When it’s Smart to Play Dumb: Managing AI Recommendations
As machine learning and artificial intelligence (AI) evolve and begin to yield interesting results, our clients are exploring how to apply these technologies to their products and services. Their goal is simple: improve the customer experience while decreasing customer service costs.
But what does this even look like? With enough information about you, the customer, AI can provide accurate recommendations. Or, it can go a step further and take action on that information and clue you in later. So when should an AI-powered digital product check in before it does something, and when should it take matters into its own hands?
With big data comes big opportunities
Our clients know more about their customers than ever before. With the rise of AI, it’s natural that they want to take advantage of this information to create more effective services for their customers. Digital experiences are ideal for delivering personalized services because they can dynamically adjust both content and personal information based on customer needs.
If a company knows what you want, should they provide you only with the options you care about and nothing else? Going a step further — if the company is so certain, why not simply act on your behalf, rather than provide options?
An AI-powered digital experience could have the data necessary — with an incredible level of confidence — to know what you’ll actually prefer. So why not be where your customer wants you to be and provide an experience that seems magical?
While working on an appointment-scheduling feature for a client, we took a step back to think about this question, considering:
- What fundamentals must this experience have to be successful, and what would its users want?
- When does AI-powered personalization cross into “creepy, thinking machine” territory?
Based on this brainstorming, we developed two approaches: “ask, then act” and “act, then explain.”
“Ask, then act”
When the back-end of the experience compares a customer’s situation against the data it has available, it will generate numerous options — and then ask the user which option they prefer. This is the default mode for many experiences. You see it in the “People who bought this also bought” on major retailers’ websites or the suggestions to add other people to a message in email applications. It’s so innocuous at this point that we hardly notice it, but the experience is trying to unobtrusively make your decisions easier.
Every time you act on the suggestion, the AI says “Aha! I was right. Make a note to improve future suggestions!” In the experience we were designing, the “ask, then act” approach had us making suggestions for a new appointment date, time or location based on previous appointments the user had created.
One potential problem with the “ask, then act” approach is that AI suggestions can trigger “alert fatigue” — such as when your smartphone tells you how long it will take you to get home when you’ve just gotten to work, or when you dismiss a dozen notifications irrelevant to the work at hand.
This noise and clutter takes effort and time to dismiss. Even with the best intentions, there can be an exhaustive level of interaction needed to simply get back to whatever you were doing before you were interrupted.
“Ask, then act” pros and cons
- Eliminates creepy factor
- Creates a validated data point
- User answers imply permission or acceptance
- It can still be wrong
- Not providing the right actions
- Not asking the right question
- When an experience admits it doesn’t know, it implies it’s not smart
- Wastes time — if you know something, why not just do it?
“Act, then explain”
In an “ask, then explain” instance, when the data in a scenario points to a clear solution, the AI simply acts. For example, when you use navigation applications, you’re given the best route, with explanations of why other routes are not optimal. Ride-sharing apps assign you a driver based on your location and your previous ratings. In this client project, our experience could have used a number of factors to schedule the user’s appointment, then explain what it did and why.
When “act, then explain” works, it’s great. But what if it goes wrong? We’ve all experienced a route from a navigation app that technically might be shorter, but is overly complicated and feels longer. Or when a driver-matching algorithm assigns you a curt driver based on your previous activity and you find yourself wondering, “What did I do to deserve this?” The whole thing starts to feel arbitrary and ill-informed. Before you know it, you’ve been alienated from an experience that was great only yesterday.
“Act, then explain” pros and cons
- Simpler experience
- Impressive experience
- Getting it wrong
- Misaligned expectations leading to alienation
How to choose: value vs. consequences
For our client project, we were challenged with figuring out which approach might work best for a given situation: “Do you want me to schedule your appointment? What day? What time?” or “Here’s your follow-up appointment. Click here to make changes.”
So, which approach is better? When should a product or service check in before taking action, and when should it just act outright? As you may have guessed, it depends. The key is understanding value versus consequences.
If there’s not much at stake but a lot of value for streamlining the experience, we say “Let ‘er rip!” Act first and explain later. This provides a streamlined and satisfying experience for users. They enjoy the added bonus of getting to their goals quicker because the experience just made a series of decisions, intelligently, on their behalf.
When the consequences of an incorrect action are higher, it’s best to ask first before taking action. If the outcome could be something like poorly invested money or damage to personal or professional relationships, then there’s not much value in streamlining the experience. Better to play it safe and ask a couple of questions before getting to work.
Illustrations by Citizen art director Doug Merritt