Creating Chatbots and Virtual Assistants That Really Work

Chatbots and their more sophisticated counterparts, virtual assistants, have emerged as a key tool to streamline customer service operations for organizations in a wide range of industries, and to help reduce costs. Around the globe, consumers interact with chatbots and virtual assistants on their desktops, tablets, smart phones, and smart speakers. In fact, a recent Business Insider study predicted that by 2020, 80% of enterprises will use this technology. And according to Juniper Research, financial institutions intend to automate up to 90% of their customer interactions using chatbots by 2022. This is projected to result in an estimated 8 billion dollars in annual industry savings. Another survey conducted by Spiceworks showed that this year nearly half of organizations with more than 500 employees plan to implement AI-powered chatbots on their corporate mobile devices to optimize communications between teams and individual contributors.

Chatbots and virtual assistants – How do they differ?

First a quick rundown of the core differences between chatbots and virtual assistants: Chatbots work with structured dialogue. They are typically programmed to answer specific replies to specific questions and generally can’t reply to complex questions that aren’t programmed into them. Virtual assistants, on the other hand, mainly concentrate on natural language processing (NLP) and Natural Language Understanding (NLU) to respond to queries. Both technologies rely on correctly trained machine learning models to work as intended.

Obstacles to top performance

While chatbot and virtual assistant use is certainly on the rise, end-users are too often confronted with imperfect performance due to insufficient quality or volume of training data. This can mean incorrect or irrelevant information, which has the exact opposite effect of what was intended. Instead of streamlining customer communications and driving better brand loyalty, poor chatbot performance can cause additional, often negative contact with human customer service personnel. When the technology causes this kind of frustration, it can result in damaged customer relationships and a lower CSAT score. A US Internet Users poll from last year identified the top reasons customers get frustrated with these solutions and near the top is – you guessed it – too many unhelpful responses. Other significant concerns identified by the same poll include:
  • Prevents customer from speaking with an actual human being
  • Redirects to self-serve FAQs
  • Unnecessary pleasantries
  • Unnecessarily long response times
  • Incomplete data about customer
Some of these are straightforward programming issues. But what can truly be done to make chatbots and virtual assistants more robust, more accurate, more nuanced in their language processing capabilities, and – ultimately – more satisfying for end-users?

A big piece of the puzzle is better training data

To train machine learning models for chatbots or virtual assistants—for customer service, eCommerce, media delivery or another use altogether—you need massive amounts of training data. And Appen knows you need high-quality, accurately annotated data so your chatbot or virtual assistant can understand the nuances of language and written text.

Case study: Appen helps Flamingo create better automated solutions

An example of this is an engagement we completed for Flamingo AI, a company that specializes in “conversational commerce” technology for enterprises. The company’s artificial intelligence-based virtual assistant platforms help financial services and insurance companies automate online customer experiences for higher satisfaction scores, better sales conversion rates, and lower costs. Flamingo has seamlessly combined instant messaging, smart workflow, natural language processing, and machine learning to automatically serve customers online, guiding them through a chat-based sales process or service experience. Flamingo worked with Appen to ensure that their virtual assistants worked right out of the gate, without the need for human operators to initially work alongside them. Appen’s first project for the company was for an Australian Stock Exchange (ASX) 100 Australian financial services firm. Appen leveraged its global network of linguistic professionals to source a group of native Australian English speakers, and asked them to provide questions, comments or responses based on a specific step in a customer journey or process such as applying for superannuation. The questions Appen provided stood in for the real-world data Flamingo didn’t have time to collect, and when fed to the virtual assistant, successfully prepared it for the interactions it had once it was live on the customer’s website. As a result, the end customers’ experiences were greatly improved by interacting with the virtual assistant instead of using a static web form. To learn more,  read the case study here.

Let’s make customer service experiences better than ever

From data collection to transcription to linguistic rule development, Appen has the capabilities to help you take your virtual assistant or chatbot solutions to the next level. Our seasoned project managers become an extension of your team and can quickly ramp up the necessary resources to start generating results right away. With experience in over 180 languages and dialects, we can help you improve engagement for customers around the world. Learn more about how Appen can improve your machine learning solutions – including chatbots and virtual assistants – with high-quality training data.   At Appen, we’ve helped leaders in machine learning and AI scale their programs from proof of concept to production. Contact us to learn more.
Website for deploying AI with world class training data