Chatbot Development: YOUR GUIDE TO GETTING IT RIGHT: The race is on to include chatbots in marketing and CRM efforts, but many companies still aren't getting it right--and these tips can help.
Remember that a chatbot is not just a way to deflect calls and enhance customer service; it will also reflect your brand, experts say. Your immediate goal, for example, may be to empower customers to self-serve and to reduce call volume. Your chatbot will be standing in as the face of your brand, and that means it needs to faithfully reflect brand messaging, product marketing strategies, and the quality of experience your customers have come to expect. However, according to a Pitney Bowes white paper, chatbot failure rates can be as high as 70%. A failure is defined as any time a chatbot cannot complete the query, sending it to a human.
With all of this in mind, we talked to several experts to get their thoughts on the key elements of chatbot development at every stage of the process.
Define Your Chatbot Strategy
"In order to ensure a successful chatbot deployment, businesses need to first define their ideal business outcomes," says Jen Snell, vice president of product marketing at Verint Intelligent Self-Service. "That might sound obvious, but in the rush to deploy conversational AI solutions, it's an essential first step that is too often overlooked. Any successful technology implementation should support your bottom line. Chatbots are no different. Your KPIs need to directly reflect your overall goals and broader business strategy."
For most, that broader business strategy will mean some combination of chatbots and live agents for call center interactions, said Verint president Elan Moriah, Verint CEO Dan Bodnar, and several other speakers at the Verint Engage19 conference in May. They agreed that the best return from chatbots is to use them to handle the mundane, routine interactions (e.g., "What is my balance?") so that human agents can handle complaints and other, more complex issues. Due to the increasing interactions from social media and other channels--some estimate a 350% increase in customer interactions--companies can't rely on humans alone but instead need chatbots to handle the increasing number of communications in an always-on era.
It's critical to understand how your customers want to interact with you, and how you can most effectively deploy chatbots to meet their needs. "Don't assume you know what your customers want, or how they might be trying to get it," Snell advises. "Look at the data, then design and deploy your chatbot in ways that address the reality, not your assumptions, about what customers want."
Get Specific to Deploy Quickly
Another consideration, according to Brent Schneeman, senior director of engineering at Alegion, is whether the company will use several chatbots, each developed for specific services--such as payment--or fewer, more generic chatbots that will handle a wider range of subjects.
Snell recommends determining which areas automation will provide the quickest return for and developing chatbots for those first. "Define the primary users and the different types of experiences that you want to offer to each user. Then organize the content. Context is the key. The better that you can develop the context, the more valuable that you make the chatbot--people will be able to use it better."
To determine which deployments will yield the best returns, Snell recommends that companies consult individuals from marketing, communications, IT, business development, and customer service when designing chatbot strategies and implementation, to ensure that the solution meets the needs of the entire enterprise and matches your vision and goals.
Snell says it's vital to clearly identify both the quantitative and qualitative outcomes a company wants. Quantitative data gathered through common search terms or web page traffic prior to designing and deploying the chatbot can be used to generate systematic analyses regarding customer behavior and expectations. "By keeping [the above] tenets at the forefront of their deployment strategy, organizations can ensure that the solutions they deploy will generate desirable and profitable outcomes," Snell says.
Another consideration, according to Schneeman, is whether the company will use several chatbots, each developed for specific services, or fewer, more generic chatbots that will handle a wider range of subjects. Some chatbots are extremely focused, designed to answer only a handful of questions, though natural language understanding enables these chatbots to answer these questions asked in different ways (e.g., "What is my balance?", "How much is my balance?", etc.). Natural language understanding is critical for chatbot success, Snell agrees. The chatbot needs to be able to understand the various ways questions are posed, using clarifying questions when necessary. The answers to the clarifying questions can be used to further train the chatbot to recognize the initial question for future similar queries.
The advantage to subject-specific chatbots is that they can be trained and ready to go live sooner than a chatbot handling a larger range of issues, says Cheryl Martin, chief data scientist for Alegion. However, these "point" chatbots also must be able to recognize when the customer is asking about a non-core (for that chatbot) subject, then quickly and seamlessly hand off the call to the correct chatbot for the customer's new query.
"For each one, the overall umbrella is: 'Is the current state of the conversation something that you know about?' If the inference is that the conversation should go to another chatbot, then it is routed to that one," Schneeman says.
But keep in mind that if a chatbot's capabilities are too narrowly defined, the complexity of managing multiple chatbots can hinder the efficiency of the automation, according to Snell.
"Organizations can overcome this by establishing a long-term business and IT vision," Snell says. "This vision should consider how chatbots and associated technologies can support business goals far beyond the first deployment. Additionally, businesses should prioritize the curation of a collaborative culture among their employees. At the end of the day, bringing technology out of its silo will only work if the organization isn't broken into isolated divisions."
When designing the chatbot, the company has to determine the logic of the message flow--how the message moves from one step to the next, says Jeremy Pollock, principal product manager for PubNub. "It's important to have a logic layer separate from the chatbot layer. As you move to a conversational approach, you have a bridge from the chatbot to all of the data."
APIs provide the necessary bridge to the data, Pollock explains. "You need to look at how you want to route the messages--the different topics and the different channels that you are going to want to interact with. Topology is very important. You need to understand the usage. You need to identify the most active chat customers."
It's also important to understand who will be the most active users of the chatbot and develop a topology that serves them best, Pollock explains.
Integrating Your Chatbot
Many chatbot projects fail because they lack the proper integration with ERP, CRM, or other systems where the underlying data lies, so they can't provide the requested data, or can handle only very limited requests before they need intervention from a human agent, says Simha Sadasiva, founder and CEO of Ushur. "Most are meant to deflect phone calls. Every time you need human intervention, it is expensive."
It also impacts customer satisfaction when chatbots fail to do their job or to find someone who can. "You need intelligent automation so that the chatbot can do more than just a simple task," Sadasiva says. "It needs to do more than just understand what the question is about, but also needs to understand workflow and what the customer is interested in."
A good chatbot will orchestrate conversations without a lot of IT infrastructure, Sadasiva says. "You need to have intelligence embedded within the chatbot."
Consider too, that a great conversational self-service solution needs to be integrated with multiple systems of record in order to self-serve across your entire site. By including stakeholders from across the organization, you can better ensure the integrity of each department's role within the overall customer experience.
Go Live, Collect Data
Once a company goes live with its chatbot, the next step is to continue to collect data from interactions to attempt to improve the chatbot, experts agree.
"Modern companies are always monitoring and measuring," Schneeman says. "They look to optimize, optimize, optimize and test, test, test. They never stop trying to improve. They are always monitoring and expanding what they can do with their chatbots."
The qualitative information will come from chat logs after deployment and can provide invaluable insights about how customers are asking for what they want or need, according to Snell. "This will help you better understand your customer's intentions and adapt your systems to better serve them over time." That might mean finding new ways to answer questions.
For instance, your chatbot should connect to available multimedia within the organization, Sadasiva adds. You'll be able to go beyond just delivering voice or text responses and start including videos, relevant articles, blogs, and other information from the company's own knowledge base.
Chatbots can also be deployed to deliver answers, knowledge base articles, and other information either for the agent's own education or for the agent to forward to the customer. For example, the chatbot can be programmed to send agents spec sheets or a how-to video on a product fix when a conversation starts moving that direction. The agent can go over the details with the customer, and forward the information if warranted.
Another element that can make a chatbot more valuable to the organization is inbound-outbound message manipulation, according to PubNub's Pollock. "Companies usually run inbound messages through their own custom code." If, instead, an open code is used, it can be better manipulated depending on how many times a particular question is asked. For chat and chatbot applications, proprietary backend code is often used for content moderation, profanity filtering, and machine learning-based language translation, Pollock explains. This adds undesired latency to the chat experience, in that typically many more network hops are involved in the chat data flow, but just as important, this approach generally results in less adaptive logic. With a serverless compute module--available at the edge of the network, where messages flow into the chat and chatbot applications--it becomes possible to separate out the low latency logic into easy-to-modify functions.
For example, a chatbot's translation feature can be easily modified if the code is running on the network. A company can modify it and incorporate routing logic so that different inbound messages can follow different paths--e.g., one may find that Amazon's translation service is better for Spanish and Microsoft's better for German, and so, based on message properties, messages of different languages are routed to different translation services.
While chatbots have proven successful for many implementations, there are still pitfalls that can lead to failure--or at least, much lower-than-expected returns.
When first deploying a chatbot, organizations often make the mistake of siloing the project rather than considering the technology as a broadly capable business solution, according to Snell. As a result, these companies confine their understanding of chatbots as purely technological assets when chatbots have the ability to actively support teams and their goals across an organization. That's why it's important to include stakeholders from across the organization in the initial development stage.
But experts agree that any chatbot deployment should start small so that any holes in the logic layer, integration between applications, and so on can be quickly recognized and corrected.
"You need to have a flexible integration framework; you need to determine what services you will need to interact with for customer business reporting and for accessing the customer data," Pollock says. "You also need to have the right analytics tracking technology. You can leverage cloud providers that have tracking in their platforms."
Continue to Refine After Deployment
Organizations need to continuously review user interactions with the chatbot to determine not only if it is functioning as expected, but also if there are ways to enhance or expand the performance, Snell says. "Evaluating performance for your chatbot can be difficult, as many operate on the scale of 200,000 unique customer interactions per day. Clearly, assessing success when the dataset is that large requires more work than any human could reasonably manage. This is why a blend of human and machine-driven analysis is crucial in any chatbot deployment."
Combining the machine and human analysis enables businesses to create accurate (for them), actionable benchmarks, even with huge data goals, Snell explains. "Consider an example when a chatbot is assisting a customer with purchasing a plane ticket. Machine analysis can determine whether the user successfully purchased a plane ticket by the end of their chatbot conversation. Human analysis can look for opportunities for future conversation enhancement, including upselling of services."
Though the success metrics for every company will be different because every company will have different goals for its chatbot deployments, there are a few basic elements central to most of the implementations. Does this chatbot drive revenue? Does it foster cross-team collaboration? And finally, does this deployment streamline existing processes? If you can answer these questions, you're on your way to chatbot success.
BY PHILLIP BRITT
Phillip Britt is a freelance writer based in the Chicago area. He can be reached at firstname.lastname@example.org.
|Printer friendly Cite/link Email Feedback|
|Publication:||Speech Technology Magazine|
|Date:||Jun 22, 2019|
|Previous Article:||DIAGNOSING DISEASE with Speech Analytics: It's now possible to detect everything from depression to Parkinson's disease with speech analytics, but...|
|Next Article:||Higher Learning: AI, ML, and Speech Tech in Academia: Academics are at the forefront of the biggest AI changes in the industry.|