This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them. Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.
Decoding Dialogflow: Proper Care and Feeding of IVAs
In the 1984 Terminator movie, Arnold Schwarzenegger, a cyborg hit man, uses a phrase that has sort of become legend: “I’ll be back.” This phrase can also convey meaning when applied to the less violent, but very important, intelligent virtual agent (IVA) cyber creations being rolled out in contact centers around the world. For IVAs to work well and to provide lasting value, they need monitoring and continuous improvement by the people and organizations that create them. Hence, any developer that creates V1.0 of an IVA needs to think in terms of “I’ll be back.”
Monitoring and Continuous Improvement
My industry colleague, Art Schoeller of Forrester Research, said this of IVAs during a Talkdesk-sponsored webinar: “It takes work to make them work.” He was referring to the effort required to get IVAs not only operational, but also useful to the end users for whom they are intended. As a strategic element in an organization’s customer engagement strategy, IVAs require continuous oversight, maintenance, and optimization.
“Bot rot” is a term one author used to describe what happens when an IVA is neglected. Unmaintained IVAs will ultimately lead to poor user experiences and low utilization rates.
Organizations thinking about rolling out IVAs need to think of them as they would any other software development initiative in terms of maintenance and improvement. To get good utilization rates and successful outcomes over time, people will need to be assigned to evaluate an IVA’s effectiveness and to make improvements. For example, when a customer is transferred from an IVA to a live agent, developers need to seek to understand why the transfer occurred:
- Was it because the bot didn’t understand the customer’s intent?
- Was it because the intent the customer wanted was not programmed into the bot?
- Was the customer’s speech or text thread unintelligible?
- Did the person not use the specific entities associated with an intent?
- If the confidence level prediction from the natural language processing (NLP) was too low to choose any of the available intents, why was it too low?
To resolve the issue of an IVA not understanding what a customer has said enough to choose an intent, IVA developer Interactions uses a patented process called “Adaptive Understanding.” When an Interactions-developed IVA does not have confidence that it has properly identified a user’s intent, it sends a small snippet of audio to an intent analyst. The analyst listens to the snippet, which is typically less than 15 seconds worth of audio, classifies the intent, and returns control back to the IVA. The IVA then sends the feedback to the machine learning algorithm, where it is trained to understand this audio the next time. This is an example of blended human-AI interaction, which Interactions has automated to a remarkable degree.
While Google CCAI and Dialogflow do not have Adaptive Understanding capability built in, they do have tools available so that people can examine when and where an IVA is not operating as expected. Organizations need to instrument their IVAs so that they can measure the success of the customer interactions, including conversation killers — i.e., customer disconnects —as well as areas where the IVA is underperforming, misunderstanding, or offering inaccurate responses.
It is normal for an IVA to have limited functionality in a first release, with more and more capabilities gradually added as usage and learning increase. The initial version on an IVA will likely be simple, yet good enough to meet some customer engagement automation needs. As confidence in the IVA increases, its capabilities will likely be expanded by adding more intents with their associated entities and training phrases.
When implementing an IVA, take care to protect its persona. A product manager or a small group of people should be responsible for assuring the IVA’s persona remains consistent. As outlined in a previous article in this series, “What Makes a Good Bot,” IVAs reflect your organization’s brand and image. The bot should behave and act in support of an overall corporate persona your brand is trying to impress on the market. The way messages are crafted, the words used, the tone of voice, the cadence, wit, and sass need to be designed so that the people using the IVA are continuously exposed to the brand’s values every time they engage with the IVA. As different developers touch the IVA, they must be careful not to adulterate the IVA’s persona.
As products evolve over time and as new products or services introduced, the IVA will need to be updated to reflect the changes. Additional intents may need to be added along with their entities and particular vocabularies. If the IVA is supposed to return product information, the knowledge base containing responses, articles, publications, and sources of information will also need updating. This will also be true for organizations using CCAI’s Agent Assist capabilities in which Dialogflow is constantly listening to conversations and in real-time submitting information that may be of value to the agent.
One simple way to gauge whether an IVA has been helpful is to ask the people using the IVA if it has been effective and how it could be improved. If the IVA can process speech-to-text, you can invite people to speak to the IVA to provide much more than simple Yes/No feedback. Inviting user input via NLP may be of value to the maintenance team charged with assuring the IVA continues to be useful and relevant. Another easy measure is whether people use the IVA more than once or if after calling in they immediately request a live agent. Re-use is an indirect indicator that an IVA is providing significant value.
Series Recap and Conclusions
When we started this series in April, CCAI was still in beta. Two key capabilities, Virtual Agent and Agent Assist, are now generally available exclusively through Google CCAI partners. The first contact center partners offering these services commercially are Avaya and Mitel. Others, including 8x8, Cisco, Five9, Genesys, Nice inContact, Salesforce, Twilio, and Vonage, will follow in 2020.
For customers of these contact centers vendors, CCAI democratizes NLP, making speech a viable and readily available interface for anyone to interact with a contact center. I predict that because NLP can flatten the IVR tree, speech will replace IVR as the preferred method for interfacing with contact centers over the next decade.
Some IVAs out on the market are pretty clunky and are not well designed; however, as NLP begins to front contact centers and as experience in conversational design becomes more widespread, IVAs generally will evolve and eventually excel at solving/automating many customer service issues.
In this series, we have discussed a number of topics that are generally relevant beyond the CCAI and Dialogflow scope. Topics that are relevant to other IVA vendors’ products and services are:
- Introduction to Building Intelligent Bots
- What Makes a Good Bot
- Engagement Reach & Customer Journey
- Collecting Data and Using a Lexicon
- Creating Virtual Agent “Intents”
- Getting to Interaction Specifics with Entities
- Enabling Voice in CCAI
- Fulfilling User Intents
- Integrating CCAI with Contact Centers
- Routing to a Human, Testing, and Going Live
- Creating Multilingual Bots
- IVA Maintenance and Continual Improvement
Although this is the final article in this series, I invite you to attend a session I will be moderating at Enterprise Connect 2020, “Conversational AI: Using Messaging, Speech, and Chat to Automate Customer Engagement.” (See the full conference program here, and register using the code NOJITTER to save $200 off the current rate.) And while the main focus in this series has been on Dialogflow and CCAI, there are many IVA vendors helping brands create high- value automation solutions for contact centers. In 2020, I will write articles on many of these, articulating their unique value propositions.