Each day the Commonwealth Bank of Australia analyses 157 billion data points through its automated decisioning technology. The AI powered “brain” allows the bank to alert customers to potential fees and benefits they may be entitled to, among other services and sales use cases.
CBA has said the missed fee revenue will likely cost it $400 million. The strategy and a multibillion dollar investment in digital technology is part of an effort to rebuild customer trust after years of scandals, according to new CEO Matt Comyn, who has his sights on establishing CBA as Australia’s digital leader.
Alex Burton, head of data and decisioning strategy at the bank’s retail arm is responsible for much of the underlying technology. He says the high bar set by Comyn is in line with customer’s expectations, who now judge the bank not against direct competitors but digital giants like Amazon and Netflix. Burton says, “That’s not an unreasonable expectation” from customers and it is how the bank now measures itself.
“We look around the local market and I don’t think we have a lot of competition [in digital experience]. So we have to measure ourselves against that experience [provided by global leaders]. And our customers are absolutely measuring experience against those other [leaders].”
But in delivering that experience Australia’s largest bank faces challenges around scale, and compliance hurdles tech companies have so far largely avoided. And in the wake of a Royal Commission CBA’s margin of error has never been thinner.
The technology the bank uses to automate decisions, provide customer alerts, support and sales tools is provided by American software vendor Pegasystems. Using Pegasystems, software which uses AI to link and automates processes, integrating them with customer engagement tools, CBA hand coded several AI models to create what it calls “Next Best Conversation” – a tool which predicts outcomes and suggests potential actions. Those NBC models were then used to train new AI models, exponentially speeding up the process, Burton says, and reducing years of data science work into weeks.
Today, over 200 NBC models feed on the 22 million digital interactions CBA has with customers every day, analysing 157 billion data points to provide automated and potential actions.
“Every time you open [your app] it’s decisioning, you swipe its decision again,” Burton told Which-50 during the Pegasystems user conference in Las Vegas this month.
“So you’re always going back and doing the most real time off all the information we get … Over the next 12 months we’ll do 3 billion decisions in mobile.”
The automated decisioning helps CBA send out half a million payment reminders each week to help customers avoid fees. The bank has also begun sending reminders for benefit entitlements from third parties.
For example, CBA data can show when a customer may be entitled to a refund from the RTA. The models automate the process, sending a notice to customers. Burton said the collective value of those RTA rebates to CBA customers will be $150 million for the year.
At the moment these types of alerts rely on an analysis of a mix of batch and real-time data analysis. But Burton says the real time integration is continuously being increased.
Of course, CBA uses the tools in its sales and services as well, identifying high value customers and their propensity to respond to offers. Pegasystems rolled out new “customer empathy tools” this year to all of its customers for those marketing and sales scenarios. When the company announced the new features, which puts organisations “in the shoes of the customer” it highlighted CBA as an organisation which had developed similar tools of its own accord using Pegasystems software years ago.
Burton told Which-50 just because a person is eligible for a financial product, and likely to buy it, does not mean it should be offered.
He said the bank had developed “guard rails” to ensure such decisions were not automated or left to AI. The bank uses three elements to asses these type of scenarios. Firstly, is an absolute eligibility – is there anything stopping a customer from accepting an offer like age requirements for example? Software does the heavy lifting in this step, analysing millions of potential customers, removing those not suitable.
Second is the “should test” – should a customer be offered a product or service. At this point the bank deliberately avoids any AI use, according to Burton.
“We still hard code those rules in,” Burton says, “To make sure they’re not AI driven, there’s not greyness in them as such. We sit down as a business and work out what’s our appetite for those sorts of things? And then we hard code it into the system.”
Finally, there is the “propensity” element – “How likely is the customer to want that thing? And that’s where we, we pull out all the stops,” Burton said.
At this point AI becomes very useful, Burton explained, predicting a customers likelihood to respond to an offer and the next best course of action for the bank. But he insists it can’t reach that point without the proper non AI checks and balances.
The current system took several years to build and mature, and Burton says finding the talent to construct it initially was hard because of a lack of local talent.
“When we first started this a couple of years ago, we scoured the world for [talent] … If you went back five years trying to find decisioning talent in general was a hard thing to do. So we hired people from Turkey, we hired from Spain, we brought [Chief Analytics Officer] Andrew [McMullan] across from RBS to help run at once it was in place.”
That skills gap for decisioning experts locally still hasn’t closed much over the years, Burton says, but the basic understanding of some of the technology is improving.
“I think more people get it – the outcome. Part of it was selling the outcome, not just the technical side of it. I think more people are getting the unified decisioning connected to all channels is the thing to do with data and the intelligence in it. And I think more businesses subscribe to it so you don’t have to do the selling internally.”
Finding “skills on the ground”, however, remains a challenge, Burton said, noting competitors were experiencing similar challenges.
“There are many people who have tried to emulate the same things we’ve done in the region, and also around the world, and I don’t think they’ve quite got there yet.”
A previous version of this article stated CBA analyses 1.8 billion data points in decisioning analysis. The correct amount is 157 billion data points.