Sapphire Now 2018 - How the Queensland Treasury is using SAP Leonardo to predict tax delinquency

Profile picture for user jreed By Jon Reed June 11, 2018
Summary:
You might not think of a tax collection office as a hotbed of digital transformation - neither did I. But the story of how Queensland Treasury is using SAP predictive analytics to spark their client-focused culture turned into a Sapphire Now 2018 standout.

predictive-ball
At SAP Sapphire Now 2018, our diginomica team coverage surfaced use cases where organizations are pushing forward with new technology. But what we really want to hear is how the organization is changing from the inside out.

Finding out how the Queensland Treasury in Queensland, Australia is using predictive tech from SAP Leonardo was interesting enough. But then Simon McKee, Deputy Commissioner of the Queensland Treasury's Office of State Revenue (OSR), said this:

The cultural change was a precursor, if you will, to the digital.

Hold up. Culture change in a tax collection office? Yep. The OSR, which is a business unit within the Queensland Treasury Office, began with SAP ERP in 2005. They added a tax revenue management system in 2009, and they moved their system onto HANA in 2015.

A digital transformation, not a headcount reduction

But McKee knew that the new technology wasn't going to stick without a plan for digital - and cultural - transformation. In 2017, The initiated a transformation program. As McKee told me and Vinnie Mirchandani:

We invested heavily in culture change to get our people match fit for the digital change coming down the pipe. We invested heavily in leadership.

One early goal? Deter employee fears this was a headcount reduction efforts disguised as "transformation."

We felt that unless our people were ready to adopt new ways of doing things, [we wouldn't be successful]. We communicated that this is not about job replacement - far from it. It was about job enrichment. That was the key message we put across.

But why the shift? What was the motivation behind it?

Well the world is changing; our ecosystem is changing. The expectations from our clients is that they are direct with us 24/7, on the device of their choice. Our staff want more interesting work to do, they don't want to move paper around a desk.

The Queensland government backed the effort:

Our government is being very supportive of this, because they want Queenslanders to be able to get on to do their business, to grow the economy. The whole purpose is to make it easier for our clients.

The digital transformation effort began with "design thinking and disruptive training" workshops, with eight week sprints for each revenue team, or line. They learned some eye-opening things about the taxpayers they serve. Lesson one? Don't call them "customers." McKee:

They don't see themselves as customers. They say, "We are tax payers; we have to interact with you, but we don't necessarily want to. What we want is a really good experience. We want to get in and out quickly."

But there's a huge caveat, also a big theme of Sapphire Now. Better customer experience? Yes. But not as the expense of compromising the customers' data:

As you heard Bill say in the opening keynote, "We don't want you to be creepy." By that I mean, "If you are going to use technology to identify us, such as voice recognition or facial recognition, we need to understand what you are going to do with that data." The commissions' number one priority is around protecting people's [data].

The OSR has now run thirty-six digital initiatives. They run these applications out of an SAP-managed private cloud in Queensland (the HANA Enterprise Cloud, or HEC). The range of projects include a virtual assistant and an SAP contact center solution. A voice-to-text project is planned; they want to allow operators to focus on their clients rather than type. The call data will be pulled into a new data warehouse.

Next up - machine learning and SAP predictive analytics

But it's the Machine Learning (ML) projects that really have McKee excited. The OSR is the first government agency in Australia to deploy a machine learning project. First up: a land tax debt project to go live in July 2018, with more ML go-lives planned by end of year. The core goal of land tax debt predictions? Anticipate when somebody might become a debtor before it happens.

In 2017, McKee's team ran a proof of concept with an SAP team from Palo Alto. In eight weeks, they crunched 187 million records from 97,000 tax payers. Three internal sources of data were used. Despite what McKee described as " limited data - not necessarily clean data," the machine predicted, with 71 percent accuracy, when someone would default before they defaulted. McKee believes with more data - and cleaner data - they'll be able to get that number much closer to 100 percent.

SAP did the heavy-lifting of building the functionality:

They worked with our subject matter experts to understand the job to be done. Then they went away and developed the solution, brought it back and said, "Is this what you wanted?" We said, "Yeah, but a bit of adjustment here and there." They went away again, came back, and it's been an easy process.

To build the tool, SAP used the HANA predictive analytics library (PAL) and a customer retention application:

The brain is in the machine, but it is custom built for us.

So how do they feel about the 71 percent accuracy rate? Is that good enough to create opportunities?

We're over the moon, because that was using three internal sources of data, and not all clean. So now we are adding more data.

McKee acknowledges they need to get that number higher:

As we add more data we expect it's got to be greater than 71 percent. Also, one of the projects we've done since then is cleaning our data. We've de-duplicated it, cleaned it, and we've got a [single customer record].

McKee showed us a screen from the land tax debt tool that indicated the tax year and the risk ratings (low, medium, and high risk). That gives their phone operators the ability to see the client's risk assessment in real-time.  Aside from improved customer service, what else can you do with that information? As it turns out, there are two main groups of non-tax-payers:

  1. One group has no intention of paying for whatever reason. Predictive can help with that group by sending out pro-active notifications, or strongly worded letters to repeat offenders, hopefully for a better result before long-term delinquency occurs.
  2. Another group wants to pay, but is unable to for a many reasons. Their is plenty of predictive work with this group, including reaching out to them for better payment terms. In a future case they are working on now, taxpayers could be granted machine-generated payment relief if they are caught in a situation beyond their control, such as a natural disaster.

The ability to segment into these two groups forms the basis of future campaigns, targeted to each. But this isn't about criminalizing people: "It's about helping those people who can't pay rather than penalizing them."

McKee wants to avoid the ethical quagmire of identifying machine versus human:

We are going to be transparent. They won't be dealing with a machine, they will be dealing with a human, but the machine will generate automatic messages, like "We'll come back to you when things get better." The machine can monitor to see whether their activity has started up again and it might go, "Hi, are you able to pay this now? If not, here's a payment arrangement."

McKee says it's not too early to see some results from SAP Leonardo: "We're seeing it now, absolutely." That includes revenue bumps on tax collections. SAP solved the problem of in-house data science talent:

Within 12 months of us actually talking to SAP, it's going to be deployed... We can't compete for Python developers, and we can't compete with the Google's for data scientists. So this offering is as a service.

End note: Vinnie Mirchandani of Deal Architect joined me in this customer discussion and helped to bring these themes to light. You can catch his Sapphire Now coverage on his Deal Architect blog.