Skip to content
Enrich

The language model for banking data

Engage

Personalization and financial wellbeing

 

Drive

Insight exploration and marketing

Assess

Creditworthiness and risk

Company

Everything you need to know about Bud.

Our agentic banking consumer agent explained

Jakub Piotrowski, VP of Product, delves into the details of our agentic banking consumer agent: what it is, how it works and its numerous benefits to banks and consumers alike.

Agentic banking at Bud

At Bud, we’ve always been focused on getting the best possible understanding of consumer finances, so that our clients (financial institutions) and their customers can make the best possible decisions. With the evolution of technology, it’s now increasingly possible to make that decision-making autonomous.

Achieving the best possible outcomes for the consumer has historically been a complicated matter for financial institutions. Partially because some mistakes were profitable for the bank (like customers staying in their overdraft for too long), and partly because it’s genuinely difficult to provide meaningful advice at scale.

Agentic banking in action: a consumer agent from Bud

With the rapid evolution of AI technologies and an increasingly sophisticated understanding of data, we’ve been able to build the first iteration of an agentic banking platform.

The first consumer agent deployed is providing a truly agentic experience – and impressive results. It can understand personal financial situations, is trained on a configurable set of instructions and can trigger simple actions to make sure the goals of each customer are met.

The understanding of personal finances comes from our multilayered understanding of data. We not only look at adding categories, merchants or payment processor data but also consider recurrence and other aspects of each transaction. We supplement this with account balances and product metadata.

The objective of the agent is to maximize the return on interest for consumers that have, as a minimum, a checking and a savings account with the financial institution deploying the platform.

This objective, in reality, is a little more complicated – as the key principle is not to make mistakes.

At the core of this capability is a reinforcement learning model, which is trained on a large set of spending profiles. This is done by providing a simulated environment in which the agent takes actions in order to reach objectives, adjusting its strategy in response to a reward generated as a result of taking actions. This allows the agent to develop a complex and nuanced strategy resulting from all of the different situations experienced across the spending profiles.

This is all possible thanks to a reward system where missing a bill payment or going into overdraft was penalized, while generating interest on savings was rewarded. The agent, at present, is able to take three actions:

  • do nothing
  • move money from the savings account to the checking account, or 
  • move money from the checking account to the savings account. 

Analysis1 of the customer base of a US bank indicated that had this agent been running, it would have generated at least $500 in gains over the course of one year for more than 27% of customers. This is a very simple way to increase a customer’s savings without lifestyle changes or a spending reduction. For a low-earner profile, the agent would have effectively protected customers from overdraft fees, resulting in an average of $460 in fees avoided – and, for some customers, avoided fees were in the thousands of dollars. The net gain of consumers in our analysis was strong across low, medium and high earner profiles with respective median net benefits of $135, $157 and $381, and with a headline population net gain of $167.

The customer sees the movement between their accounts, and can at any point stop it or reverse any transaction (or move money between those accounts on their own – it’s fully transparent and accessible). The money is always in the customer’s accounts – and, in the initial phase, those are accounts sitting with the bank deploying the agent (but open banking-based scenarios are fully feasible, too).

There is also flexibility in the model. Consumers can be allowed to not only decide whether or not to use the agent, but also set some parameters such as a specific amount of free cash to always be available.

The principles covered here are applicable to multiple scenarios. We can easily extend the set of objectives and constraints, allowing for a natural language definition of those. Our level of understanding of personal finances means it is straightforward to translate those into a reward system. We can also think of more actions to be taken.

Ultimately, the approach we have taken allows us to extend the agent with capabilities that would cater to a broad set of consumer objectives. Optimizing cashflow, enhancing the credit score or getting the most out of available financial products are entirely feasible.

Next-gen personalized banking

Effectively, what our exciting agent is offering in its initial use case is supercharged payment rules. The fundamental difference is that it can capture and encode complex relationships and dynamics that are difficult to identify, represent and maintain when using a rules-based approach. Improving and extending the system, therefore, becomes easy rather than the increasingly time-consuming process required under other approaches. Also, because of its base in high-quality data processing, it doesn’t require any additional data cleaning or preparation, making it more robust.

But the biggest advantage of the agentic approach is that the rules can be increasingly sophisticated meaning that, at some point soon, we might be looking at models seeking to boost customers’ credit score, consolidate loans or seek the best financial products fitting the requirements or lifestyle of the individual.

What’s more, we’re already looking at more potential applications that will change the way financial institutions and their employees operate.

Find out about agentic banking at Bud



1
Analysis conducted over a random sample of 1,000 accounts from a US bank where the included consumers had both a primary checking account and at least one savings account in the dataset. Low-earner profile was defined as someone with income below the 30th percentile and a high-earner profile was defined as someone with income above the 90th percentile. All numbers rounded to the nearest dollar. All numbers presented are related to gains/impact seen on customers after the simulation of running the agent for one year.