In this basic article you can read everything about chatbots: from definition to the importance of a clear objective and from implementation to pitfalls and results.

Read the full article or quickly jump to the topic of your preference via the links below:

What is a chatbot?
What do you use a chatbot for: what goal do you want to achieve?
Chatbots are good at one job. Which tasks do you choose?
What are the pitfalls when using chatbots?
What does an implementation process generally look like?
What costs should you take into account?
What is the payback period?

 

What is a chatbot?

A chatbot is an automated software application that can independently handle specific questions – usually via text messages, sometimes via speech. A chatbot consists of two parts: a part aimed at recognizing the question, a second part aimed at performing the question/answer game aimed at that specific question or intention. The question is often translated into the intention (the intended goal of the customer) or an outcome (the solution). Chatbots can contain different levels of intelligence. At its core, there are two flavors: script-driven bots and AI-powered bots (see below). Chatbots exist since the XNUMXs, but have undergone considerable development in recent years.

What is the difference between FAQ, Q&A, an IVR and a chatbot?

A Q&A is a pre-formulated question with an accompanying answer. Often questions and answers are on a web page. FAQs are based on 'frequently' or 'most frequently asked' questions.

The customer must look up the answer to the question in a list. Sometimes the customer can provide a search query, so that the most obvious question/answer combination emerges based on word recognition. An underlying database can also use synonyms for this.

The IVR does not provide answers, but is a routing mechanism in which, based on keywords (specified in advance and linked to figures: choose 1 for administrative questions, or by means of speech recognition), it is determined which employee (or which competence) is required to answer answer the customer.

Two types of chatbots

Chatbots can be roughly divided into two types. The most common variant is also the most basic and comes down to automating a form with questions, where the customer is supported in filling in answers. In this application, a simple chatbot runs through a predefined script with standard questions. This often also means that only the requested information is processed from the customer's answer. Word recognition can be used intelligently (based on probability and learning ability), but there is usually no intelligence in the response.

In all cases, a fixed process is completed in a fixed sequence. The difference with a normal web form is that the chatbot presents the questions one by one and determines the next question based on the answer. As a customer, you cannot deviate from the conversation.

The chance of frustration on the part of the customer is quite high, because the customer expects to conduct a dialogue, while the system primarily enforces a certain course of action and does nothing with additional information provided.

Recognition based on one word: “order”

Another question with answer recognition is necessary: ​​"what do you want to know about your order?" (warranty, return, delivery, error in delivery)

Recognition based on intention: “I received the wrong size and would like to exchange”.

Several separate counter questions are needed to clarify the problem: was there a mistake when preparing the order or does the customer want to exchange because the product does not fit?

 

The second variant chatbot is intelligent, because when processing input from the customer, it can process all the relevant information present in that response and uses machine learning to do so. The smart chatbot is trained with a dataset and gets better by dealing with new data. If a customer already indicates at the start “I have received the wrong size and would like to exchange, my customer number is 1234567”, for example, the chatbot does not have to ask questions about the customer data or about the customer's wish (exchange). The intention of the customer seems to be 'exchange' in this case, but it may also be that the customer wants to know the return conditions.

Artificial intelligence can be used in three ways: when offering Questions to the customer, when processing reply of the customer and the combination of both. In the second case, the chatbot uses more information than is available in its own systems and the bot adds knowledge about this to its own systems.

From dumb to smart(er) chatbots

Targeted chatbots for arranging narrowly defined products or matters (e.g. airline tickets or reservations) must be able to ignore irrelevant information. The advantage is that such a more or less closed system knows exactly what to look out for.

The disadvantage is that no natural dialogue is built up, because the system cannot deal with customer initiatives. A chatbot that helps you book a ticket may not be able to respond to a question about booking a rental car or a hotel.

Many organizations therefore consider it a safer solution to limit the application of AI to recognizing customer demand. In doing so, they opt for control over the answer that the chatbot produces.

As a chatbot solution knows more about things like properties, structures and rules, it is better able to conduct a normal dialogue. For example, an appointment has properties such as day, date, time, duration, location and participants. For example, times can be reduced to certain structures: if you ask for an appointment for a week from now, you need to know what today's date is.

What do you use a chatbot for: what goal do you want to achieve?

Organizations that want to use chatbot technology would do well to answer the 'why' question correctly. What do you want to achieve with it: higher customer satisfaction, cost savings, shift in the use of channels, or gain experience with new technology? And what do you want NOT ALLOWED achieve with?

With a chatbot you automate a number of tasks that are otherwise performed by employees.

With a chatbot you can also serve your customers outside opening hours, without (initially) waiting. In addition, a chatbot provides consistent answers and there is asynchronous communication - it is not a problem if one of the two interlocutors waits a while before giving input. Furthermore, the customer can remain in the same channel: there is no need to call, texting is sufficient. These benefits can contribute to higher customer satisfaction.

Chatbot as a replacement for the IVR

(Chat)bots can also replace the IVR. The classic IVR – regardless of the input method: speech recognition or dialing – is aimed at the customer with one specific question. The customer can only make one choice. Making a choice with a multiple question becomes difficult. A bot or conversational IVR can recognize and act on multiple intents, for example by assigning the message directly to the employee with the right expertise.

Other goals: save costs, increase customer satisfaction

Chatbots can contribute to cost savings, because the automated handling of questions leaves room for agents to handle more complex questions. The NPS of the channel in question can also be increased if the chatbot is sufficiently effective in handling the appropriate questions. The use of a chatbot can also lead to higher employee satisfaction, because employees are relieved of routine questions.

You can measure customer satisfaction by asking the customer specifically on NPS or CES or satisfaction on the chatbot channel - preferably in combination with the handling by an employee.

After the why question comes the what question: determining which use cases are suitable for bot handling and creating an initial business case. After this you also have to determine on which channels and/or where on the website you will offer the bot.

Chatbots are good at one job. Which tasks do you choose?

For example, if there is already communication between customers and employees via live chat (on the website), WhatsApp or Facebook Messenger, you can use a chatbot there. You look at the conversations on the channel where you are going to use the chatbot. The questions asked by customers are clustered by subject and the implementation team identifies which questions lend themselves well to handling via a chatbot. This concerns the recognisability of synonyms, so that a chatbot is easily trainable, but also clearly structured processes and flows. Consider, for example, the question about returning a product, asked in 100 different ways by 100 different customers. Two thirds of these questions are used to train the chatbot, one third is used to test the chatbot.

You can automate multiple questions in an environment.

How do you make it clear to the user/customer what your chatbot can and cannot do?

If a chatbot cannot process an opening question from a customer, you can transfer that conversation directly (and in fact unseen) to an employee. If you have not implemented this feature, it is advisable to clearly explain what the bot can do for a customer, such as 'book a ticket' or 'make an appointment'. If a question arises that the chatbot cannot handle, let the chatbot explain this and make sure that the chat history is visible to the employee together with the conversation.

What are the pitfalls when using chatbots?

According to Forrester brands make the wrong decisions when using chatbots. They focus too much on automating employee tasks and cutting costs. Usually this is at the expense of the customer experience. According to Forrester, it is naive to think that chatbots will contribute to a reduction in FTE. The Forrester analysts reason (based on research) as follows: the number of digital interactions will increase. Simple and routine issues can indeed be automated, but the number of complex tasks will also continue to grow. Contact center decision makers expect the staff size of their customer service operation to continue to increase in the coming years. The deployment of employees remains the largest cost item, so it is smarter to also focus on the use of chatbot technology to support employees. Chatbots can therefore relieve the pressure on the personnel budget.

A chatbot contributes to a good customer experience – Forrester concluded on the basis of research that 88 percent of companies using chatbots see a decrease in volumes of email, phone calls and live chats. But only 12 percent report an improvement in the Net Promoter Score.

Dead-end bot – a chatbot that is unable to escalate to an employee. This means that for questions that your chatbot can't handle, a customer will see messages such as 'I'm not trained for this', 'I can't help you with this', followed by the advice 'contact customer service. ' Despite an investment of time by the customer, a dead end bot does not provide a solution, but it does provide additional contact through other channels. Companies that deliberately create a chatbot that should be able to do everything without human cooperation are making a strategic fallacy.

No warm handover – when a chatbot crashes, it is essential that the conversation history and any context is transferred to the agent who takes over the contact. If you omit this integration, it will lead to longer conversation times and frustration for the customer. Outside opening hours, you can have the chatbot offer the customer the option to pass on the conversation to an employee who will call or text back the next day.

No knowledge, time or money for chatbot maintenance and management – when information, products, prices, processes or procedures change, the designed dialogues must be adapted. A chatbot also needs several optimization steps at the start.

Underestimating the impact of chatbot on your customer service staffing – when a call needs to be taken over by a live agent, a live agent may not be immediately available. If the waiting time (whether or not visible to the customer) increases, this is not conducive to the customer experience: it then contributes to the second disappointment. When the customer initiates a live web chat session, a response time of 30-60 seconds is considered maximum; the customer usually sits looking at their screen and waiting. With WhatsApp or Messenger, the consumer is used to not immediately receiving a response from a live employee.

Not acknowledging that a chatbot is bounded when the chatbot has handled a question correctly, a customer may want to ask a second question. Depending on your design, a chatbot may or may not handle a follow-up question or a second, different question. In certain cases, a chatbot session can therefore also lead to a live interaction if it is successfully handled.

If you want your chatbot to be able to handle a second question, it is important to have the chatbot close with "Is there anything else I can help you with?" and then start again at intent recognition.

What kind of technology do you use to make a chatbot?

A chatbot must be able to perform three tasks: 1. understand what the customer means with his or her message (intent recognition), 2. manage the dialogue and 3. be able to formulate an answer in natural language. In addition to chatbot technology, a good platform also contains facilities for integrating chatbot conversations into other systems. Think of CRM systems, models for specific languages ​​and specific knowledge (jargon) and concepts-devices (ontologies) or facilities such as location recognition. For most transactional dialogues, a chatbot will need to be able to exchange customer data with a CRM system or another type of system containing information about the customer and, for example, order status.

There are several 'low code' platforms for developing chatbots, of which Google Dialogue Flow is the best known.

What does an implementation process generally look like?

In addition to a clear objective (which answers the why question) and a project plan, prior to implementation you also need:

• procedures for handling chat sessions (opening, response time, waiting time notifications, keeping open, closing, saving, identification of the customer, etc.).
• indication of costs and benefits
• proposals for the management of the chatbot
• knowledge and expertise around data (Json) and (choice and use of) software platform
• an image of the integrations required with surrounding (CRM, order system) or underlying (knowledge bases, databases) systems
• an evaluation method including timeline and criteria for success

What data do you need?

To be able to build a chatbot you need conversational data – literally: written conversations between customer employees, where those conversations have taken place in the channel where you want to replace part of the employee tasks with a chatbot.

Do not use phone calls for chatbot dialogue design, as an excess of irrelevant information leads to a lower degree of recognition. If no data is available from chat conversations between employees and customers, it is advisable to start with live chat (web chat) first, in order to collect this data. In most organizations, the necessary volume of conversational data will be collected within a few weeks. These conversations are analyzed and classified, so that the most important goals for the chatbot (what customer intentions should the chatbot be able to handle) can be determined. Then you will work out the intentions in dialogues. This conversational data serves as a basis. Then you will test the chatbot with previously unused dialogues. Then you test the chatbot with the help of employees who will ask questions. In the final phase, the chatbot goes live and additional optimization steps follow.

In many cases, bot solutions offer standard integration options with knowledge bases, so that the bot can search the knowledge base for the right answer for non-transactional questions. This way you only have to manage your content in one place and you can also use your chatbot for information questions in addition to transactional questions.

How do you put together an implementation team?

The team that will take care of the chatbot implementation preferably consists of the following roles or competences:

• an architect or IT manager who can assist with the integration
• two to three experienced customer service employees, who are familiar with dialogues with the customer
• privacy or security officer for compliance
• marketing or customer service for the tone of voice of the chatbot. For example, you will have to determine how formally the chatbot communicates and to what extent you incorporate humor. Chatbots are often tested for funny questions.

How much time should you allocate for the implementation?

When data is available, the process from analysis to building a simple chatbot with one intent (task) usually takes several months, including integration and compliance checks. In some cases, building a bot takes significantly less time than the evaluation and learning process that comes afterwards.

Who does the management?

Keep the management of an existing chatbot – which takes about a few hours a week – in your own hands and make sure you have trained two employees. This way you can switch quickly in the event of calamities or make adjustments. The appropriate persons for this are experts or team leaders within customer service. This also keeps you the knowledge carrier of the content and the tone of voice of the chatbot.

What costs should you take into account?

• implementation costs
• management costs (FTE)
• costs for the project team that periodically expands the chatbot with new intentions
• costs for the chatbot platform such as, OBI4wan, Dialogflow, Inbenta,
• costs for integration with solutions such as Topdesk, Zendesk, Freshdesk, Genesys, Avaya, Steam, etc.

How do you arrange the escalation to a live agent?

You can consider the chatbot as a 'digital employee'. When the conversation needs to be transferred to a live employee, a transfer is made to a queue in which live employees are active. In principle, you send along the dialogue that has taken place. Depending on the settings, you can prioritize such a transferred call and assign it to specific employees. Nothing changes for the customer; on the other side of the chatbot window, the conversation is continued by an employee. This is often clearly stated, with or without a custom avatar and name.

How do you arrange the storage of conversations?

Chat sessions in the open part of the website usually do not contain customer identification data. The conversation then remains anonymous until customer data is requested.

Chat sessions in a private my environment are linked to the customer's identity and can be added to a CRM record in this way. Chat sessions via WhatsApp contain at least the mobile number; that can be used for identification.

Part of the implementation is describing procedures about the storage and retention period of chat sessions and the duration with which a chat session remains open (active or inactive) or is closed. It is also possible at the end of the chat session to e-mail a record of the entire conversation to an address specified by the customer.

What is the payback period?

The payback period is determined by the number of questions that are automated via the chatbot. By combining the volume of these questions and their average call duration and any after-work time (also in the back office), you get a global indication of the number of FTEs of customer service employees that you can save per year. You will have to correct this with the number of escalations and the number of follow-up questions.

Sometimes it is possible to use a bot as a planner for a live conversation between customer and employee. Such a bot can, for example, search for information based on intention recognition or verification layers and present it to the employee. This can, for example, shorten the average call duration.

What is the future of chatbots?

What will come after the current generation of chatbots? Technology is developing rapidly. Existing chatbots have a lot of data, but the chatbot technology itself is also getting richer. For example, voice bot solutions such as those from Alexa, Google or Siri can analyze intonation and emotion. In addition, the voicebots of the virtual assistants from Apple, Google and Amazon will gradually be integrated into customer service environments. This allows customers to ask their questions to the voice assistant instead of having to contact a contact center or go to a website for a web chat session: "Hey Siri, ask bol.com where my package is?"

Chatbots will always remain a custom solution to a certain level and will therefore have to be trained in the domain of the organization, just like real employees. However, it is expected that pre-trained chatbots will come onto the market in the near future that are equipped in advance with basic knowledge about a certain process, a certain market or sector.

Would you like to know more about practical examples, cases and the latest news about chatbots? View the Chatbots Dossier

This article was produced in collaboration with OBI4wan, Glue and Effecta

Dossier, Featured, Technology OBI4wan