The promises of today’s chatbots are almost dreamy if you are charged with leading customer support and service. Features like advanced cognitive technologies, native machine learning, natural language processing and generation sound very appealing. “Such technology can handle large data sets to process, evaluate, and respond to inputs, mimicking human conversation,” according to a November report by Deloitte on Conversational UI. “This enables numerous customer engagement applications, including common uses such as billing support, customer authentication, and FAQ responses, and more sophisticated applications, such as technical support.”
And it’s working for some organizations. 83 percent of respondents in Deloitte’s State of AI Survey said they have achieved moderate or substantial benefits from their work with AI technologies, and 94 percent said AI is very or critically important to their success.
However, organizations still have plenty to learn about using chatbots in customer support and services scenarios, experts told CMSWire. And they're still making mistakes with such implementations. “One common mistake is many organizations start with ‘proof-of-concepts’ that are designed poorly and fail to support a company’s business objectives,” said Derek Top, research director and senior analyst for Opus Research. “Many firms underestimate the resources and staff necessary to provide a successful chatbot implementation.”
Ensuring Accuracy of Chatbot Engine
A company’s first-order concern should revolve around the accuracy of the chatbot “engine,” meaning its ability to correctly recognize end-user intents consistently and at sufficient scale, according to Top. For many companies, accuracy is a moving target and highly-dependent on how much the data they have to ingest in the form of chat transcripts and other conversations, he added. "Contact centers are on the hook to manage a growing number of service channels, and chatbots are becoming a first-tier support option for helping augment the live support and drive efficiencies in the contact center,” Top said. "The goal is to engage customers, when appropriate, to support and answer user questions and take action to decrease costs and optimize service channels.”
The basic metric for chatbot success is like it is with all other tech-implementation initiatives: ROI. “Cost savings are core to the initial business plan that justifies intelligent assistance and chatbots,” Top said. “But chatbots are indeed proving their worth by making customers happy, making live agents or technicians more productive and helping to grow the topline.”
Related Article: Building a Useful Chatbot: 3 Questions to Ask
Support the Human Supporting the Chatbot
Seth Earley, founder and CEO of Earley Information Science, said too many organizations look for the magical technology solution that offers a chatbot or intelligent virtual assistant and hope it will automate customer services interaction once and for all, without much hand-holding. The better approach, he said, is “having a human in the loop” at the outset to handle complex challenges your bot can’t.
Earley calls it a hybrid approach. Don’t take a bot and turn it loose. Rather, carefully model use cases, take your existing content and knowledge base and transform it, by breaking it into components and tagging it appropriately for ingestion. “So what we're doing is we're identifying those different pieces, those different components of the intent,” Earley said. “And then we're matching it up with an ontology, with basically an information architecture and using that to retrieve a very specific piece of content. You can say it's a question-and-answer system, but it's a retrieval mechanism.”
Customer support agents should always be empowered with tools and resources that help them support the bots, especially in the beginning stages of the chatbot support implementation. “Once you get consistent behavior you can release it to the wild,” Earley said. “When you're actually interacting with a customer, you have to think about how that interaction is going to happen, and where you can have the human in the loop. The bot can interpret key intent. The human can check if it’s correct. And that becomes part of the learning.”
Start to evolve the chatbot's role once you get a sense of the repeatable tasks. “You can do all the mining, process analysis and look at prior call dialogues, recorded chat transcripts and call transcripts," Earley said. "And you can use that to identify the things that are the low-hanging fruit."
Getting It Right Takes Time
The initial learning stage with the bot should not come with expectations of solving challenging, open-ended tasks from customers, according to Earley. Don’t tackle the most complex and difficult scenarios at the outset. Use knowledge engineering tactics and distinguish between task complexity and knowledge domains, he said. “You’ll get into different levels of complexity, both in terms of task and the domain,” Earley said. “You don't want to do these open-ended questions. Make sure you have the ability to escalate to a human at any point, and make sure that you have a mechanism to correct the bot.”
Organizations will struggle with implementations like these for the long-term. A common mistake they’ll make if they want the short answer, and they want the quick fix. “They want the vendor that has all the magical solutions and none of that's true,” Earley said. “But to make this work, it has to be a concerted, well-thought-out, long-term roadmap. Too many organizations are trying to fix their knowledge management initiatives or their cognitive projects in three months.”
Related Article: Why I Hate Customer Service Chatbots
Determining User Needs, Bot Channels
One of the most common mistakes is not properly understanding the end user’s needs and designing an inadequate conversation, according to Matias Burcheri, VP of technology conversational interfaces, mobile & UI engineering studios at Globant. “Not having a clear goal for the chatbot design leads to an incorrect intent detection, so the bot ends up getting confused and the accuracy the bot has on solving queries drops significantly,” Burcheri said.
Learning Opportunities
Webinar
Mar
29
The CX Commute vs. the CX Joy Ride
Is your CX traffic holding up your customer's journey?
Webinar
Mar
29
Is Your Microsoft 365 Data Recoverable? How to Survive the IT Hotseat When Data is Lost
See why MIcrosoft recommends third party backup
Webinar
Mar
30
An Optimized Customer Contact Strategy Combines Transparency and Intelligence: The State of Outbound Communications
Learn from the Neustar-commissioned Forrester Consulting study on challenges and solutions for contact centers
Webinar
Apr
5
Adopt a Modern Digital-First Marketing Strategy
Learn how to use a composable DXP to your advantage
Another widespread mistake is not deploying the bot to the proper channels, Burcheri added. Organizations, he said, need to be where their customers are, whether that is website chatbots, mobile, WhatsApp or other options. “Call centers were born in an age where the main communication channel was the telephone, but nowadays, new ways have appeared, and the relevance each channel has is different,” he said.
Do You Have the Right KPIs?
Not having proper KPIs and metrics around the chatbot is another common mistake, Burcheri said. Often, users will know they reduced the number of phone calls in call centers but lack information and metrics on improved user satisfaction, increased demand or customer retention. “Thinking a chatbot is just a way of reducing human agent’s occupancy is a fatal error for many organizations,” he said. “As with any other communication channel between the organization and its customers, you need to have a clear picture of how it is performing and that you are being able to satisfy your customers’ needs.”
Related Article: What Comes Next for Chatbots and Customer Experience
Purpose, Scripts and Flow
Implementing a chatbot entails three steps: Establish the purpose, build the scripts and build the flow, according to Burcheri. All these must be aligned to a set of KPIs that will measure the performance of our chatbot, he said. “When we are establishing the purpose, we need to ask ourselves, ‘Why will people want to use it?’” Burcheri said. “We need to define what a user will be able to do and what is going to be left out of the scope. We must understand which business cases are more reasonable to have an assistant for. We must empathize with how the assistant will be invoked by users, in what channels it makes sense for the app or skill to be available.”
In building scripts, design the conversation between the user and your assistant. Determine how this conversation will flow and the information required while using the assistant. “We should be aware if the channels provide visual support or only text or voice and design accordingly for each channel,” Burcheri added.
As for flow, that’s when implementation turns to Natural Language Processing (NLP) and session management. Determine and build the required end points on your backend to fulfill the requests of our users. Write down behind-the-scenes decisions the system logic will have to make and finally build test cases and a test plan. Chatbots require specific skills for testing and during testing, Burcheri added, and new scenarios may arise that should feed the design process or even retrain our NLP model.
“While building the bot you must add all the necessary data capture points that will guarantee you are collecting the information needed to keep track of the KPIs you defined,” he said. “Not having the proper metrics could lead to customer dissatisfaction and disengagement."
Chatbots Are a Retrieval Mechanism
The bottom line is that a customer support chatbot is an information retrieval mechanism. “When I talk about chatbots I say they are channel,” Earley said. “It is a retrieval mechanism. Search is a retrieval mechanism and an intelligent virtual assistant is a retrieval mechanism. And at the end of the day you need some mechanism that says, 'Tell me what you're looking for, or give me a signal.' And the signal is the intent, but the intent can also be enriched by other signals.”
And it isn’t easy work. "There's a lot of work in refactoring the content," Earley said, "and building the right information architecture and right content... You have to use knowledge engineering approaches.”
