3.

Building a Chatbot Solution

Business Considerations: Technological advancements may allow us to implement advanced
chatbots and other solutions, but how they add value to the business is a very subjective call. The
business needs to evaluate all factors to ascertain how a chatbot is good for their business.
 Chatbots vs. Apps: The key considerations of chatbots vs. apps are
• Chatbots are preferred to get quick answers for questions and 24-hour access.
• Apps are preferred for ease of use and convenience.
A survey in the report lists some factors that a business must check with their current needs:
• Quick answers to simple questions
• Getting 24-hour service
• Convenience
• Quick answers to complex questions
• Ease of communication
• Ability to easily register a complaint
• Getting detailed/expert answers
• A good customer experience
• Friendliness and approachability
• Having a complaint resolved quickly
 Growth of Messenger Applications: Another factor driving the need for chatbots is the
increasing usage of messenger applications and a stable growth rate of having a mobile-focused
approach by companies. The customer is now connected 24 hours to the Internet through their
mobile handsets and wants to access services via easy-to-use interfaces.
In early 2011, messenger applications started coming up and had good adoption rates as
mobile devices, internet connectivity, and cloud computing also picked up at same time.
Sometime in early 2015 we got to the point where messenger applications, specifically WhatsApp
and Facebook Messenger, were at meteoric adoption rates and equalled the activity on social
networks. As the trend suggests, people are more active on messenger applications compared to
social media networks.
This trend points out the shift in consumer behaviors where chat is a preferred mode of
communication. And this implies that businesses need to enable this channel of communication
with customers as well using either chatbots or human chats.
 Direct Contact vs. Chat: The increasing use of messenger apps has shifted the way customers
want to interact with businesses. In the early days of messenger apps, the studies showed an
increasing preference for contact via chats. The consumers believed they could solve their issues
faster over chat than calling a customer care representative. They also felt more comfortable with
chats because they could keep them as a record for follow-ups, unlike phone conversations.
Business Benefits of Chatbots: Considering essential business aspects allows a company to
decide whether it wants to go ahead and build a chatbot or improve current apps/ channels. If the
company decides to go ahead with a chatbot solution, it needs to understand the key value creation by
chatbots. Chatbots are now accepted by users with different percentages in different industries, with
different adoption rates.
 Cost Savings:The cost savings makes the best case for bringing chatbots into the service
delivery channel of company. While there are lot of other strategic and growth benefits, the cost
needs to justify all the efforts and resources used for chatbot development and maintenance.
The highest expenditure area for salaries is the customer service representatives,
where the savings are also highest. The cost reduction is clearly visible and attributable to the use
of chatbots across the functions in insurance sales, reporting, sales, and customer service.
  Customer Experience: The customer experience brings a multitude of values to the business,
not just limited to direct sales or savings. The benefits of good customer experience include
• High brand value and recall
• High lifetime values (improved engagement)
• Brand differentiator from competitors
There are many other derived factors due to a happy customer experience. A loyal customer base
is a recipe for long-term success.
The Chatbots Magazine summary points out features that contribute to a good and unique
customer experience derived from chatbots. The key points are listed below for reference:
• 24-hour service
• Getting an instant response
• Answers to simple questions
• Easy communication
• Complaints resolved quickly
• A good customer experience
• Detailed/expert answers
• Answers to complex questions
• Friendliness and approachability
The variety of new features attracts customers and creates unique value for companies.
Success Metrics: Success metrics are important to define at the start of any chatbot
development. The metrics work as a compass to direct the solution and the intended benefits of the
chatbot. The success metrics need to be manageable and measurable with a simple explanation to the
business. We will discuss a few metrics that can be used to track and manage the success of chatbots.
The metrics focus on success when you compare a chatbot interaction with a human interaction.
 Customer Satisfaction Index: The Customer Satisfaction Index (CSI) measures a customer
service representative’s quality of interaction by following up with the customer with a small
survey and capturing their experience of the interaction. CSI is one of the most impactful metrics
to monitor because it provides not only the satisfaction scale but also the areas of improvement.
 Completion Rate: The Completion Rate (CR) is defined as the proportion of interactions
with a chatbot that ended as the solution resolved for the customer. This metric tells us how many
times the chatbot can complete a conversation and deliver the required responses to the user. A
higher completion rate indicates a more efficient chatbot service.
 Bounce Rate: Bounce rate (BR) can be defined as how many users move away from chat
after typing in one or two inputs to the chatbot. A high bounce rate means the chatbot is not
successful in engaging the user and this must reflect in some of the customer feedback.
Along with bounce rate, we also measure the reuse rate (RR), which refers to how many
customers come back and use the chatbot again. BR is a perfect metric to identify those people
who more tend to use chatbots and target similar customer segments.
Managing Risks in Chatbots Service:The risk is to be understood, communicated, and
mitigated before general customers are allowed to use the chatbot services.
1. Third-Party Channels: Banks and other financial institutions are very much aware about
opening a new channel for users to access financial services. While it adds convenience to the
customers, it brings some risks as well. Technology is growing way faster than the risk
frameworks we have. By the time a general user is able to figure out the risk with usage or best
practices for using chatbots, they may already face a security breach. For example, a customer
interacting through Facebook Messenger is interacting with them bank systems using the
Facebook platform, which may have vulnerabilities and is not designed for banking operations,
just generic chats among people.
2. Impersonation: Impersonation can result in similar looking chatbots, or humans having
conversations using fake windows, hacking social media, and other sources of impersonation.
 The banks already face a lot of fraud due to criminal ingenuity from fraudsters and spend millions
in education for phishing, vishing, and other impersonation attacks. Two-factor authentications
are one possible way to reduce the impersonation attacks by having two-step verification from
two different sources. In most cases, hackers are not able to crack both factors of authentication
and are less likely to be successful in fraudulent transactions.
3. Personal Information: It is challenging to control users who may accidentally enter their
personal information to get access to a service. As the chatbots are driven by natural language, the
chances of revealing personal details are high. The chatbots need to make sure that they use as
little information as possible. It’s better to create a PIN with adequate access control so that the
user never needs to disclose personal information; they can just use the PIN. Educating the users
is an essential step to make sure the user is aware and alert for any fraud.
4. Confirmation Check: Confirmation is the most impactful and sometimes last resort to make sure
transactions done via chatbots are legitimate. Fraudulent or mistaken transactions are possible
using chatbots. As new technology comes into user service, it takes time for the users to
understand the right use of the service, and in this process, they might do some illegal
transactions as well. For any transaction that seems to be an anomaly or unexpected, it’s always
good to call up the customer and ask for confirmation before processing it. This check saves the
user and bank from fraud.

Generic Solution Architecture for Private Chatbots: Solution architecture is a practice of

defining and describing an architecture of a system delivered in context of a specific solution and
as such it may encompass a description of an entire system or only its specific parts. Definition of
a solution architecture is typically led by a solution architect.

Fig: Solution architecture of the 24x7 Insurance Agent

The following subsections expand the solution architecture with other vital information to explain it.
Workflow Description
Here is the workflow:
1) Conversation interface: We will develop the interface from scratch; it does not depend upon
third-party interfaces (e.g., Skype, Telegram, etc.). This will help us create customized interfaces for
the organization’s need and extend other features as required. The interface can be an entirely new mobile
app for the iOS, Android, and Windows platforms.
2) MFA and Active Directory: The authentication system will be built at the back end to authenticate
devices (MS Intune), users (Active Directory), and applications (by PIN). We will make it on a node
environment to allow integration to other identity management services as well. In the natural form,
we will only have PIN verification to access the application.
3) NLP engine: The NLP engine will be built to accept text inputs from a queue and extract intent and
apply context to the incoming query. Once the question is broken into the required components,
it’s sent to the bot logic.
4) Bot logic: This is the core handler of the incoming request. It will have two core inputs before
processing the request. The bot logic does not call information services until it has satisfied the process
set as per the below two methods. If there isn’t enough information for the bot logic to reply, it’ll
ask the user for more information.
a) Policy interaction flows: These flows are bank expertdesigned workflows for the incoming request
(e.g., if someone asks for an update of address, what are the essential steps for a reply?) The steps will
make sure the user complies with the steps to get an answer. This ensures that all policies, statutory or
internal, are followed by the bot, just like an informal HR. Also, policies and FAQ can be defined here.
b) Machine learning: The request that requires a machine learning algorithms to improvise the output is
requested from here (e.g., can I request a statement from March 15 to March 18?) This needs a machine
to apply appropriate logic to extract dates, an employee id, and an existing account balance to create the
right query to the information system. Further, sophisticated features like the mood of the employee,
the urgency of the request, and the sentiment of the question will be built here.
5) Information service: This is the place wherereal-time information is fetched for the employee
request. This service handles all appeals and prepares responses AND can also send a request for
actions as well. The information service will talk to three core data services:
a) Data lake/databases: If some data needs to be fetched from some database or data lake of corporate.
b) Third-party APIs/ODBCs: Interacts with HR systems via ODBC/APIs or some other method that
exposes itself with REST APIs.
c) Human HR: If there’s a low score of confidence in the reply, it will transfer the request to HR for a
response through the chat interface.
6) Actions: The information service will also route the requests asking for some work to the actions
item management queue (e.g., a request for leave for tomorrow). This request requires an update
in PeopleSoft or the HR system. All such change requests will be routed to an HR approval (can be
direct as well); once approved, either by real HR or by a policy-based rule, it will be sent to an action
queue. HR can support broadcast here, and they can be delivered by an actions workflow.
7) Updates in real systems: For authorized and approved requests, we update them directly in the
HR system and trigger a notification to the chatbot user and also trigger emails and other built-in
process flows of the banking system.
 Key Features: Below are the key features:
• Built for you: We will make the bot for specific needs, not fitting those needs to existing bot
frameworks.
• Data privacy by design: The bot is developed with data privacy by design. It will be fully compliant
with local laws and internal laws.
• Developed with a microservices architecture: The entire application will be based on principles of
microservices and hence will allow future-proof design and also advanced application development on
top of the framework.
• Options for deployment: We can choose which components we want to deploy on-premise or on
the public cloud. Based on needs, we can create a deployment plan.
• Extensible: We can integrate as many APIs or AI/ Ml features as deemed fit for use. All the new future
changes can be consumed as APIs in the framework.
 Technology Stack: Now the technology stack:
• Core engine: Java and JavaScript
• Backend server: NodeJS and other JavaScript
• Front-end server: Mobile apps based on native frameworks
• Log management: Cloud store of a small Hadoop cluster. Also, these stored conversational logs
provide data for AI/ML model training.
• Visualization: Can be custom created using D3; if self-service is required, then Tableau/PowerBI
integration with the logs.
• Search: Elastic Search to search the conversation logs.
 Maintenance:
There are two critical streams in terms of maintenance scope:
• The technology: Application uptime will be
maintained with the help of on-premise engineers and
an on-call channel for all queries.
• The AI/ML brain: This will be done offsite by a team
of data scientists and updates will be pushed to the
systems when ready.