Continuous learning
Those adaptive models help the system learn from customer responses, elevating actions that resonate with customers while gradually deprioritising those that fail to generate engagement. It’s a process that allows the bank to continuously refine how it communicates with customers while maintaining consistency across channels and business units.
“You can ask questions, you can understand the deeper context, and that’s been one of the key foundations of decisioning: how you feed that information in. I think it gives us a richer pool of information to understand our customers and helps us tease out more intent,” Cuthbertson said.
When it comes to the next-best conversation, those conversations are pulling from that foundation of decisioning. “It helps our agents engage with customers in a different way, but ultimately, I can still trust the decisions that are coming out of it,” she continued.
“For us, all the layers have to work together. If one of them is weak, there’s going to be a weakness throughout the whole foundation. It never changes how we act, but decisioning determines whether we’re acting correctly. That’s a really important distinction in the three layers from a decisioning perspective.”
But the success of the platform created a new problem for NAB. As the number of actions grew, the bank began asking whether it was identifying enough customer moments to justify them.
The question led to a review of the home lending portfolio. NAB had already invested heavily in helping customers choose the right mortgage product and become established after settlement. It had also built programs designed to identify customers who might be preparing to leave. But the bulk of the customer journey in between those moments was being ignored.
“We were doing a really good job of engaging customers. We’d built a lot of actions, but we were focused on the moments of truth when a customer comes onto our system. How do we help them choose the right product, and how do we get them set up with all their online banking and digital needs? We also did a bunch of stuff at the back end, when we thought they might leave us. Then we realised we weren’t really speaking to them in the middle of that journey,” Cutherbertson said.
“By being able to step back and look at it from a top-down perspective, we found that we were only engaging about 20% of our customer base. We used a blueprint to take that top-down view and came up with a series of actions to fill the gaps.”
The result was a dramatic increase in customer coverage.
“We’re now talking to 75% of our customer base. Simple things like little milestones and nudges when they’re paying off or paying down their mortgage help keep them engaged throughout the lifecycle,” she said.
Some of the new interactions are relatively simple. “It just keeps them engaged throughout the lifecycle.”
Smarter team decisioning
Of course, mapping journeys and identifying opportunities can consume significant time and resources, but the bank is finding better news on that front also, as tools mature.
“Where I see agents playing a role is helping us spot those opportunities more proactively, rather than somebody having to go in manually or having to do a workshop with 20 people to map out that journey,” Cuthbertson said. “On the front end of planning, it’s usually around the number of people involved in coming up with ideas, the quality of those ideas, and the time to market. If you think about a normal marketing ecosystem, when you’re planning what you want to design the system to look like, that could be at least a week of workshops with lots of different opinions. Now, you can just pick a Blueprint (part of the Pega toolkit). We did it a couple of weeks ago in a workshop, where we pulled it up and had all of that mapped out in about a minute.
“What you’re then doing is having a debate about what’s the priority and where we think we want to play, rather than everybody trying to come up with a bunch of ideas.”
Cuthbertson sees a future where speed-to-design accelerates, and changes how many people are required to ideate.
“Where we haven’t yet done the work, but what’s coming in the Pega roadmap, is around time to market. It usually takes us about three days end-to-end – not full capacity, but three days – for building, testing, and deploying actions. I want to get that down into hours.”
She also believed AI can help automate much of the operational work involved in maintaining large decisioning environments.
“I’m looking forward to a world where agents are building the decisioning rules. For me, they’re still deterministic rules, but we can automate that. We can take time away from humans having to build and code those rules by using agents to do it. That means less time managing the system and more time, as I said, on the front end thinking about what we’re doing, what we’re building, and improving the quality of the decisions. It’s really about automating those routine tasks and elevating the thinking around what we’re putting into the system.”
According to Cuthbertson: “I think being more customer-obsessed is about being able to take that step back and look at both sides of it: making it more efficient so people can focus on the right things to put into the system. So that’s what I’m thinking about now. I do think having agents backed by good decisions is going to be a key differentiator for us.”
NAB is also expanding the types of information available to its customer brain. The first generation of the platform relied primarily on structured customer data such as transactions, accounts, balances and product holdings. More recently, as the data lake infrastructure has evolved, advances in AI have made it possible to extract value from less structured sources.
“We started off with our core customer structured data: Who are your customers, what are their transactions, what accounts they hold, what services that they use, that’s your core foundation. Now, what we’re starting to really be able to discover at scale is your more unstructured data. We are looking at call centre transcripts, web chat, and some of the new tools and techniques make it faster and a lot simpler. Actually, the quality of what you get out is much better,” said Cutherbertson. “We look at things like [when] customers have called in lodging something like a transaction dispute. We can kind of get that context right away and feed that back into the engine.”
Actions are evaluated based on customer engagement and the outcomes they generate, she says. Those that fail to deliver value can be modified or removed.
“If it’s not working, why bother running it? Switch it off.”