Member preview

Conversational UI: Five Danger Zones

Common pitfalls in chatbot design, and how to avoid them

Whether you’re building a chatbot or extending a voice assistant, we’re still in early times for conversational UI. As I’ve prototyped my own bots, built skills, and consulted on others, the guidance I give to other designers has coalesced into a few key areas where particular caution is warranted. Let’s explore five common areas of difficulty found while designing for conversational UI, and a few considerations for emerging from these danger zones unscathed.

Proceed with caution when approaching these five dimensions of conversational UI.

1: Proactivity

As with many before you, you may be tempted to lean into proactive prompting: dialogue delivered to a customer out of turn. Traditional conversational UI is delivered in immediate response to a customer request, but proactive dialog might start a conversation instead of responding.

But before leaping into the proactive void, consider the context. Are you building a bot that lives within another app, like Slack, Teams, Skype, or Messenger? If so, keep in mind your customers probably CANNOT disable notifications for those apps — they need those notifications for the app to provide value. Consequently, if you don’t respect your customer’s boundaries, they’ll have no choice but to cut your bot.

Generally, proactive notifications are only appropriate when the customer has signaled their desire for that notification beforehand. Learn from Clippy: wait until asked to help. And if you must be proactive, be smart: do so based on signals from your customer that they may be receptive to your engagement.

Ghosts of conversational paperclips past…

2: Verbosity

Keep in mind that the conversational apps that will host your bot are usually in highly constrained environments, often running on a small mobile screen or a partial-screen desktop snap alongside more productive work.

Also consider that the conversational expectation in these environments is typically crisp sentences, and no more than 1–2 sentences per prompt. You can chain prompts if necessary, but long prompts seem out of touch and are hard to scan.

Resist the temptation to overwhelm the customer with text. The terseness of voice UI can be re-applied to good effect here.

Instead of providing all of your content up front, apply the principle of progressive disclosure. Retain the context of your customer’s original request, and use that to inform follow-up responses.

3: Graphical UI

Yes, most conversational apps provide graphical patterns like buttons, cards, and even pickers. Use them cautiously. If your chatbot relies on these elements heavily, you may run into several problems:

  • For desktop applications, you’re forcing your customer to remove their hands from the keyboard, which is inefficient and counter productive.
  • For customers using a screen reader, over reliance on graphical elements may defeat the benefits of providing a text-based interface.
  • There’s very little “conversational” about these controls — if you’re going to attempt personality, these controls may ruin the illusion.

Make sure every question gets a complete text-based answer, and is only supplemented by GUI. Never require customers to switch from keyboard to mouse.

A red flag should be any time your conversational prompts refer directly to UI. If your prompts use phrases like “menu”, “click”, or “swipe” — you’re now describing GUI, not having a conversation.

This GUI toolkit for Facebook Messenger includes many useful visual affordances — but the more we rely on them, the less conversational our interactions become.

4: Specificity

Are you assuming your customer will refer to objects by name? This might make sense when referring to a person or an airline, but what about customer-named objects? “Ecommerce_Prod_WestUS2” doesn’t exactly fit in a conversational context.

If your conversational UI requires selection and manipulation of objects or concepts, consider more context-driven ways to allow customers to choose objects.

Once they’ve chosen an object, remember that context until it’s replaced or times out. Be intentional about that timeout — under what conditions will your customer expect you to remember content they’ve specified, and when will it be acceptable to forget?

Context is key to helping customers manipulate objects at enterprise scale. Don’t assume your customer can name the sprinkle in the pile — help them find it another way.

5: Authenticated use

The conversational app your chatbot connects to will have a large effect on your authentication story.

  • Is your customer’s account already tied to their credentials for your chatbot’s host app, or do they face an extra authentication step?
  • Does the level of customer trust in your chosen host apps affect your customer’s willingness to authenticate with your own experience? (see: Facebook and Cambridge Analytica)
  • Have you provided a strong enough value proposition to convince customers to jump through that authentication hoop?
  • What can you do if authentication fails, or hasn’t completed yet? Is there any value you can provide unauthenticated?

Don’t forget to build an onboarding experience for your app, giving customers a “why” and a “how” to get them to a fully authenticated state.

Bring your conversation to life

Throughout it all, view your conversation with a user as a journey. Otherwise, we’re just throwing conversation hearts at them, hoping we’ve chosen the right one.

A cautionary tale, told in stale sugar and disjointed phrases

None of these pitfalls is an unsolvable problem, but they ARE traps if not approached with considerable caution. They are risks… and opportunities. Let’s keep the conversation going.

Cheryl Platz has worked on a variety of voice user interfaces including the Echo Look and Echo Show, Amazon’s Alexa platform, Windows Automotive, and Cortana. She is currently Design Lead for the Cloud + AI Admin Experience team at Microsoft. As founder of design education company Ideaplatz, Cheryl is also touring worldwide with her acclaimed natural user interface talks and workshops.