the following months.

It has also reviewed the terms of your credit card, informs you that you have lounge access in the majority of airports in the area. It asks if it should book you into the lounge at the airport destination logged in your Uber trip.

All of this could be actioned with one simple ‘yes’. This is finances working for you, not you working for finances.

What is Bud’s role in this? The platform is the layer of intelligence serving transaction and analytical models, building a foundation of accurate data that tailored LLMs can be layered over. Bud’s fintech, neobank and small and large financial institution clients license this technology, taking data analytics work from weeks to seconds and making these outcomes a reality for their customers.

Leave a Reply