By 2021, more than 50 percent of enterprises will be spending more per annum on bots and chatbot creations than traditional mobile app developments.
Gartner said in the report of Top Predictions for IT Organizations and Users in 2018 and Beyond.
Gartner’s stats are a clear indication where the industry-shifting is if you had any doubts over the future of chatbots. It’s now been more than two years we are providing chatbot development services we have built many chatbots across various domains and use-cases.
In this blog post, we will discuss our experience building a chatbot for Tax Filing solutions. Our learning and experience can act as a catalyst in building a great chatbot for your enterprise and financial services.
We aim just to do one thing very well; help our clients answer customer questions automatically.
Our client deals into full-fledged financial services and wanted to simplify tax filing for the end-users.
How We Started
We started working on a project with the goal of building something like KPMG LLP for tax filing guidance for an individual as well as multinational companies. We thought to build a database of well-researched answers and advice so that anyone could ask a question to tax filing chatbot like the difference between an exemption, credit and deduction, benefits of filing early, when to hire an accountant, etc. and get an immediate response.
We thought to shape a chatbot that could perform a role of “Tax Adviser”; guiding and educating taxpayers on tax laws and best practices and a trustworthy place for “income tax calculation”; that help taxpayers calculate income tax online.
There are many different chatbot platforms available such as DialogFlow from Google, IBM Watson Assistant, Azure Bot Service from Microsoft and others. We choose to work on Dialog Flow as it has machine learning backed training that checks user requests, help to identify the need, and match that with intent. Therefore, the platform assists in error identification and change approval in real-time, making chatbot development fast, efficient, and easy.
After spending a considerable amount of time, we came up with a pilot program for tax filing chatbot. We launched a pilot that allowed end-users to start communicating with the chatbot to get the appropriate solution. From this moment, we get the real sense of heat that comes to dealing with customers while helping them in tax filing and other related services.
Lesson 1 – Identify Key Performance Indicators
We were supposed to build a chatbot that will help and support users online. As chatbot is used at the first line of contact for users, it has to answer any questions users have. Our approach to creating a database of well-researched answers and advice for any questions fallback badly.
Shocking? It was for us too.
The end-users expect a chatty human-like tax filing chatbot that could answer queries profoundly to their particular needs. What did we deliver to them? Another good search engine for financial services and solutions.
How Did We Improve?
Our approach to dump content had failed us.
AI and ML-based chatbots can deprive relevant answers from the database. What is required is that we provide our chatbot a capability based on its KPI. As the first line of contact, a chatbot is required to support and promote “Self-Service.” In self-service, a user tries to solve the issues by themselves by following the guided solutions.
Here, the tax filing chatbot needs to be engaging with users. Therefore, we shifted from UX to Conversational UX. Conversational UX design allows the bot to ask open-ended questions and engage in ways that can make customers comfortable opening up about their issues.
Chatbot : Hey! Welcome To XYZ.
Last Date of Filing Tax is 31/01/2019
You must have filed it.
User : [Yes ] [No]
If [Yes} = Great !
Chatbot : Would You Like To Check/Verify The Status Of Return?
If [No] = How May I Help You?
Chatbot : What Are Tax Filing Process? Or Fill My Tax Return Form
Here a chatbot not just guides users through the process, but also proactively propose the next step to be taken to complete a task.
Outcome – The bot not just increased the interaction rate, but enhanced the experience of users who engaged with it.
Lesson 2 – Be Real For Chatbot Conversations
You already know what you want to serve to your end-users and fortunately, your data related to knowledge management to train your chatbot to respond to predictable questions. Still, why have chatbots not seen such great success till date?
OnGraph Team did a tremendous job to deliver true value to end-users via chatbot conversations. Still, our developed chatterbot missed the mark and conversation ended up as terrible experience for the end-user.
After some analysis, we know why
There was a lack of solid fallback experience in a chatbot to come up with a reply when the user responds in an unexpected way. Every time, the bot failed in predicting what a user wants to know and how will they ask it.
Though we try to teach a chatbot to use every system and service, every digital asset a finance company uses to help chatbot answer people’s questions. But our bot was not prepared to handle the unexpected conversation.
The tax filing chatbot lacks the capability to differentiate the intent from utterances; thus was not able to provide the expected answers.
How Did We Improve?
We focused on the three most important components of a chatbot: intents, utterances, and entities that make it contextual and behavioral smart.
Image Source: Tangowork
Intent, Utterances, and Entity are the core components of a human speech that create meaningful communication.
Content 1: A user may ask a question in simple words like:
What income do I have to pay taxes on?
Context 2: A user may ask a question as he thinks, like for instance:
I pay federal tax and I also paid state income tax for some. Perhaps I paid tax to two different bodies. Now, what taxable income should I consider to pay taxes on?
For a chatbot, answering the question of context 1 was easy with the database provided. To answer the question of context 2, we prepared our bot to respond to the unexpected as precisely as possible.
Note: We also integrated programs to respond in situations when a chatbot couldn’t interpret intent from utterances and suggest to transfer request for human assistance.
Chatbot : Sorry, I don’t understand what you are trying to say. But, not to worry! Our human experts are going to help you. 🙂
Human Assistant : Hey Max,
Alexander this side! Enrolled Agent and Tax Preparation Expert
The tax filing service you requested is actually paid. Would you like to opt it?
Customer : Am I really talking to a human assistant?
Human Assistant : I know we have an automated chat system, but rest assured I am a real person!
I just want to check if you saw my previous question?
Customer : 🙂 Fine! I would like to opt for the paid service. Please let me know how to proceed.
Outcome: Today, people are smart and you can’t fool them. Letting a user know about the current status does help a business win a customer’s trust.
Lesson 3 – Have A Mechanism In Place To Understand User’s Sentiments
A thought like “I build an awesome” may go out quickly when the bot begins to converse with people. Different people have different expectations and not every time a chatbot serves them well. But there is always scope for improvement. The best way to improve on the system is to modify it as per the user’s feedback.
We had launched our bot with mechanisms in a conversation where users can indicate that an answer was helpful or unhelpful. This was our way to collect feedback and improve the system.
On contrary to this, we noticed that users rarely bothered to share feedback even in most simplified ‘Yes’ or ‘No’ format. If they did, it was generally negative.
How We Further Improved the ChatBot Experience?
Though we have not achieved our true objective, we are improving by incorporating other advanced technologies that were available like ‘sentiment analysis’. It will help us identify a user’s sentiment towards the chatbot based on the conversations they will have. It wouldn’t require users to opt-in to share feedback.
The new mechanism will also look for terms like users typing response, vulgar language, overuse of exclamation marks and all those phrases that indicate satisfaction, happiness like ‘that was helpful,’ ‘thanks,’ ‘you are awesome,’ etc. It is hard to measure everything and it will take us time to come up with a completely functional system that helps us improve the mechanism as well as user experience with the chatbot.
Chatbots are not as difficult to build as people think, but the hard part actually starts just after that. You have to train and test the bot till the point it would start serving your purpose thoroughly.
The use of tax filing chatbot is still in an experimental phase at the moment. But technologies are improving day by day. Try new things like NLP, work on classifier accuracy, make it respond faster, add more logic to handle more use cases, but remember you can improve a whole lot of things in your chatbot.