Top 8 Learnings After Years of Conversation Design

By Replicant
November 23, 2021

Written by Benjamin Gleitzman, Chief Technical Officer and Co-Founder at Replicant

Conversational AI remains a nascent discipline, and although Replicant sits on the leading edge of this transformative technology, the field is evolving faster than I could have imagined even a few short years ago.

The concept of user experience design, centered around the physical ergonomics between humans and machines, has been a legitimate field of study since the early 1940s.

With the advent of modern computer equipment, user interface (UI) design is also of venerable provenance, most iconically on display in Douglas Engelbart’s “Mother Of All Demos” in 1968 that simultaneously introduced the application window, graphics, video conferencing, and the computer mouse (not to mention word processing, revision control, and real-time collaborative editing).

From Xerox PARC and the Apple II all the way to Windows XP and mobile devices, the “language” of graphical interface design is largely a book already written. We all know, intuitively, where the “OK” button should be placed and how to close a digital window.

Conversational AI has gone from a relatively unknown technology to a mainstay in customer service, and we can see the evolution of this discipline play out specifically in contact centers.

It’s changing the way customers get their issues resolved and bringing new efficiencies to the forefront while drastically improving the customer experience. It’s also become a game-changing tool for contact centers to unlock deep customer insights.

But this evolution hasn’t been without its lessons.

Unlike UI design, there is no already-written book or one-size-fits-all approach to implementing conversational AI in customer service. Expectations and best-case scenarios that may work in a sandbox environment can create more problems than solutions when rolled out to customers at scale, leaving callers frustrated and developers scratching their heads wondering where they went wrong.

I’ve learned that successful conversation design, implementation, and deployment comes down to thinking like a customer and creating conversations that don’t force callers to fit a square peg in a round hole.

It’s a methodology that doesn’t put words in the caller’s mouth and adapts to accommodate even the most unique voice on the other end of the phone line. It’s a unification of the aesthetic utility of conversation design with the actual utility of the powerful AI models that allow the machine to listen, think, and speak.

Here are a few of my learnings and best practices for implementing best-in-class conversational AI:

Trust is key.

Tell the truth about speaking to a machine, with no gimmicks or crutches to cover latency or inaccuracy. A machine that introduces itself with a human name, or uses computerized typing sounds to mimic a call center agent, only confuses the caller and adds time to a conversation – not value.

You don’t know your conversation drivers until you start measuring them.

You should be making decisions based on data rather than relying on hunches or incomplete agent dispositions. Successful conversation design learns from every interaction to better understand what customers want and provides an unbiased view into what use-cases and flows to automate next.

The “happy path” is only the beginning.

According to Norbert Wiener, MIT professor and originator of the field of Cybernetics, “the great weakness of the machine…is that it cannot yet take into account the vast range of probability that characterizes the human situation.” As a corollary, if you only design for what you expect to hear, your callers will end the conversation more frustrated than they began (and likely never call you or use your product again). You need to be able to change the subject, handle multiple intents, collect data while on hold, think ahead, and reference things you said earlier. These are “easy” for humans but very difficult for machines, and can make all the difference between success and failure. A well-rounded conversation design accounts for every scenario to avoid escalations or confusion.

Each customer has a “frustration meter” and when you’ve filled it up they’re mashing 0 or asking for an agent.

Suffice to say that before Replicant, no one had ever had a very good conversation with a machine. Callers may get frustrated simply after hearing they’re speaking with a machine. Rather than lie or hide the fact, you should push back with the machine’s true value (“I can get you to an agent, but it looks like there is a 15 minute hold time. In the meantime, can I get your policy number?”) The customer will be in and out before they could even get to an agent and will avoid being locked in “bot land”. They might even learn a thing or two about how smart and fast a well-designed machine can be. If customers want to escalate, let them, but ensure you pass context to the agent so callers don’t have to repeat their account number or why they are calling (I hate when I have to do that!).

Meet the customer where they are.

I think of the family in Detroit that can’t speak to their smart fridge because it doesn’t understand their accent. Or the cashier in Quebec with a broken POS system who must wait on hold for customer service from a human agent because the automated machine doesn’t understand their vernacular. Meeting the customer comes in many forms that may include jargon (intent models that handle colloquial affirmative phrases like “you got it, boss”), accents (custom transcription models that are well balanced and equitable in understanding different pronunciation), common entity extraction (that’s Gleitz – G as in glass, L as in lightbulb, E as in eagle…, or even phonetic fuzzy name matching for people with names like mine that may be difficult to spell correctly), and preferred channels (switching to SMS and Flex Form for credit card collection). A conversational AI is only as good as its ability to handle the most unique customer, not the most common.

Deploy continuous improvement…

…to learn from every conversation, better handle edge cases, and prevent conversational flows from getting stale. This industry moves rapidly, and so do customer expectations. Conversation designs should automatically improve through machine learning (a process we call “kaizen”) so a customer never runs into the same issue twice.

Without test cases you are lost.

When it’s time to release a new version of your conversational AI, deployments should be validated with a full suite of regression tests gathered from previous successes and failures. Even simple changes in scripts and call flows can confuse customers and tank success rates. When testing changes, gather feedback from team members or friends who don’t speak or think like you do. If you want to try something new, A/B test it to validate your assumptions. It’s the most effective way to turn theories into action and ensure your designs are based in truth and equality, not biased opinion or personal experience.

Containment isn’t everything.

It’s easy to build a terrible experience with excellent containment. Planned transfers and escalations should be part of your gameplan. Measure how successful you are at getting customers where they need to go as efficiently as possible, and make sure when the live agent takes the call you are passing context and providing an automated screen pop so agents can quickly get up to speed where the machine left off.

*Bonus: Outbound Calling*

Make sure there is a business case.

Or, no cold calls. Customers often have a negative view of dialing machines as it is, and every word spoken by conversational AI should provide value to customers in solving a problem, not creating a distraction or nuisance via outbound cold calling.

Give something to get something.

Don’t open the call demanding personal information. Explain who you are (a machine calling on behalf of…) and why you’re calling. Include information that only the business would know before asking for any personal identifying information. There are many out there using conversational AI for nefarious purposes and it’s important to train those you call to practice good habits.

Make sure you can navigate IVRs that may answer the call.

Before you get to a person, your conversational AI might need to press some digits. Make sure this can be done dynamically. Before we know it machines will regularly be speaking with machines. They should both use language – it’s easily understood and auditable. Voice is the API.

Be quick!

Collect multiple pieces of data on the same turn. Be understanding if the person you call doesn’t have time to talk. If necessary, don’t force the issue. Be polite, end the call, and try again later – you’ll probably reach a different person at a better time.