Meet Charlie, Belgium’s most helpful customer support chatbot

Meet Charlie, Belgium’s most helpful customer support chatbot

Argenta logo
Charlie is Argenta’s virtual assistant. She lives within the Argenta mobile banking app and helps answer customers’ questions and complete transactions. But what makes her so helpful? She only chats when she’s confident she can answer the question.
The problem

How can we make a chatbot that improves customer experience?

Argenta is the fifth-largest bank in Belgium and its customer support team receives 20,000+ messages per month. It was impossible for their team of 23 agents to keep up. They were overloaded and overworked.

During the discovery phase of the project, we identified three areas where the agents were losing the most time:
    1.Basic FAQs
    2.Basic intent detection and data capture for specific transactional flows
    3.Conditional data detection and capture for increasing daily limits

Argenta is known for its excellent customer service and wanted to ensure that customer satisfaction remained stable or improved when introducing the bot to its customers. With this specific challenge in mind, we introduced a few custom solutions that take Charlie from being a basic bot to a great bot.
The solution

Charlie is a chatbot with a clear scope and near-perfect NLP model

Charlie only chats when she’s confident she can help

Latching onto the concept that Charlie should maintain the status quo or improve customer satisfaction, we pivoted and approached her introduction differently: She’s a very respectable lady and only introduces herself when she’s at least 85% sure she has the right answer.

 Conversation where Charlie is 97% sure of the right answer

Conversation where Charlie is 62% sure of the right answer and immediately hands it off to a live agent to support
How is this possible? We developed a unique introduction and fallback system so that if she knows the answer, she shares it. If she doesn’t, the conversation gets routed to a live agent, and the customer has the same experience they would normally have pre-Charlie.

The only change for the customer is that if their question can be automated, they receive an instant answer instead of queueing for a live agent. It’s a win-win for the customer and the support team because both save time! The customer doesn’t waste time queueing for a live agent and the live agent doesn’t waste time answering a question that Charlie could easily automate.

Charlie has pre-defined areas of expertise

By not introducing herself unless she’s confident in the answer, Charlie doesn’t need to know everything. Together with the customer service team, we defined a clear scope of easy-to-answer, frequently asked questions. This translated into 30-40 intents covering approximately 15% of all incoming questions.

Beginning with a small set of intents enabled us to test, train and monitor Charlie’s natural language processing (NLP) model to near perfection. When we first introduced her to real customers, we reviewed every interaction to ensure that she said the right thing and created a net positive customer experience. After a few weeks of intense monitoring and some minor tweaks, it became clear that Charlie knew what she was talking about and was ready to work independently.

Charlie clarifies doubt instead of defaulting to “I don’t know”

The vocabulary associated with banking is complex and nuanced. Since many banking intents overlap with one another, we usually recommend working with entities, such as debit card, credit card, limit, etc. Using contextual entities, we built a unique entity fallback system. If Charlie’s between 70-85% sure of the intent, then she rephrases the intent as a question to confirm her suspicions. If she recognizes an entity, she responds with a custom menu that proposes the most asked intents related to that entity.

Conversation where Charlie is 74% a keyword (entity) in the user’s question

Conversation where Charlie is 75% sure of the intent and confirms her understanding with a question

Charlie automates data capture to save everyone time

Basic data capture is a time-sink for a live agent. It can take up to 10 min. for a user to respond with their name, address, email and phone number. That’s valuable time that an agent could use on more complex conversations.

One of customers’ most frequently requested items is a new card reader. The batteries expire over time and customers ask for a replacement. This question is easy for a bot to recognize, so instead of bothering a live agent with this simple request, Charlie detects when a user is asking for a new card reader and responds with a data capture flow that captures the user’s name, address, email, etc. Charlie then forwards that information and the new card reader request to a specific team within Argenta, where they check the request and prep the card reader for dispatch.

Converting a frequently requested manual process into a conversational flow saves Argenta time and provides users with an immediate solution

Another frequently asked question is how customers can increase their daily transactional limit. For certain amounts, it’s an easy fix that Charlie can automate. But for amounts that exceed Є25,000, the customer needs to answer some questions. Charlie can detect when a customer wants to increase their daily limit and can also detect if the amount exceeds Є25,000 so she can already get answers to the questions before handing the conversation off to a live agent to make the final limit adjustment.

Charlie helps a user adjust their limit to Є13,000

Charlie helps a user adjust their limit to Є30,000

The Results

Charlie is a chatbot loved by colleagues and customers

Charlie saved the customer support team 24 days (192 hours) of work in the first month and a half, and they’re grateful for her help. Independently, Charlie manages 20%+ of the incoming customer conversations end-to-end. 

Our new colleague Charlie is doing really well. Because of her quick reaction to simple questions, we can focus on the real work 😉

Ilke Schiltz, Customer Support Agent @ Argenta
The most triggered flow was increasing daily transactional limits, followed by requesting a new card and then a new card reader. As Charlie continues to engage with and help customers, we’ll broaden her banking knowledge so that she continues to save the customer support team time.

💾

24 DAYS

OF WORK SAVED IN THE FIRST 1,5 MONTHS

💪

+20%

CONVERSATIONS HANDLED FULLY BY CHARLIE

😊

+95%

CONTACT CENTER
CSAT

Other work that will spark your interest

Sabam & Unisono | Login Help Assistant

Sabam logo

A chatbot that helps you troubleshoot
your login issues

A lot of the issues that a user faces at the login screen can be solved using a handful of different troubleshooting tips. We took these tips and built a chatbot that helps solve your own login problem in less than 3 minutes
An iPhone user that is unable to login to their Sabam account
The problem

Lost at the login screen

Sabam is the Belgian association of authors, composers and publishers, and Unisono is their online platform where users can apply for licenses. With a membership base of more than 40.000, they wanted to reduce the volume of customer enquiries.

After reviewing the data, it stood out that most of the requests were about problems logging into the platform. To help members login to Sabam and Unisono faster, we developed a chatbot that guides users through some troubleshooting exercises and ultimately login to the platform.

The solution

Virtual support when you need it

Working closely with Sabam and Unisono, we identified the most common login errors (ex. CAPS LOCK turned on) and used them to build a conversational flow. As soon as a user fails to login, the bot pops up and asks the user if they need assistance logging in. The bot then guides the user through the different conversational flows to help them troubleshoot their login error. 

In addition to the troubleshooting flows, we integrated 30+ other frequently asked questions in French and Dutch that users can trigger via the chat.

Screenshots of the different login chatbots that Campfire A.I. built for Sabam and Unisono
The results

Faster logins and less time wasted

The login chatbots handle 3.000+ conversations a month and solve 97% of the login issues in under 3 minutes. Not only are users happier and able to login in faster, but the Sabam and Unisono customer service representatives have more time to focus on answering questions that require human follow-up.

💬

3,000+

CONVERSATIONS A MONTH

97%

OF LOGIN ISSUES SOLVED

> 3 min.

TO SOLVE AN ISSUE

Other work that will spark your interest

The natural evolution of mobile banking into conversational banking

The natural evolution of mobile banking into conversational banking

Rawbot answers Rawbank customers frequently asked questions
Rawbot is Rawbank’s virtual assistant that customers can chat with to complete financial transactions, as well as ask questions. But what makes Rawbot so popular? He’s easily accessible via the website, WhatsApp, Messenger, Instagram and Twitter.
The problem

How can we replicate mobile banking within a WhatsApp chatbot?

As the largest bank in the Democratic Republic of Congo, and with WhatsApp now the most-used messaging app in the country, Rawbank was looking for a way to bring traditional SMS banking to WhatsApp.* It’s important to note that the financial landscape in Africa is unique because it’s historically built on mobile banking. With very few physical banks and unreliable internet connections, SMS banking was and continues to be the norm for sending and receiving payments.

The project began purely focused on centralizing and automating questions coming from its +200,000 customers. However, Rawbank’s digital team quickly realized that they could also introduce financial transactions. The challenge then became how can we expand the MVP to:

    1. Enable secure financial transactions via WhatsApp
    2. Meet its customers where they are to answer their question

With these two criteria in mind, we worked together with Rawbank to develop an omnichannel, conversational banking strategy.
The solution

An omnichannel chatbot approach for mobile banking

Evolving from SMS to WhatsApp

When evolving from SMS to internet-based financial transactions, we first needed to overcome a few technical hurdles. Most importantly, there wasn’t an easy plug-and-play API to connect natural language processing (NLP) and bot-building software to the existing infrastructure. Instead, we built a customized microservice that acts as a translator between the newer cloud models and existing telco infrastructure.

One of the easier transitions from SMS to WhatsApp is that the mobile number remains the customer’s unique identifier. The other benefit is that by using WhatsApp there’s additional security and encryption built into the platform. It’s very difficult to hack WhatsApp and steal someone’s login– you’d basically have to steal someone’s phone.

For the MVP, we built two transactional flows. The first flow allows customers to check their balance in Congolese francs, euros and US dollars. The second flow allows customers to check their most recent transactions. The next version of Rawbot will include transactional flows that allow customers to send and receive money via WhatsApp.
Bank balance conversational flow

Introducing conversational AI to mobile banking

To start with conversational banking, we agreed on a scope of 50 intents that cover the most frequently asked questions from Rawbank customers. Since the vocabulary associated with banking is very specific and because many banking intents overlap with one another, we usually recommend working with entities, such as debit card, credit card, limit, etc. Using contextual entities instead of match entities, allows us to create a fallback system, so that instead of responding with “I don’t know,” Rawbot can clarify any doubts himself.

Intent and entity fallback system

If Rawbot’s between 70-85% sure of the intent, then he rephrases the intent as a question to confirm his suspicions. If he’s between 50-100% certain of the entity, he responds with a custom menu that proposes the most asked intents related to that entity. This approach enables Rawbot to resolve a majority of the incoming questions end-to-end so that only a small percentage of complex cases get passed on to Rawbank’s internal support team.

Tailoring the conversational design to match each channel

For Rawbot to be a success, it had to be easily accessible by Rawbank customers. The only way to achieve this was by adopting an omnichannel approach so that Rawbot was available on channels where Rawbank customers were already active.

Technically speaking, connecting a bot to different channels is quite simple. However, designing an omnichannel conversational strategy is much more difficult. It requires modifications and redesigns per channel so that the user has the best possible experience.

Every channel and platform is different. From the way the user interacts with the channel, to the context, interface, visualizations, technical restrictions, etc. For example, a chatbot begins the conversation by immediately introducing itself when the web chat is opened. However, on WhatsApp, the user must begin the conversation, which means that we had to foresee two different introduction flows.

Rawbot’s Whatsapp and Web introduction

Other differences are related to the channel interface and technical abilities. In a web chat, we can use carousels to visually highlight different options, but that feature isn’t available in WhatsApp, so we need to use a menu to present options.

Rawbot’s Whatsapp menu

Rawbot’s web menu carrousel

Or on Messenger, we can use buttons to help the user make quick choices, while on Twitter that feature doesn’t exist, so we need to use numbers.
The Results

Users are making the switch from SMS banking to WhatsApp banking

Rawbot chats with +6,500 customers a month. Half of those customers interact with him on WhatsApp, while the other half is divided among the website, Messenger, Instagram and Twitter.
These numbers continue to grow each month and we see that more and more customers are using WhatsApp to check their balances and most recent transactions. The transactional flows are triggered +800 times per month, and the next Rawbot release will include transactional payment flows where users can send and receive money through WhatsApp.

🔑

800 +

MONTHLY TRIGGERS FOR TRANSACTIONS

💬

6500 +

MONTHLY ACTIVE USERS

🖥️

4

ACTIVE ON DIFFERENT CHANNELS

Other work that will spark your interest

European Commission | Travel Chatbot

European Commission Logo

A chatbot can support travelers 24/7

Traveling is an activity that happens in different time zones and outside of the traditional office hours. However people sometimes need instant support. Artificial intelligence allows us to easily overcome the limitations of different time zones and traditional office hours by creating a chatbot that's available to support users 24 hours a day, 7 days a week, and 365 days a year.
Expand your comfort zone with Discover EU
The problem

Travelers need immediate support

Discover EU is an initiative led by the European Commission (E.C.) that helps 18-year-old E.U. citizens discover Europe by train. Many of these young Europeans are inexperienced, first-time travelers who find themselves in situations where they need immediate help (ex. I missed my train, what do I do now?). However, it’s costly and challenging to staff a large team of support agents outside traditional office hours and across different time zones.

To guarantee affordable, real-time support for young travelers, we worked with the E.C. to develop a chatbot that’s always available to answer young travelers’ questions.

The solution

A chatbot with a customized handover

We began the project by reviewing Discover EU’s database of user questions and identifying the 20 most frequently asked questions. This set of questions and answers became the foundation of the chatbot. 

 

We strengthened this chatbot implementation, by customizing the handover between the bot and a live agent. In cases when the bot can’t solve a user’s problem, we transition the user from the conversation with the bot to a conversation with a live support agent. During this transition, we use A.I. to tag the conversation with keywords and relevant information. These smart tags help the agents quickly understand and prioritize the conversations.

The results

Automated support that's always available

In the first month, the chatbot solved more 700+ questions and handed over approximately 150 questions to a live support agent. By automating 83% of the support requests, we solved travelers’ problems faster and gave the support agents the time and space to focus on more complicated requests. 

💬

700+

QUESTIONS SOLVED IN THE FIRST MONTH

🤖

83%

OF SUPPORT REQUESTS AUTOMATE

Other work that will spark your interest

European Parliament | Crisis Relief Chatbot

European Parliament logo

A smart chatbot can scale your crisis communication

Fast and accurate communication is critical during a crisis. But, it can be challenging to manage the mass influx of questions that accompany a crisis. We worked together with the European Parliament to build a smart chatbot that answers employee questions in real-time.
A screenshot of a Facebook post from the Crisis Center in Belgium about the Covid-19 crisis
The problem

Quickly communicating at scale

In a matter of hours, the Corona virus forced a majority of the European workforce to leave their offices and set-up a desk at home. Many companies lacked an official remote working policy, which only added to the confusion and stress caused by a global pandemic.

 

The European Parliament (E.P.) immediately recognized the importance of accurate communication during a crisis and briefed Campfire to build its first smart chatbot to help streamline and accelerate communication with its 11.000 employees stationed across Europe.

The solution

Real-time answers to questions

A.I. (artificial intelligence) is the best technology for answering a high volume of questions. By layering a basic chatbot with a robust A.I., we can ensure that the chatbot is “smart” and understands a wide variety of user questions. 

 

In less than a week, we published a first version of the chatbot the E.P.’s intranet. This bot could answer 50+ different questions about the Corona virus, the Corona crisis and the E.P.’s remote working policy. 

 

Because the pandemic evolved rapidly, we monitored the conversations daily. This allowed us to identify gaps in the bot’s knowledge database and improve the A.I. with continuous NLP (natural language processing) training. We also worked with health department experts to regularly update the bot with new information and regulations.

 

The results

Quick adoption and even quicker answers

In the first 6 months of the crisis, the bot answered 15.000 questions from E.P. employees. 

Using A.I. to build a smart chatbot meant that the HR Team and other internal experts could ensure that they answered employee questions in real-time and with accurate information. It also gave the E.P.’s internal team the flexibility to focus on more complicated cases that required individual follow-up. 

🔁

Weekly

CONTENT UPDATES

💬

15,000

QUESTIONS ANSWERED IN THE FIRST 6 MONTHS

Other work that will spark your interest

SIT | Virtual Tutor

A virtual tutor powered by Artificial Intelligence (A.I.)

Online learning offers students flexibility, but often lacks the guidance and support that's available in a traditional classroom. We worked together with university staff and used A.I. to develop a virtual tutor that guides students through their online course material.
THE PROBLEM

Remote studying can be a struggle

Every summer, the Singapore Institute of Technology (SIT) runs an online chemistry course for incoming first year students. The course is designed to put everyone on a level playing field. To help students learn chemistry online, we developed a virtual study assistant to complement the online chemistry course.

THE SOLUTION

A virtual tutor that's available 24/7

In collaboration with the professors and researchers from SIT, we reviewed the different course chapters and developed a hierarchical structure of hints to match the evaluation questions at the end of each chapter. The first hints are more generic, while the second and third hints are more specific to the question.

 

We also added the answers to 100+ frequently asked questions that cover more generic chemistry content, such as explaining Avogadro’s Constant, the formula for calculating moles, and the definition of enthalpy or entropy.

A screenshot of the flows used to build SIT's virtual tutor
THE result

Engaging students in meaningful conversation

During the user research phase, 57% of the students indicated that they found the chatbot useful in answering questions. They commented that the chatbot is convenient to use and that having a conversation with a bot is more engaging and interactive than simply googling a question.

 

56% of students also expressed that they would prefer an imperfect chatbot that’s available 24/7 than an instructor available for a limited period of time.

 

As part of a larger research project from SIT, we’ll continue optimizing the bot and their team of professors and researchers will continue investigating the impact of a human tutor vs. a virtual tutor on course completion rates. If you’re interested in learning more about SIT’s research, you can read the full research paper here.

✅

57%

OF STUDENTS FOUND THE CHATBOT USEFUL IN ANSWERING QUESTIONS

🤖

56%

OF STUDENTS PREFER A CHATBOT THAT’S AVAILABLE 24/7 TO AN INSTRUCTOR AVAILABLE FOR A LIMITED PERIOD OF TIME

 

 

Other work that will spark your interest