Companies are awakening to the value of conversations with consumers. Simply put, when consumers can easily converse with a brand, they are more likely to develop affinity for that brand and to buy from them, thus the term “conversational commerce.” Brands are using conversational interfaces to deliver customer experiences for sales and service, but there's a catch: designing these conversations may be harder than they initially thought.
For consumers, chat and messaging are often easier to use than elaborate visual interfaces, especially on phones and smaller-screen devices. In the home, in the car, and on the go, voice assistants like Amazon Alexa, Google Assistant or Apple Siri offer the convenience of hands-free input and less-distracting output. As such, Gartner predicts that 25% of customer service and support operations will integrate virtual customer assistant (VCA) or chatbot technology across engagement channels by 2020. Forrester calls this technology “conversational AI,” and it is one of the key drivers for customer service automation.
Many tools and platforms are now available for building chatbots and speech bots, from large tech giants — such as Google, Amazon, Microsoft, IBM — to startups with novel development paradigms or industry-specialized applications. These tools offer the promise of automating conversation development, including aspects such as natural language processing, flow management and speech recognition. Despite these technologies, conversation design, the crafting of human-to-machine dialogue that accomplishes a desired task, remains a human-led discipline.
Conversation Design: A New and Scarce Skill
Conversation design is critical to realizing the benefits of conversational commerce. Without design thinking, conversational systems can result in failures that frustrate customers, erode the expected ROI and possibly damage the brand. Unfortunately, conversation design skills are scarce, and not readily transferrable from the widespread web and mobile visual design skills.
Before the recent rise of conversational interfaces, authoring dialogue and prompts was a niche skill in most enterprises, confined to Voice User Interface designers responsible for the IVR (Interactive Voice Response) system. To expand the pool of designers, companies have turned to practitioners in adjacent fields with experience with human dialogue and verbal communications, such as video game designers, copywriters and script writers. Without access to these resources, companies will have to partner with specialist service providers, or build up internal design skills through trial-and-error.
Related Article: Calling All Linguists: The Messaging Bots Need Help
Start By Understanding Customer Intents
In customer service, good design is the ability to resolve requests at a high rate with high satisfaction. Successful resolution has two phases: intent determination —understanding the customer’s reason for the contact — followed by intent handling. Although domains such as entertainment or social communications may have an almost unbounded set of intents, customer conversations with a particular brand have a finite set of intents (in our experience, usually less than 200) that depend on the brand’s industry, products and services, and business model.
During the intent determination phase, conversation design requires one or more of asking, clarifying, disambiguating and confirming the customer’s intent. To understand intent, machine learning models are trained from examples of actual conversations that have been tagged by human analysts. These analysts, who are subject matter experts in the business, read call and chat transcripts and determine the appropriate intent for each customer utterance or message. Contact center agents can also be used to supplement (or replace) business analysts by tagging conversations as they are taking place, or soon after they finish.
Beyond conversation data, intent models can also be trained on additional data such as behavioral signals (e.g., pages the customer has viewed), enterprise signals (e.g., status of an order or account) and exogenous signals (e.g., weather or local events). These richer models enable smarter intent predictions: for example, the utterance “I have a problem with my order” when the user is also adding a discount to her shopping cart would indicate the intent “Order/Promo/Apply,” whereas the same utterance when the shipping system shows an overdue delivery would indicate the intent “Delivery/NotReceived.”
Related Article: What Is Conversational User Experience?
Handle Intents by Learning From Agent Conversations
Once the intent is understood, the conversation needs to be guided to a successful outcome, which may be answering a question or completing a transaction. During this process, the conversational interface should address both the consumer’s expectations and assumptions as well as the company’s priorities and constraints.
The data-driven approach to conversation design is to analyze a large sample of human conversations for each intent. These transcripts reveal the different ways that customers express problems and respond to instructions, and the ways that agents diagnose and resolve issues. Unsupervised machine learning techniques exist, and can be used to mine conversations and identify the common patterns of dialogue flow between customers and agents. To remove noise and avoid misguided training, the sample set should be filtered to select conversations from the best agents, as measured by resolution rates and customer satisfaction ratings. The discovered flows then give conversation designers a starting point for writing dialogues.
Many conversations have more than one intent: the consumer may want to perform multiple tasks, or may see one intent as a path to another intent. For example, “Password/Reset” is a common intent, but the reason for resetting a password is almost always to perform some other action. In the retail industry, we have found that the intent “Order/Status” is usually followed by another intent, such as changing the delivery address, changing order items, switching to store pickup, canceling the order, asking for a refund, or any one of over 20 possible follow-on intents. To fully automate an interaction, conversation designers must incorporate intent sequences in their bot design. If the bot is unable to handle the second and subsequent intents, the customer will have to escalate to a human agent, which increases the cost of the interaction. And if human agents are not available, the customer is left with a partially complete interaction which is probably even worse than no interaction at all (because he may feel his time was wasted).
Conversational AI is about creating a human-like experience that helps consumers feel connected to the brand, whether interacting in voice or digital channels. AI is essential, but not sufficient. Thoughtful conversation design is required to make AI intuitive and truly conversational, and to turn visitors into engaged customers.
Learn how you can join our contributor community.