close

Welcome to Synpulse’s digital reading experience – Please choose your region of interest

The Magazine
Management. Expertise. Inspiration.

Date: 14/11/2017

Title: How to Efficiently Implement Chatbots

Teaser: Recent advances in artificial intelligence enable the development of efficient chatbots that understand natural language. The financial services industry is becoming increasingly aware of the emerging opportunities.

Button: Read more

Image:

How to Efficiently Implement Chatbots

Chat is becoming more and more relevant as an interaction channel between customers and businesses. Recent advances in artificial intelligence enable the development of efficient chatbots that understand natural language. The financial services industry is becoming increasingly aware of the emerging opportunities.

Author: Weili Gao | Karim Attia

Chatbots are computer programs that can chat with people automatically, e.g. via Facebook Messenger. This means that customers can interact with a company via chatbots just as they would with a human customer service representative. Chatbots have debuted in 2016 for business use, and we expect them to become the primary way customers engage with companies within the next few years.

Benefits of chatbots

Chatbots offer many benefits that enable financial services providers to fundamentally reshape and improve their customer interaction. Chatbots are much easier to use than customer portals or mobile apps as the customer doesn’t need to search for the right webpage or go through complex menus to find relevant information or to place a request. Instead, customers can just start chatting with their bank or insurer.  It will automatically navigate them through the correct process – just as if they were talking to an insurance broker, financial advisor, or customer service employee, but with 24/7 availability and no waiting time.

Furthermore, chatbots can serve as a single point of contact. This means that customers are able to access all the available services of a financial service provider as well as various third-party services via a single messaging app. Chatbots allow banks and insurers to automate most of their customer interaction, resulting in a significant reduction of labor costs and processing time for customer requests. Additionally, a chatbot is easy to scale as it can communicate with thousands of customers on different messaging platforms at the same time. Finally, it is possible to program efficient chatbots with relatively low cost and time investment due to the large number of available chatbot development platforms. By using pre-trained artificial intelligence (AI) models from providers such as Microsoft, IBM, Google, Amazon or Facebook as a basis, it is possible to reduce the amount of effort required to train a chatbot significantly.

How modern chatbots work

As soon as a chatbot receives a message from a user, it has to identify the intent of the user’s request. It then guides the user through the relevant business process by providing him with information and determining required parameters by using so-called entities.

Thus, when developing a chatbot, it is necessary to first specify the intents and entities that the underlying AI should be able to identify when receiving a request from a client. With the help of a few sample user inputs and the machine learning approach, the chatbot is then able to link previously unknown user input sentences to the correct intents and entities. Comparable to a child who naturally makes many mistakes when it first starts to learn how to speak, a chatbot initially makes many mistakes when processing natural language. We can improve the accuracy of intent and entity identification by training the chatbot with additional annotated user inputs. This training effort can be reduced significantly by using a pre-trained «natural language processing model» (NLP model) from AI-as-a-service providers.

How to avoid pitfalls when introducing chatbots

Financial services providers are still hesitant about investing too heavily into chatbot technology. Before committing more resources, they first want to assess the reaction of their customers. They typically start experimenting with chatbots by initially automating only a few selected processes involving customer interaction. Accordingly, the customers of the financial services providers need to be informed that they will only be able to access a very limited number of services when using the chatbot. The answer to any customer request that exceeds the abilities of the chatbot is a response such as «Sorry, I could not understand you». Thus, there are two problems arising from this proceeding:

  • Chatbots with limited scope will only gain little acceptance among customers due to the poor user experience. Therefore customers must constantly make the mental effort to consider how to phrase their messages in a way that the chatbot will understand them.
  • These chatbots are deployed to work without any human supervision. Hence, they need to be 100% accurate in processing user input to avoid errors in the business process. However, attaining 100% accuracy often requires a lot of training for the chatbot, and in some cases it might even be impossible to achieve that goal with the current state of technology.

Gradual automation

We therefore suggest a different approach: Human agents work closely together with chatbots to achieve an optimal customer experience ( fig. 1):

  1. First, introduce chat (without chatbot) as a new interaction channel through which customers can access the full array of available services. At this stage, incoming requests are still processed by human agents, typically in traditional customer service centers.
  2. Next, a basic chatbot is deployed into the chat channel established in the previous step. This chatbot will only be able to handle simple requests initially. Whenever it receives a request it cannot process, it automatically hands over control of the chat to a human agent who will reply to the customer.
  3. Use the incoming customer requests and the corresponding responses and decisions of the human agents as training data to improve the chatbot continuously. With time and training, the chatbot can handle more and more requests autonomously, i.e. without human intervention.
  4. Expand the chatbot with further capabilities, such as voice-based interaction, or modify responses based on customer analytics and sentiment analysis.
graphic graphic
Fig. 1: Gradual automation of customer interactions via chatbots

Handover between chatbots and human agents

An important requirement of the gradual automation approach is the smooth transition between chatbots and human agents. Unless general artificial intelligence steps out of the realm of science fiction, there will always be times where the chatbot needs to reach out to a human being.

Whenever the chatbot is not sufficiently confident about the processing of user input, it sends its proposed response to a human agent ( fig. 2). The agent then decides whether to directly forward the chatbot’s response to the customer, or to modify the response first before sending a reply. The agent’s decision is then used to train the chatbot for similar future requests.

graphic graphic
Fig. 2: Dynamic handover from the chatbor to a human agent

Virtual assistants

In 2015, Facebook launched a virtual assistant called «M» on their «Messenger Platform» that adopts a chatbot-human hybrid approach to process user requests, just like the concept we suggested above. In contrast to completely AI-based virtual assistants like «Siri» or «Alexa», the responses generated by M are always sent to human agents initially. These agents then decide whether to directly forward M’s responses to the users or whether to intervene before replying.

For example, users can make a reservation at a restaurant via M. The human agents simply call the restaurant for the customers. The users of M do not know if they are currently chatting with a chatbot or a human agent. Facebook records all interactions between the customers and M and uses this data to further improve the NLP capabilities of M’s AI.

Current virtual assistants like Siri or Alexa cannot communicate with restaurants or insurance companies to make a reservation or to purchase an insurance product on behalf of customers. Imagine they could do this by forwarding requests to the businesses in natural language. Which restaurant would the virtual assistant select to make a dinner reservation? Which insurance company would it choose to purchase travel insurance? Certainly, not the one with which it cannot communicate.

Conclusion and outlook

Chatbots are not just a current hype but will become an important part of future business models. They enable businesses to offer their customers a new type of interaction to improve sales performance and to reduce costs. Customers will become used to communicating with companies via messaging apps, virtual assistants, and smart home devices. We expect that companies will even start designing products specifically tailored to chatbots.

Despite their many benefits, chatbots still require a well-thought-out strategic initiative so that they can be successfully deployed in customer interaction. Devising such an initiative requires in-depth knowledge of chatbots as well as detailed industry-specific know-how.

Contact

graphic

Silvan Stüssi

Cookies help us deliver our services. By using our services, you agree to our use of cookies. Find out more.
OK