How We Created the Heartbeat of TrueAccord

By on May 29th, 2018 in Industry Insights
TrueAccord Blog

The genesis of Heartbeat—the machine learning engine that makes TrueAccord debt collection a reality—is a story that demonstrates our commitment to consumers who hope to take control of their financial future. Heartbeat is how we create a more humane and thoughtful collection experience.

So What Is Heartbeat?

Heartbeat is a fully automated and reactive decision engine that uses a combination of machine learning and data-driven heuristics to determine the optimal way of interacting with each individual debtor. It tells us when we should contact them, how often, through which channel, with what content, and what specific types of offers we should provide. Most importantly, it is the engine that replaces an agent’s phone-based collection activities with a data-driven strategy, ultimately making the same decisions, but automatically, more quickly and with a bigger heart.  

How Was It Built?

We built Heartbeat with three key tenets in mind. The first is compliance: to create a pre-approved boundary for what should and shouldn’t be said to each debtor. The second is performance: how we leverage data-driven heuristics to test our assumptions and continually improve the performance of our debt collection system. And the third is the customer experience: how consumers engage with the product at every phase and how we ensure we’re seeing positive reactions. Put these all together and they constitute the core foundation of Heartbeat.

We Start with Data, Then We Test

The process starts with data, and lots of it. We’ve collected years of historical collections data to help determine the optimal way to communicate with consumers and generate the best collections model. We then set up an experimentation engine to test and refine the process for continuous improvement.

For the testing to be relevant, we ask a few key questions: Is this a problem that we can define well enough to solve and take action? Do we have enough data from different segments of our population to solve the problem for all of our customers, not just some? And does the result add value to the process? Otherwise, it’s not worth putting the time into it.

Once we decide to move forward, we establish a hypothesis (e.g. paydays are the best days to set up recurring payments) based on the intuition of our domain experts who know what an ideal customer journey should be like. We test our hypothesis with an A/B experiment to see if it performs better than our current status quo. The data we collect from these experiments shows us what tactics work best. To ensure we’re optimizing for various audiences,  we re-target the test to new segments until we have enough data to apply the new treatment to the broad audience.

Then Machine Learning Takes Over

The biggest challenge is that data and heuristics are not enough to offer highly personalized treatments at scale. At some point we have to transition, by taking the learning outcomes based on all of our initial data and programming them into a machine learning model. The goal here is to replace human heuristics with an automated decision-making model that continues to learn from multiple samples at scale. A human agent is prone to biases, such as using non-compliant language in their calls when pressed to make their monthly numbers or using the wrong tone based on a previous conversation that may have impacted their mood. A machine learning model doesn’t fall prey to these biases.

The more data we collect, the better the system gets and the more accurately it represents edge cases and special needs. Now with more than 2.5 million consumers and tens of million of interactions, we’re seeing great results and constant improvement. The larger sample sizes also allow us to reach a statistically significant result faster in large experiments, often in only 30 to 60 days.

What’s Next for Heartbeat?

Right now, only two percent of our customers still need to interact with one of our agents. That’s already a pretty impressive number, but we still want to reduce it even further.

We constantly scale the technology behind Heartbeat and improve its intelligent self-service capabilities that feature our three key tenets: better compliance (Heartbeat can navigate the legal restrictions with less risk than a person), better collections performance (50-500% better than our competition), and a better customer experience where consumers are empowered to manage their debt in a way that puts control back in their hands and treats them the way they want to be treated.  

At TrueAccord, we’ve always been committed to providing the best customer experience for a behaviorally complex debt collection process, and Heartbeat is true to its name in working to that objective.

If you’re interested in learning more, check out this interview with one of Heartbeat’s creator’s Sophie Benbenek!