In a previous blog post, I discussed using AI in the call center with virtual assistance. In this post, I want to go further to dissect the various pieces, components and decisions that need to be made to roll out a successful chatbot solution.
Contact Center Changes
For a long time, organizations relied on call centers to support and engage with their customers, relying on technology that was invented in 1876 ― the telephone. This invention opened a whole new world of communication possibilities and changed the customer-business relationship to a point where customers did not have to travel to a business’s brick-and-mortar location to ask questions or receive support. The telephone revolutionized how we communicate, but it has its limitations.
For instance, an agent can only talk to a single customer at once, leaving customers waiting in a queue once all the support agents are engaged. Organizations also tend to have lower per-call talk time and high call volume. This ratio typically means that agents are performing rather menial tasks. This could include looking up an order’s status, making a payment or scheduling an appointment. In order to meet fluctuating demand, organizations are in a constant cycle of staffing up or down depending on call volume.
Chatbots offer a way for organizations to meet more simple customer needs, allowing your organization to focus your contact center personnel (and budget) on higher level customer needs. Follow these steps as you evaluate a chatbot rollout in your contact center.
Step 1: Identifying Suitable Chatbot Workstreams
A great place to start is to review call transcripts or recordings for calls that have a short duration, particularly if there is a high number of calls with short durations. This is a very important step as it will dictate how and where the chatbot will be deployed. Also, something as simple as interviewing the agents to understand what sort tasks they’re performing most frequently, and also to help determine whether or not a human is needed to perform those tasks.
CDW offers a business analyst solution to assist organizations with this task. With this solution, a CDW business analyst visits with members of an organization’s IT and contact center leadership and reviews their operations, all the way down to evaluating how the agents handle customer requests.
Step 2: Choosing the Correct Chatbot Model
Will an agent need to be available for each chatbot session? Will the chatbot be available after hours? If so, which tasks can it perform? Here are the three modes available for chatbots:
- Autonomous/interviewer: Able to handle the entire interaction without agent assistance by not offering to speak to an agent. If customers need additional assistance, they’re directed to call, email or even request an agent to contact them via another method. This type of chatbot interviews the customer by asking some simple questions to guide the customer to make a purchase decision, lookup order status or basic troubleshooting steps.
- Live handoff: This mode has all the same features as the autonomous, but offers the ability for the customer to be connected to an agent if the chatbot cannot assist or doesn’t understand the request. Once the chatbot has gotten to the end of its capabilities, it will queue the customer until an agent can take over the conversation. Once the conversation is transferred the agent reviews the conversation and takes over. This can lead to customers waiting in queue after the chatbot has reached its conclusion. The issue, though, can be something as simple as insufficient training or a nomenclature issue that can stop the entire workflow.
- Hybrid chat: This is the newest method of introducing chatbots, which includes the best of both worlds. With hybrid, a chatbot handles the conversation (from a customer prospective) from start to finish. Initially the chatbot operates in an autonomous mode with one caveat, it’s being supervised by an agent. If the chatbot should get to a point where it no longer can handle the conversation, the agent monitoring the chats gets alerted. Agents are given suggestions on the response to send to the customer or they can intervene and take over entirely.
Step 3: Deciding on Chat Platforms
There are many chat platforms available but not all platforms are created equal nor are they all relevant to every customer base. For example, an organization that has a direct-to-consumer product with a strong social media presence may opt to include Facebook messenger. Organizations that want to handle sensitive information during these chat sessions may opt to use a secure chat platform such as WhatsApp where the entire conversation is encrypted.
Step 4: The Chatbot Framework
The chat framework can route the chat messages to various natural language understanding platforms (more on this to follow), provide reporting and can bolt on to an existing contact center solution. This can be a home-grown solution or a packaged solution (yet customizable) such as ExpertFlow.
ExpertFlow is a great option here as it features an open API, meaning customization and integrations can be done to fit just about any use case. It’s compatible with the Cisco on-premises contact center solutions and features reporting tie-ins to Cisco Unified Intelligence Center (CUIC). ExpertFlow also has tie-ins to just about every chat platform available and can create a connector if one doesn’t exist. You can learn more about it here: https://www.expertflow.com/chat/.
Step 5: Choosing a Natural Language Understanding Platform
A Natural Language Understanding (NLU) platform evaluates a text string and attempts to decipher the author’s intent. An intent is the most basic task that is being requested by the customer. A customer may need to do something simple such as paying a bill or scheduling an appointment. Since a chatbot is a form of AI that attempts to emulate human behavior, it must be able to decipher what a customer is requesting or validate data being given without having a list of keywords or phrases.
The chatbot can first ask a simple question such as “How can I help?” If the customer were to make a statement “Can’t login to my email,” the chatbot must be trained to know if it is dealing with a technical or sales issue and be able to access the information to offer the customer troubleshooting steps or where next to direct the customer.
Another function of the NLU platform is to be able to set customer input into a pragmatic format to be passed to backend systems and then present data back to the customer in a human, readable format with context. A simple example of this is a customer requesting an appointment for “a week from today.” The NLU platform will look at the current day of the week and convert to a datetime stamp.
There are several cloud-based NLU platforms out there. One thing to be mindful of is that like most cloud-based services, there is a subscription cost associated based on usage. These prices are generally very cheap, but is still a recurring cost to factor in. Here are some platforms to consider:
- Google DialogFlow (Google also has a larger suite of features in their Google CCAI package such as sentiment analysis or live transcription.)
- IBM Watson
- Amazon Lex
- Microsoft LUIS
For on-premises only deployments, RASA is available and similarly featured as a cloud offering.
Step 6: Chatbot API Connector Service
An API connector service is what allows the NLU platform to perform an action. The most common method is by creating a webhook service for the NLU platform to send variables parsed from customer input which then is processed by the API connector service.
These actions are generally API calls to external systems, which then formats the response that will be directed to the customer mimicking a human response. Likewise, the NLU platform can detect ambiguity and either ask clarifying questions or respond saying the customer request is not understood. This portion will require custom development by a partner or a staff developer. Note: If using a cloud-based NLU platform, this service must be reachable from the internet.
Step 7: Dialogue Design
This step entails producing a script or template for the way the conversation will flow. This includes how the chatbot responds to the customer to make it seem not so obvious that the customer is speaking to a computer. This includes scripting responses with multiple sentence syntax to be used at random for each intent (Examples: “Hi, how are you?” And “G’day! How’s your day going?”).
While this step may seem trivial, it’s the data that will be used to train the NLU platform to recognize requests that allows customers to speak freely without a cheat sheet of key words. The dialog design changes over time by evaluating NLU historical data to find alternate wording for requests or maybe acronyms that were not accounted for during the initial deployment.
Chatbots are a great way to augment an agent workforce and offer your customers a friendly alternative method for self-service. Using hybrid chat, customers have a uniform experience, with agents monitoring and guiding the conversation for a consistent experience.
One additional tidbit ― the same conversation structure designed for that chatbot can also be used for voice by adding automatic speech recognition, transcription and text-to-speech services. The customer request is transcribed into text and then passed to the NLU platform to parse the intent. Responses from the NLU system are then spoken by the text-to-speech engine.
For customers to have a pleasant experience and for you to realize your return on investment, it’s important to do proper discovery, planning, deployment and tuning after you go live. Reach out to your account manager to get more information on how CDW can assist with chatbots.