John Zimmerman, Changhoon Oh, Nur Yildirim, Alex Kass, Teresa Tung, Jodi Forlizzi
AI and UX design have grown up as quite different disciplines. But we're now starting to see that small bits of AI can enrich a UI in interesting, useful ways. Adaptive user interfaces (AUIs) employ elements of AI to improve user experience. AUIs recognize and automate frequent tasks, such as when an email recognizes a phone number and lets users initiate a call with a tap on the number. These bits of low-risk AI free up a little time for consumers and maybe make them a little happier.
Similarly, workers spend millions of hours every day interacting with business process applications, often performing certain repetitive tasks exactly the same way. Procurement officers order the same supplies from the same vendors; nurses retrieve the same clinical information describing patients; lawyers search for and open contracts they are writing based on incoming emails; and instructors check to see which students did and did not turn in their homework, due the night before. AUIs can learn from prior actions and automate the most frequent, repetitive tasks, freeing workers' time and attention to focus on decisions that require real human intelligence. Interestingly, while AUIs are becoming more common in consumer-focused applications, adoption lags for enterprise applications, where the return on investment in worker time-savings and increased happiness is more easily measured and valued.
We are a group of researchers and designers from Accenture and Carnegie Mellon University exploring how AUIs can streamline enterprise business processes. We propose that UX designers are the right innovators to incorporate AUIs into enterprise apps. UX designers are uniquely positioned to recognize situations where an AUI might help because they view the world through the eyes of the user. They can anticipate situations that will occur frequently, impact a large number of users, and require the worker to execute the same sequential set of actions to complete a task. However, there is currently no systematic approach that UX designers can use to identify and assess AUI opportunities. UX designers receive no training on AUIs. In addition, current design processes and tools used to envision, wireframe, and prototype interfaces do not scaffold designers in choreographing how an interactive flow should collapse as the system learns what users usually do. We want to change this. As a start, in this article we offer a draft of a design process and detail some common AUI design patterns. This is a first step toward turning UX designers into effective AI innovators around the use of AUIs.
Advances in AI have lowered the barriers for providing a range of AUI capabilities. These systems monitor users' actions and begin to automate actions as confidence in predicting user actions grows (Figure 1). AUIs can learn collectively from users, allowing adaptation to happen quickly, and they can learn the actions of individual users. For example, a ride-sharing app might quickly learn that most people want to go home when they access the app from their local airport. Over time, the same app might learn that you most often travel to work when you arrive at your home airport on a weekday before noon, and that you always want to share a ride when going home, but not when going to work.
|Figure 1. AUIs can learn to automate a frequent task by adapting to groups of people. They can also learn to automate by adapting to the behavior of an individual.|
One powerful benefit of AUIs, compared with other AI investments, is that they keep the user in the loop. They integrate a machine's ability to log actions and automate tasks with a person's understanding of context and logical action. This collective intelligence allows for the benefit of automation while lowering the risk that a machine's inference might inflict serious harm. For example, when a Web browser infers that an online form needs the user's name and contact information, it typically presents the user with a list of possible choices for how to fill out the form. When the user selects which set of contact information to use, the browser fills in as many form elements as possible. It then relies on the user to notice any remaining elements and to notice any elements it may have filled in incorrectly. As users complete the form and repair errors, additional training data is provided to help the AUI improve its future performance.
The idea of AUIs as a way to improve user experience has been around since the 1990s. However, advances in AI and the transition to online interfaces that log user actions have created the perfect environment for this technical innovation to move into the real world. This technology is just now starting to spill over into enterprise and business process systems. Email applications, which have a strong overlap between consumer and enterprise, are one example. Many email clients can now recognize and highlight actionable entities in messages, including descriptions of events that can be turned into a calendar entry, phone numbers that when tapped can initiate a call, and contact information that can create or update an entry in the user's contacts. These specific advances have recently moved more strongly into the enterprise through customer relationship management (CRM) systems, which now also recognize and highlight common entities like phone numbers and meeting information, speeding workers' interactions.
While AUIs are becoming more common in consumer-focused applications, adoption lags for enterprise applications.
Businesses can innovate with AI through predictive analytics, advanced decision support, next-best action recommendations, robotic process automation, and CRM systems with customer-facing personalization. These innovations garner attention but may require large investments to develop. Any of the above AI approaches will generate enough value to justify the investment. But unless the enabling conditions are right, they can be high cost and high risk. They often require AI that is quite complex, more plentiful labeled data than is immediately available, and process changes and skill upgrades that the organization may not always be ready for. In contrast, AUIs are often simple to implement and have an easy-to-calculate return on investment (savings in worker time). Additionally, AUIs complement other innovation approaches, offering an additional path toward more efficient and error-free information processing.
Innovation with AUIs takes place at design time, when development teams are creating new interfaces or performing a major overhaul of existing interfaces. They are different from many other AI innovation initiatives because the opportunity to use them is most easily recognized by UX designers. Instead of data scientists leading the collaboration, designers can discover opportunities for AUIs when conducting user research, when synthesizing data, and when wireframing new interaction flows.
The following provides a sequence of steps for UX designers who want to innovate their project with an AUI.
Step 1: Recognize opportunity. UX designers first need to recognize situations where a user will likely do the same thing over and over again. Opportunities show up at the context level and also at the detailed interaction level. Design teams should examine tasks that workers do daily. When a procurement officer first accesses their system in the morning, they might always look at an overview of all outstanding orders. An AUI could land them on this screen. Later in the day, they might always access an order based on the email they most recently viewed. An AUI could use the sender's information to predict what order they want, eliminating the need to search and select. When sketching a wireframe, the UX designer should start to notice that specific controls lend themselves to adaptation. For example, a pulldown menu can start with an intelligently selected value, and it can display a ranked list of likely items, before showing an alphabetical list.
Step 2: Qualify the benefit. For each opportunity, the potential benefit of the AUI should be qualified. How many times of day might the AUI help; how many workers would benefit; how much time would the collapse of a transaction save; and how much does the worker's time cost? AUIs work best for frequent tasks performed by lots of people. But they also make sense for workers whose time is really valuable—professionals like doctors and lawyers, where the cost of each transaction is much higher. Teams should try and estimate the return on the investment to help determine if it makes sense to create an AUI.
Step 3: Gather resources. AUIs, like all forms of AI, require labeled data for training the system to make an inference. Design teams should request access to telemetry data showing users' interactions with the system they wish to automate. Teams need to see if the data needed for the inference is available and if it is labeled or unlabeled. For example, if the team wants an AUI to recognize urgent incoming messages, they most likely have a collection of unlabeled data. They have a bunch of messages, but the data needs to be labeled to clearly indicate urgency. Alternatively, if teams want to land users on screens they most likely want to access next, labeled data in trace logs that show where users go immediately after entering a system may be available in the system.
Step 4: Estimate cost. Next, design teams need to estimate the time and cost for implementing the AUI. This initiates a collaboration with a data science or development team when these skills are not currently represented on a UX team. Design teams should share their AUI concept, share the data they think will be needed to train a system, and share their estimate of how the return on the investment will grow over time as workers increasingly save time doing their work. Repeated trips through this estimate step will help teams learn to recognize situations that will likely produce more value than cost and situations where this is less likely to be the case.
UX design teams are in an excellent position to innovate around AUI opportunities and to measure when it is worth the development investment.
Step 5: Construct details. Teams next work out the details of the interaction, using wireframing, design patterns, and other tools to explore ways to present an AUI to a user. Sometimes the interaction should deemphasize the automation, keeping the user's attention on their current task. For example, spam filters sort incoming messages as spam and not spam, subtly communicating that messages have been flagged as spam by increasing the count next to a folder icon. In contrast, search engines actively display a ranked list of what they think the user wants right at their locus of attention, making it faster for the user to select an item than to continue typing their query.
Step 6: Communicate to the development team. In the final step, teams document their AUIs for delivery to a development team. They need to communicate how an interactive flow collapses over time, as the AUI learns what the user wants. Documentation should detail both how the interaction works and also the conditions under which the adaptation should be triggered. How many examples should the system observe before taking action? How confident should it be before trying to help? How does it recover from error? These are critical questions for a team to investigate and communicate.
We have noted that common interaction design patterns exist for AUIs. We present three familiar patterns, noting that they have a broad range of forms.
Rank choices the user can select. Many AUIs present users with a ranked list of items they predict users want (Figure 2). This pattern helps users to easily see and select things from a ranked list. It most often appears right at the locus of the user's attention. It works well in situations where the inference is far from perfect, but where the top several choices in the list include the item the user wants. In most cases, logs from the previous interface will work as a labeled dataset, making these fast and easy to develop. In addition, the interaction of having the user explicitly make the choice produces additional labeled data that will continue to improve the quality of inferences. The interaction keeps the user in the loop.
|Figure 2. Examples of Ranked List of Likely Targets (clockwise from top left): Google Search's suggestions of likely search terms, Facebook's friend tagger, word suggestion for text messaging, email applications suggestions for message recipients.|
Highlight entities the user can act on. AUIs can detect entities, and they will often highlight these entities so users can take some sort of action (Figure 3). This AUI most often appears outside the user's locus of attention. The AUI highlights the item, making it easy to notice, but it does not force the user to act on the entity, making it easy for the user to ignore if they do not wish to take the suggested action. This AUI is quite different from other types of automation that often trigger alert screens that users must address before they can continue with their work. In most cases, this AUI is applied to an existing unlabeled dataset. For example, you might have an archive of email messages with event information in them; however, people will need to mark up the event information so the AUI can learn what it needs to detect.
|Figure 3. Examples of Highlighted Actionable-Entities include Grammarly's spelling-error detection, entity detection in email, and detection of time or date in messaging apps.|
Fill in forms the user can complete. This pattern appears in most Web browsers, which autofill people's contact information into most new forms they encounter. It also appears within most calendar programs, which automatically fill in event information when coming from an email or other messaging system (Figure 4). This AUI reduces time spent repeatedly entering the same information. It's like a barista that remembers your order and starts making your coffee when you walk in the shop. However, innovation teams need to be careful when selecting it, as people sometimes fail to recognize when the AUI fills in the form with frequent information, but not the information they needed for this transaction. This pattern is most often used where a labeled dataset already exists based on the history of earlier transactions. It can be slow to implement, especially for services a user does not use frequently, such as filling in a form every day.
|Figure 4. Examples of Web forms that need contact information and calendar entry from entity recognition of an event from email.|
UX design teams are in an excellent position to innovate around AUI opportunities and to measure when it is worth the development investment. AUIs offer low-hanging fruit, where a little bit of low-risk AI can have an impact. They automate repetitive tasks, making workers more efficient and freeing their time and attention for activities that require human intelligence.
Our collaboration is just beginning. We plan to advance our efforts in this area by developing new design methods and new tools that make it easier to recognize opportunities, qualify potential value, estimate cost, and sketch and prototype the interaction design.
John Zimmerman is the Tang Family Professor of AI and HCI at Carnegie Mellon's HCI Institute. He conducts research on human-AI interaction and has designed AI interfaces for more than 20 years. He teaches classes in UX design, service design, and the design of AI products and services. firstname.lastname@example.org
Changhoon Oh is a postdoctoral researcher in the HCI Institute at Carnegie Mellon University. His research on human-computer interaction and user experience focuses on human-AI interactions and how the intersecting UX/HCI and AI/ML perspectives have yielded meaningful and synergistic results in both communities. email@example.com
Nur Yildirim is a Ph.D. student and design researcher at Carnegie Mellon's HCI Institute. She designs human-AI interactions and develops tools and processes for designers to engage with AI as a design material. firstname.lastname@example.org
Alex Kass is a fellow and principal director at Accenture Labs, where he currently leads the Future Technologies R&D Group. Alex has a Ph.D. in computer science (AI) from Yale and over 30 years of experience working at the intersection of information technology, human cognition, and the future of work. email@example.com
Teresa Tung is a prolific inventor at Accenture, with more than 180 patents and patent applications. She specializes in bridging the enterprise gap, taking innovation from startups, academics, and digital natives and making it work at enterprise scale. She leads R&D in cloud, AI, and data architectures. firstname.lastname@example.org
Jodi Forlizzi is the Geschke Director and a professor in the HCI Institute at Carnegie Mellon University. Her current research interests include designing services and human-AI collaboration. Her design research has made contributions to social problems including eldercare, accessibility, human assistance, and overall well-being. email@example.com
AUIs offer an excellent opportunity for UX-design-led innovation with AI; however, designers should approach this with a commitment to ethical design. UX comes from a history of computing as work automation, a history that sometimes leaned toward Taylorism and a focus on productivity mattering more than a worker's humanity. In addition, working with AI raises a host of new ethical concerns. Below, we touch on four critical lenses for thinking through the design ethics of AUIs.
First, attend to the felt experience of work. Pay attention to workers' agency, their feelings of being in control of their work. Adding automation into a familiar task will impact feelings of control. Pay attention to their identity as a worker, asking if the use of AUIs makes workers feel better at their job: Does it make them feel more capable? These issues of agency and identity converge in self-efficacy, their belief that they have the capability to do the work and do it well. AUIs work when they automate actions that workers consider boring, repetitive, and tedious (BRaT).
Second, attend to the symbiotic relationship between AUIs as learning systems and the workers they learn from. Designers should notice if automation creates an unhealthy dependency. In addition, you want to notice who is working for whom. Does your design make workers feel that the AUI serves their needs, or do they feel that they work in service of the AUI?
Third, watch for unintended consequences. AUIs save time, but where does this time go? One promise of AUIs is that they can free people from tedious tasks so that they might focus on work that requires more human intelligence. But is this what will actually happen? Take a systemic approach and use a model of current practice to predict what will likely happen.
Fourth, monitor the emerging best practices for the design of AI systems. Several sets of guidelines have recently been published that discuss details of user privacy and system transparency, that note critical issues of algorithmic bias, and that talk about the need for AI explanations so users might more appropriately trust the systems they collaborate with. Privacy for enterprise systems raises specific issues, such as workers not wanting AI systems to infer their productivity.
UX designers should lead AI innovation in the enterprise through the creation of AUIs, but they should do this in a designerly way. UX has long championed a concern for the user and has expressed concerns for the larger impact of whole systems. Design works to produce a preferred future, and work with AUIs should deliver on this promise.
Copyright held by authors. Publication rights licensed to ACM.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2021 ACM, Inc.