What AI-Enabled UX Design Frameworks Tell Us
When Congress authorized $ 284 billion to reopen the Paycheque Protection Program (PPP) in January, small businesses and the self-employed scrambled to apply.
But many applicants ran into questions about eligibility, loan cancellation, and terms of service that were not easy to answer in public forums like the US Small Business Administration (SBA) website.
To further complicate the application process, a separate SBA relief program – the Economic disaster loan (EIDL) – was announced in April as part of a $ 484 billion COVID-19 relief plan. Together, the programs comprise a dense jungle of atomized information that is difficult to sift through – at least without sound advice and an extraordinary amount of patience.
Intuit, a company that built its brand on distilling financial data and tax laws, recognized the opportunity to bridge the information gap. In April, he launched Help Help, a compliance-driven financial tool to help small businesses and the self-employed navigate the requirements contained in hundreds of pages of the Coronavirus Aid, Relief and Economic Security Act (CARES).
Designed and built in a matter of weeks, the program offers insight into how well-designed information architecture, paired with automation, can convert long, obtuse documents into a digestible format. The tool helps users determine their eligibility, estimate loan amounts and obtain funds. In a sense, it’s a design framework that is being built.
“Two years ago, before [information architecture] became a thing, everyone just said, “Oh, this is Twitter Bootstrap.” People confuse it with design systems. It really is much wider than that.
Clarence Huang, software engineer at Intuit, said that Aid Assist’s notable achievement is not so much the interface components themselves, or the asset libraries that describe them, but the way the page looks. displays with each other. The automated system predicts when users are confident, confused, or tired and makes screens calibrated to reflect those mental states.
To borrow a phrase from Radiohead’s Thom Yorke: “Everything in its right place. “
“Two years ago, before [information architecture] became a thing, everyone just said, ‘Oh, this is Twitter Bootstrap’, ”Huang told me, referring to a popular front end user interface toolkit. “People confuse it with design systems. It really is much wider than that. This is a cross-cutting concern in software development that affects product, engineering and UX design. And I think the more we, as an industry, know, the better it makes us. “
A generative design framework
To develop the framework for a complex construction with 400 unique screens, the design team started with three main goals:
Basic Design Goals of Aid Assist
- Reduce cognitive load. Users needed to understand the requirements of the CARES Act and how it could benefit them.
- Build a mental model shared with the user. This model would determine the navigation architecture of the application, as well as the decisions that guide users on particular paths.
- Reduce action bias. For example, a user may receive advice on whether a PPP loan or an EIDL is preferable based on their financial profile. However, the best option may not be to pursue either of these loans.
Huang said these goals are built into a generative design process that determines the screen sequence and interactions for each user. Think: Machine learning responsible for creating custom flowcharts. These pathways vary depending on factors such as monthly payroll, number of employees in a business, and loan disbursement plans.
“We designers haven’t really designed the product page by page, but we have designed a framework. And that framework basically has all of these two components – PPP and EIDL – integrated. So the computer uses this frame as a kind of factory to eliminate screens, ”Huang explained.
While the expression “information architecture” may seem higher than what he typically describes – a site map – in which case he aptly describes a clever design framework that Huang calls “a skeleton for the house.”
“The algorithms, like the factory, are kind of put on the walls, the roof and the stucco,” he added. “Then he smashes a lot, a lot of houses after that, based on that same frame. And this framework has some built-in things, which we learn by testing users. “
Here’s what the design team found:
Lessons Learned from User Testing of Aid Assist
- People are more likely to complete tasks if you present one question, and only one question, per screen.
- When giving people useful information, it’s best to gradually disclose that information instead of giving it all to them at once.
- After five or six questions people start to get decision fatigue and you need to give them an affirmation. Aid Assist does this through automatically inserted interstitial screens that let users know they’re on the right track.
- When users switch between screens, it’s a signal that they’re confused, and it’s time to offer some more advice.
Progress bars are gold, even if they’re just estimates
Arguably the most illuminating finding from the design team’s user testing was that a progress bar was critical to user engagement and task completion rates. However, in an adaptive system where almost all applicants follow a slightly different user flow, it has proven difficult to estimate a user’s progress.
Some small businesses have less than 10 employees. Some have close to 500. Nonprofits, veterans organizations, tribal businesses, and food and hotel businesses have separate eligibility criteria that map different flowcharts.
“It’s actually a very tricky question, because you can think of the flow in something like Aid Assist as a decision tree. And you don’t know in advance how long a user is going to take to browse this tree. But this is not good, because we have to give users a sense of progress, ”Huang said.
“Users interpret [the progress bar] move as a sign of our work for them. It’s so deep and strange.
The solution the team came up with – a predictive algorithm that approximates a candidate’s progress – is not entirely linear, nor strictly accurate on a percentage basis. But that doesn’t really matter, Huang said. Most people just want to make sure that someone, or some machine, is mining the data behind the scenes.
“With one of the users we tested, I asked, ‘Why do you even like [the progress bar]? ‘ Huang said. “They say to me, ‘Oh, you work hard for me.’ Thus, users interpret [the bar] move as a sign of our work for them. It’s so deep and strange.
Interstitial screens increase task completion rates
Intuit’s discoveries add to more and more evidence that it is not the minimization of clicks that makes navigation efficient, corn if the user has the right breadcrumb trail to complete a task.
“Thanks to some of these longer flows, [users’] the cognitive load is building up. And when we tested that, adding interstitial screens gave them an immediate result that lowered their cognitive load, ”Huang said.
At present, the once revered three-click rule has been largely demystified, and complex documents like the CARES law illustrate its inability to adapt to complex use cases – such as tax law or drug management. But what‘is particularly striking about Intuit‘User research shows how it disproves popular beliefs about limiting the time it takes for users to reach their end goals.
“A lot of these compliance-driven experiences can be visualized as a graph or tree. But you can’t show it to the user, because it’s just incomprehensible.
In at least one test stream, the addition of interstitial screens resulted in a “double-digit” increase in end-to-end conversion, according to Huang. What emerged was a roller-coaster pattern: user engagement waned over multiple question screens until a timely congratulatory message gave them a boost of confidence and rekindled their motivation.
“Many of these compliance-focused experiences can be visualized as a graph or tree,” Huang said. “But you can’t show it to the user, because it’s just incomprehensible. “
Instead, he told me, you “flatten” the data and present it in a linear fashion. The algorithm works as a support coach, giving users a boost when they need it. This does not result in the shortest distance between two points, but it is the route that users are most likely to take.