Want to delight customers? Apply design thinking to your AI products

by Michael Radwin, Vice President, Data Science, Intuit

Poor Microsoft Clippy. He meant well — and he embodied artificial intelligence that was quite sophisticated for the late 1990s — but he was destined to become a punch line.

Why? Because even the most sophisticated technology can’t guarantee a great customer experience. That’s why the best AI products are a team effort by data scientists and designers. Great data scientists work at the intersection of math and statistics, and software engineering. Great designers bring a unique combination of creativity, aesthetic sense and deep customer understanding.

Together, they provide a perfect blend to achieve the best user experience.

Of course, bringing these disciplines together isn’t as simple as putting them in the same room and hoping for the best. Data scientists and designers speak different languages and hold different mindsets.

To bridge the gap and drive innovation here at Intuit, we’re staunch advocates and practitioners of a Design for Delight methodology. D4D for short. It’s a simplified approach to what the Stanford University d.school calls Design Thinking. This method for creative problem-solving has been broadly adopted by world-class design firms like IDEO and many of the world’s leading brands like Apple, Google, Samsung, and GE. And, it’s been instrumental to Intuit’s success in developing AI/ML-driven software and services that help consumers, self-employed individuals, and small business owners make better financial decisions.

D4D is based on three core principles:

• Deep customer empathy Define the problem based on a deep understanding of our customers and their pain points, not our own technology or company.
• Go broad to go narrow Think broadly about solutions and generate a lot of ideas before choosing the best one.
• Rapid experiments with customers Prototype, test, and iterate to keep projects rooted in the real world and the results we’re actually delivering.

Following are three real-world examples of how we’ve applied D4D here at Intuit to help our financial services customers make smarter money decisions.

Helping Mint customers avoid overdraft fees

A couple of years ago, we discovered that Mint customers had paid $250M in overdraft fees, just to the 5 largest U.S. banks. Even worse, half of these fees were triggered by transactions under $50 (according to the Consumer Financial Protection Bureau, a majority of fees incurred are on transactions of $24 or less.) When a $3 cup of coffee ends up costing closer to $40, it’s a real problem for anyone living paycheck-to-paycheck.

So, we decided to do something about it.

By applying a blend of data science and design, we set out to create a new service that would help Mint users to avoid exorbitant non-sufficient funds fees.

First, with our customers’ consent, we analyzed real-world financial transactions by Intuit customers to learn as much as possible about the circumstances most likely to lead to an overdraft. Next, we used these insights to build a machine learning model that could accurately predict overdrafts. The gradient-boosted tree model was based on the 20 most important parameters, ranging from the person’s average income and account balances to recent overdrafts and debit charges for less than $50. Then, we optimized it for high precision to trigger alerts up to a week in advance, giving users enough warning to take preventive action: a transfer, deposit, or not buying that cup of coffee.

While the model itself worked flawlessly, encouraging users to take action was something else entirely. Through rapid experimentation, we learned that customers were most responsive to simple, short text messages with tailored, interactive visual calls to action. Just as importantly, we validated that loss aversion is a strong motivator for action, given that people typically “prefer avoiding losses 2X as much as acquiring gains.” Altogether, this informed the approach illustrated in the “after” screen image below.

Ultimately, we applied a combination of data science, behavioral economics theory developed at the Hebrew University, and design thinking to our development of the new Mint Overdraft Prediction Service, launched in March 2018.

With a full year under our belts, the results are encouraging. As of March 2019, Mint sent out more than 650K messages warning customers about a possible overdraft. Best of all, users acted on many of those alerts, saving nearly $1M in fees, and we’re just getting started.

Making it easier for QuickBooks Self-Employed customers to deduct business travel

The IRS lets self-employed U.S. taxpayers deduct business-related trips at $.58/mile, as of January 1, 2019. That can really add up, but first, they’ve got to do the legwork: a tedious process of manually categorizing, reviewing and annotating each trip.

Starting in 2016, we took steps to dramatically simplify this for our QuickBooks Self-Employed mobile app users, giving them the option to track and categorize trip miles by simply turning on auto-tracking in the mileage tab, and swiping left or right to mark trips as business or personal.

In 2018, our data science team set out to find an even better way to help our customers identify tax-deductible business trips. We used a frequent pattern mining technique to group likely business trips. For example, for a particular user we could identify from historical patterns that all weekday trips, from location A to location B at a certain time of day, should be marked as business. Based on this data, we were able to predict that new weekday trips, from A to B at a certain time of day, would most likely be business-related. By extending this across all historical trips for a particular customer, we could begin to predict how new trips should be categorized.

Design thinking helped us translate this business-versus-personal categorization into an awesome customer experience. We performed a series of customer interviews to better understand their thought process when reviewing trips, and then designed and tested prototypes to learn how they perceived the idea of a bulk review of business trips. This helped us tweak our designs and models to make them as clear and simple as possible.

Now, our customers can choose to use the QuickBooks Self-Employed mobile app to automatically track miles and categorize them as business or personal for bulk review and annotation, further simplifying the time-consuming administrative task of adhering to IRS regulations.

Saving time while maximizing refunds for TurboTax customers

For many Americans, the tax refund is the biggest paycheck of the year. Most tax filers will get more back from the IRS by using the standard deduction. But they want to make sure they get every penny they deserve, so they often put in endless hours itemizing deductions anyway. To help them make a confident decision without such painstaking effort, TurboTax asks users a few simple questions to determine whether standard or itemized is best for their situation.

The data science behind the Intuit ExplainWhy feature was rock-solid. We built a machine learning model (with access to over 600 features) that used data from the first part of the tax preparation process to make a recommendation with incredible accuracy. Surprisingly, this level of sophistication and immediacy backfired when we tested it with customers. They simply didn’t believe we had enough information about them to make an accurate prediction. In short, they wanted us to “show our work.”

Through qualitative studies and rapid prototyping with real tax filers, we learned something that data science alone never would have predicted, but was discoverable through our D4D process. Ultimately, we added a few up-front questions that consumers could easily answer, giving them a greater sense of control over the process and more confidence in our results. We also made it possible for users to bypass the model’s recommendation, so they could go through the full itemized-deductions process anyway and compare this outcome with the standard deduction. While only a small number of users take that step, we know this level of flexibility is important to them.

Taxpayers needed us to “show our work”

How did we do? The numbers tell the story. Of 15 million eligible TurboTax customers, we estimate that nearly 90 percent of them will now take the higher standard deduction, saving the time it takes to file their taxes.

Intuit has spent many years applying lessons learned from blending data science and design thinking to deliver awesome experiences to customers of our portfolio of AI/ML-driven software and services: QuickBooks, TurboTax, and Mint. To learn more about this methodology, and how you might apply it to your products, this Stanford d.school reading list is a great place to get started.

This has been an Intuit contributed blog for The Hive Think Tank, Medium.