Banks have been using machine learning longer than pretty much anyone else, and for good reason. Identifying the lowest risk borrowers, pricing loans, and forecasting losses all require the use
of predictive models. Sometimes these models are simple scorecards. Other times, they are highly complex machine learning models. Many banks have at least some of these core models built, with a process in place for monitoring and maintaining them, but outside the handful of core models, the need for embedding artificial intelligence (Al) and machine learning at a deeper level is an unfulfilled need at most banks.
At least part of this deficiency in AI and machine learning can be explained by the capacity of data scientists and availability of data. Traditionally, data science teams have had their hand’s full building and maintaining the core models of the bank. Add to that the challenges involved with managing, cataloging, and assembling data, along with the obstacles associated with implementing these models into production, and it’s no wonder that progress has been so slow.
To complicate matters, the number of players in the banking sector has increased dramatically in the last several years. Fintech companies are
hard at work gobbling up bits of market share, and they’re doing it by means
of innovation and Al. Banks still have the advantage in terms of data and expertise, but that advantage won’t last if bankers don’t start innovating.
When it comes to AI, some of the lowest hanging fruit in banks today are the so-called relationship businesses. These are typically high-touch, high-profit, complex relationships that have dedicated relationship managers and generate a mountain of revenue for banks. Business banking and corporate banking businesses along with wealth management teams are good examples of this type of relationship business.
Several years ago, I was at a sale offsite for a Fortune 100 company. The attendees were relationship managers that were all compensated on sales. The first half of the day was set aside for discussing prospecting strategies. Salesperson after salesperson stood
up and told about the newest event that they had hosted that was really
“working” for them. Box seats at sporting events, breakfast meetings, cold calling, and many other in-person events were touted as “the way” that was going to produce results.
At the time, I remember being struck by how unscientific it was. I even made a comment about collecting data to see what was actually working. I was quickly reprimanded: “This is a relationship business,” one experienced salesman told me.
I’ve heard that same sentiment many times since then. “This loan is too complex for a pricing model.” “This relationship is too unique to be lumped into a dataset and learned from.” “This account is much too important to be driven by your models ”
Fast forward a few years, and these relationship-driven businesses are beginning to change their tune. Businesses, who are impacted by ultra-low interest rates and increased competition, are
now open to exploring new strategies to grow and deepen their relationships. The opportunity to do so is huge.
The use cases
All sales teams use CRM tools to track leads, opportunities, and accounts. That means that all sales teams have incredibly valuable data that tracks where they’ve been successful and where they haven’t. This data is the foundation of building an Al solution to grow a business. Prospecting for the best new customers
In the world of sales, prospect lists abound, driven by inbound leads, cold calls, and third-party data sources. Sorting through it all to know who to contact, and who to focus on, can be bewildering — particularly if you don’t have access to any reporting or analysis showing what has been effective (and what hasn’t).
Building an Al solution for prospecting involves first identifying what a “good” prospect looks like. The most obvious definition of a good prospect is anyone that is likely to buy what you’re selling, but that’s far from the only example.
Bankers might want to try to predict the credit quality of prospects based on publicly available data to identify the clients where they want to spend their energy. Or perhaps they might want to predict which clients are likely to produce the most value for their business, either by growing in the future or by their needs for a wide variety of products.
In any case, once a good customer is defined, then it’s just a matter of matching up external prospect data with a bank’s existing client base to
build a training dataset. For example, I might take the data I have available for my prospects (my features) and match them up to my current clients where I know the target value; e.g., risk rating or profitability (my prediction target).
I might try to predict profitability, product appetite, or margin. Once the models are built, I can then score my list of prospects in order to rank them for potential sales efforts.
Deepening relationships with your current customers
The number of products and services that a customer needs is strongly correlated with how profitable that customer is. Al can improve the effectiveness of relationship managers by identifying particular customers that are in need of particular products. If a relationship manager can reach out with a specific product to the right customer (and not reach out with the wrong products), then that customer will be happier, sales will be higher, and the relationship will grow stronger.
Fortunately, if the bank utilises a CRM system, the data needed for this project has probably already been captured. It’s a simple matter of identifying which customers were offered which products and whether or not they bought them. Which customer and product attributes will be predictive of purchasing likelihood will depend on the product.
Once I have my models in place, it’s relatively straightforward to identify the best prospects for a marketing campaign or to have an offer in the “back-pocket” for every customer interaction.
Pitfalls to avoid
These are just a couple of examples of ways that AI can increase the effectiveness of sales teams. Building the solutions is not complex from a technical perspective and deploying them is also pretty straightforward. That said there are a few blockers that you will want to avoid as you build these solutions.
Don’t fail to record information about lost sales. Salespeople hate data entry (doesn’t everyone?). They especially hate entering data about deals that they’ve lost. Capturing data about lost deals, though, is just as important as capturing data about won deals. Machine learning requires data that the machine can learn from. That means positive and negative examples. In order to build models to predict customer behavior, all the possible outcomes have to be captured.
Don’t expect it to work perfectly the first time. These solutions, even though straightforward, take some iteration. I worked with a sales team at a large bank
to predict which clients were most likely to need a foreign exchange (FX) product to lock in conversion rates for cross-border payments. It took us six iterations to get to a good model, but when we finally did, we increased the conversion rate from 2% to 10%, which meant a SIM+ increase in sales for the organization.
Prospecting and deepening models are two ways most banks can start exploring how Al can impact your organization. These use cases are low risk, quick to build, easy to implement, and have a high ROI, and advanced tools like automated machine learning now make developing these solutions accessible without the huge upfront cost required in the past; e.g., massive time investment, hiring and retaining large data science teams, tricky
manual deployments, etc. Whether you’re a large bank or a credit union, with sizeable complex deals or simple term loans, Al can provide a straightforward way to be more targeted in your sales efforts.
Don’t wait until your data is perfect. Everyone has data issues, but it’s pretty rare to come across a data set-up that is so bad that nothing can be done. Starting
with the data that’s currently available is the only way to get started with Al. Waiting until everything is perfect means never getting started at all.