A chatbot built from the support queue up

Advisors waited on hold for answers to routine questions, while the Customer Support Unit (CSU) worked through a growing pile of calls and emails. With the product manager, I analysed 4 years of call, chat and email data and read 1,000+ customer interactions to find the patterns: the most common topics and questions, and how requests were escalated to a person. Those patterns also set Ely's tone and style. Then I designed Ely, a chatbot in English and French built in Dialogflow (Google's platform for building chatbots). Ely handles the routine questions and hands the hard ones to a person.

Problem

Routine questions tied up phone lines and the support team.

Deliverable

Ely's persona, chat window and 500+ conversation paths in English and French, plus a support-trends dashboard.

Tools & timeline

Dialogflow, Google Data Studio, Jira

May to Aug 2020 · summer internship

Outcome

• −30%customer service requests
• +140%chat usage
• +340%successful chats

All year over year, compared with 2019. A successful chat is one the customer rated helpful.

From data to dialogue

Read the receipts

• Pool the data

  4 years of advisor calls, portal chats and emails, from 2017 to 2020.

• Read the conversations

  1,000+ interactions, sorted by topic, by frustration, and by when a person had to step in.

• Keep it live

  A support-trends dashboard in Google Data Studio, so the team could see what was rising.

The biggest groups, finding forms and documents and how-to questions, became Ely's job. Anything that needed an ID check or a long answer stayed with people.

Persona

Meet Ely

Name

Ely the Empirebot

Role

Automated customer service support

Age

Legal working age

Gender

Undefined

• Enthusiastic
• Empathetic
• Professional
• Helpful
• Informative
• Extravert
• Eloquent
• Diligent
• Adaptive
• Functional
• Accessible

Voice, from the real screens

Hi, my name is Ely! I'm the Empire Life chatbot.

I am trained in making your search for forms and products easier.

Discovery

What the data said, and what it changed

• Most requests were about finding forms and documents.

  What the data said

  What it changedAn opening line that says so: "I am trained in making your search for forms and products easier."

• Some questions needed an ID check, a long answer or a person.

  What the data said

  What it changedA Contact Us department and live agent support.

Conversation map

Four departments, and a way back from anywhere

Testing

Three versions before launch

We tested three versions of Ely, first with internal testers and then with a pilot group. Each version changed three things.

• Decision tree

  How the conversation branches, and how many steps it takes to reach a form, an answer or a person.

• Tone

  How Ely sounds: the same answers in different voices, checked against the persona.

• Number of options

  How many choices Ely offers at each step.

A version won on two things: what testers told us, and the share of chats rated helpful.

Handoff

What designing Ely taught me

• Start from the queue, not the wish list

  • The data told us what Ely should do and, just as clearly, what to leave to people.

• Personality matters as much as the answers

  • A friendly, steady voice makes a focused bot feel helpful rather than limited.

• Two languages means designing twice

  • Some lines need a different tone in French, not a straight translation, and French-speaking customers sometimes prefer support to be handled differently. Every path had to read naturally in both.