How a 100k+ Employee Company Launched an internal Workplace by Facebook Chatbot in 1 Week
Chatbots have been all the rage for retail and customer support for a while now, but a new trend is emerging where companies, especially larger ones, are using chatbots to help their own employees as well.
A chatbot is a computer program, a robot, designed to interact with human users, usually in the form of a conversation — hence the name, “chat-bot”. With a chatbot, you can easily assist employees in their everyday job by automating repetitive, low added-value tasks. Things like assisting with computer issues, giving guidance on new processes, helping new hires find their way in the organization or answering frequent questions about health insurance and vacations — these are challenges that Clevy is helping customers solve most often.
However, building a chatbot from scratch can be a bit scary. Chatbot platforms and technologies require some very specific skills (development, conversational user experience). This means that customer-facing chatbot projects usually take many months to launch and cost up to several hundred thousand dollars. So while it is perfectly understandable to spend a lot of time and money on projects that generate revenue, it can be more complicated to justify these big numbers for internal projects.
Clevy is a platform that helps large companies very easily build, launch and monitor internal chatbots that will answer the questions their employees ask most frequently. It requires no technical skills and no custom development. Anyone can create and manage a chatbot on Clevy very quickly.
In this article, we will highlight how one of the largest food companies in the world, Danone, successfully launched a chatbot to its employees worldwide on Workplace by Facebook in one week and what it helped them achieve.
1. Design — Defining a scope
The very first stage of designing a good chatbot is to define a very clear scope: who it is for and what specific problem the chatbot is going to solve. A company-wide chatbot about pay and benefits can not work in a global organization, as the rules are not the same in each country. A chatbot that would be able to answer questions about health insurance and IT support is also confusing: users are accustomed to talking to different teams for asking these questions and don’t understand what they can or can not ask to the chatbot.
A chatbot must answer questions within a very well-defined perimeter to work well. The chatbot that works for everything and anything is a sci-fi myth.
How Danone did it:
Danone was looking for a solution that would promote internal understanding and engagement with Danone Communities. The fund was created by Nobel laureate Pr. M. Yunus and Honorary Chairman of Danone Franck Riboud, and aims to alleviate malnutrition, give access to safe drinking water, and break the cycle of poverty in the countries in which it operates, by investing in social businesses as minority shareholders, providing capital, and technical and managerial expertise.
The scope was simple and well defined from the beginning. By answering questions about Danone Communities, the chatbot should inform Danoners about how they can contribute, either by giving time (from a few hours to a few months) or money (from their profit-sharing) to the businesses supported by the fund.
2. Training the bot
The most crucial part of building a chatbot is the training. Chatbots are not magical and we are still a long way from what you can see in SF movies like Her or Ex Machina. For example, training a bot to understand “I’m thirsty” doesn’t mean that it will also understand “let’s go for a drink”.
The initial dataset should at least include a few question reformulations with:
- Common figures of speech
- Synonyms
- Acronyms and business language
Your bot must also be able to handle common small talk (“hello”, “who are you”, “thanks”…), which Clevy bots handle natively without any additional configuration.
When training a bot, beware of a phenomenon we call “Trainer’s Bias”. Coaches (bot managers) are generally experts in the topic, so they tend to train the bot with the questions they would like users to ask them (or to which they would like to provide an answer), but those are not necessarily the actual questions real people ask.
The only way to counter this bias is to get real users to use the bot, so that they can ask questions that the bot’s coaches didn’t think to include in the database at first.
A chatbot can not be considered ready to launch until it has been properly user-tested and its knowledge base augmented with real questions asked by real and representative end-users.
How Danone did it:
Training a bot is not an easy task. It takes time and commitment from the beta testers, so being prepared is key — especially in a large organization such as Danone.
After setting up a first version of the knowledge base in Clevy and augmenting it with our managed Smalltalk module, we ran a 1-hour training session with a select group of people. Everybody was given the task to ask the bot as many questions as possible on the topic, while the trainer could enhance the database in real-time.
The fact that the bot was about a social business helped finding beta-testers. We usually recommend having these (absolutely necessary) sessions over lunch or during a longer afternoon break, with some food or candy and (soft) drinks provided for everybody. Also, depending on the size of the bot, it may be necessary to organize two or three sessions with different groups of people.
When you are satisfied with the KPIs, your bot is ready for launch!
3. Deployment
There is no single best way to launch a bot, but there are a few things you want to communicate to your users:
- Where can I find the bot?
- What can I ask?
- If something goes wrong, what can I expect to happen?
- Examples of some best practices using the bot
There is such a thing as good and bad questions when it comes to bots, as people are used to either explaining their needs in a much too detailed way (“I was at the beach the other day with my daughter and my cell phone fell from my pocket into the water, now it’s wet and doesn’t work any more. Who can fix it?”: too much irrelevant information!) or use only keywords like in Google (when you just say “vacation”: there is no way to know if you want to take a vacation, find out how many vacation days you have left, look up the company policy for sabbaticals, etc.). You need to tell your employees to use simple, direct sentences: “my iPhone is broken, how do I fix it?” or “I want to take a vacation”. This calls for at least a simple communication plan, and if you can, embed a few examples directly inside the bot channel.
Speaking of which, you also need your users to find the bot. If possible, use an existing platform, something they are already familiar with. For example, rely on your corporate social network, or an existing chat platform.
Build a plan to handle the questions you don’t yet have an answer to. Clevy enables you to manage these situations asynchronously — when the bot doesn’t know how to answer a question, it will store it in Clevy’s back-office inbox, and your bot coach will answer it later and add it to the bot’s knowledge base. It’s an ongoing process. Your specialized bot won’t know everything straight away. You can explain what a user can do via the “I don’t know” message.
A chatbot only works if people know where to find it and how to engage with it. The more intuitive the better: use the tools they already use like an existing intranet or your corporate social network, engage with users in the most natural way for them, and make sure they understand what they can or cannot do before they start using the bot. This will prevent disappointment.
How Danone did it:
Danone is one of the first and largest customers of Workplace. Workplace is a collaboration and communication tool that helps people work faster and smarter together. Workplace’s open API and integrations with the business tools that organizations already use make it easy to launch custom bots.
Danone has already received a few awards from Facebook for their launch strategy of Workplace, so they are well-versed into deploying tools to their whole workforce. Here are some highlights of the deployment of this chatbot to the whole workforce on Workplace:
- No additional development costs: linking Clevy with Workplace takes just a few clicks.
- The bot was launched at the 2018 corporate wishes. The bot even had a booth at the event!
- It answers questions instantaneously on Workchat, Workplace’s equivalent to Facebook’s Messenger which is the standard chat solution at Danone, where they can find the bot the same way as finding any human colleague.
- A dedicated Workplace group has been created to announce the bot, and engage Danoners prior to and during the launch event.
- Danoners could directly contribute to the project by voting in a Workplace poll for the name of the bot, which helped raise awareness about the bot.
- Workplace features a convenient “Featured Bots” section at the top right corner of the famous feed. Danone activated that option especially for this bot, which has made it extremely simple for any Danoner to find and use this bot.
In the first month after the initial launch, with about 50 items in its knowledge base, the bot was able to answer about 750 questions directly in Workchat, from many countries and across timezones.
Danone Communities is now preparing a second launch, with an extended knowledge base to communicate about their impact3 Program: impact3 gives Danoners the opportunity to contribute with their time and skills directly to Danone Communities and the social businesses. Keeping things moving and interesting is surely a good way to stay on top of users’ minds!
4. Monitoring — continuous improvement
Handling unanswered questions is very important to drive better bot performance. You’ll find that the type of questions people ask evolves over time. Add missing questions to your knowledge base so that users continue to use the bot and the bot continues to be relevant and useful.
Monitoring your chatbot’s performance will also help you find out the hot topics and questions. It will also help you quantify how much time you are saving by not having to answer all these questions by email.
Once the chatbot is live, keep an eye on it. Questions will change over time, depending on many factors (time of the year, current news…). For your bot to remain effective, make sure it is always up-to-date with your actual users needs.
How Danone did it:
Danone is disciplined when managing the bot inbox. This means the performance of Danone’s bot continues to improve over time. Most questions get an answer within a few hours so the bot stays up to date with the needs of the community.
Also, using Workplace’s native APIs, it is easy to notify the user that their question was finally answered. This new notification usually drives the user back to the bot, where they often continue the conversation.
Finally, Danone continued to promote the bot internally. They arranged physical events with booths and screens that helped explain the project to Danoners and give them the opportunity to try it live.
Conclusion
There are four steps to building a successful chatbot: design — training — deployment — monitoring. The main KPI to measure is the engagement score. Do people use the bot? How often? A chatbot is also a powerful analytics tool allowing you to understand what are the most frequently asked questions, and identify the ways you can improve internal communications.
The success of a chatbot can be also linked with its ROI. In the case of Danone, answering these first 750 questions by email would take about 90 hours (assuming a rate of 7 minutes per email plus interruptions). That’s nearly two-thirds of a full-time employee’s time every month, which can be spent doing more added-value tasks.
One Danoner is able to manage this particular chatbot, maintain the knowledge base, and answer questions. Using Clevy on Workplace is helping reduce the general workload. And it helps to serve more people — more quickly — than ever before.
By harnessing the power of Workplace and Workchat, the team at Danone Communities has built the fastest bot we have ever seen at Clevy. We are used to seeing bots built and deployed within 5–6 weeks, sometimes even 3 if everything goes well: it took Danone only 6 days from zero to launch!
After this first experience with an internal communications use case, Danone has further plans to release at least 2 other chatbots using Clevy and Workplace this year. They will focus on broader topics such as IT support and change management. The scope is once again well-defined and the deployment channel is already embedded in the work-life of employees, so there is no doubt the launch will once again be a success!
Find out more about how Workplace can help transform your organization. And see how Clevy builds and deploys chatbots that can revolutionize the way you work.
