While there is both hype and promise surrounding the potential of AI assisting humans in the Contact Center, I believe the inverse is actually the most effective starting point.
Whereby humans assist and train AI Digital Agents to improve their interactions with, humans. We refer to this as Human-Assisted Conversational AI, or Human-in-the-Loop, and its a key reason why Xaqt's Conversational Service Automation platform and Virtual Agents continually outperform the market in terms of accuracy and the quality of customer experience we deliver.
Steepen the (Machine Learning) Curve
The most common feedback we hear from companies that have attempted to build their own chatbots or natural language IVR with products like Amazon Lex, Google Dialogueflow or Twilio Autopilot is that they require a lot of training data to make the bot effective enough to handle actual customer conversations. Most organizations lack this critical data and skills. They also find it difficult to know exactly when the bot is working as intended and when it's not.
Despite what several vendors tout, Virtual Agents have not yet reached the point where you can just turn one on and expect it to work effectively. Voice automation is hard and, much like a live agent, there is a learning curve that each bot must go through before it knows enough to interact with an actual customer. Depending on your use case and the complexity of the call type, the data and training requirements can vary wildly.
With Conversational AI, you want to accelerate the learning curve and achieve optimal knowledge for each customer interaction. This is not a one time process, but rather a continuum. Just we humans continually learn, so do natural language understanding models.
What is Human-Assisted AI (Human-in-the-Loop)?
Unsupervised Conversational AI and Virtual Agents are like Unsupervised Live Agents. Left to their own devices, they can wreck a lot of havoc on your customer experience but when provided the right coaching and training, they're an incredible asset.
The difference between training a Virtual Agent and live agents is that you are training one "agent" that can handle an unlimited number of interactions rather than training an entire call center full of people. The economies of scale and financial impact are profound.
Xaqt's Human-Assisted AI process is similar to that of a Quality Assurance Analyst in the traditional contact center sense combined with that of supervisor or trainer.
With a Human-in-the-Loop, the bot's interactions are reviewed, scored and labeled for supervised learning. Typically, all interactions that fail predetermined criteria or checkpoints are manually reviewed. Additionally, a statistically relevant sample of successful interactions are manually scored as well to ensure positive reinforcement.
Unlike vendors that do not have the capability of recording calls in their Conversational IVR, we do. Our QA process ensures that the recorded conversation can be reviewed along side the transcript produced to identify exactly where issues arise. Many times, voicebots fail due to errors in the speech-to-text transcript, particularly with similar phonetically sounding words such as “where and “wear” (see case study below). This also enables us to improve model accuracy for things like understanding different dialects and accents that often get lost with typical speech-to-text engines.
The "Human in the Loop" AI process ensures that any exceptions are reviewed to determine their root cause, such as either a failure in the model or an unknown process request. Our workflow orchestration engine can automatically route process flow change or knowledge requests based on the expert workflow defined. Much like an agent might escalate a call or question to their supervisor.
Our unique approach combines best in class technology with a robust process for providing the virtual agent with qualitative feedback and ensuring that it can react to changing environmental variables.
Case Study: The COVID-19 Mask Conundrum
Information changes quickly. Not only do bots need access to the most up to date information, but they need to understand how callers are going to ask for that information in order to surface the correct and most relevant answer.
As example, let's take a look at an all too familiar scenario created by COVID-19 and the changing guidance from the CDC and local communities about wearing masks. Initially the CDC stated that masks were only needed by healthcare workers. However, two months later masks are now required to enter most establishments.
Xaqt deployed several COVID-19 related IVRs and chatbots on our Conversational platform to support our City customers in addressing questions from citizens.
Let's look at a sample of questions that citizens asked, and how they evolved over the course of just a few weeks.
- Do I need to wear a mask?
- Where do I need to wear a mask?
- Where do I get a mask?
- Can I get a free mask?
- What type of mask do I need?
- Does my child need to wear a mask in the park?
Most bots would pick-up on the phrase "wear a mask" and provide callers with a standard answer. However, providing the correct answer to these questions requires the natural language engine to understand the nuances of each question.
Rapid response virtual agents or generic bot templates developed by the cloud providers (ie. Google Dialogueflow, Twilio Autopilot) were developed on the CDCs initial guidance, but as the environment changed the bots were not updated, maintained or trained on the new information thus leaving citizens with outdated information or requiring Cities to allocate staff to monitor and retrain the bots themselves.
Enter Human-in-the-Loop. With Xaqt's Human-Assisted AI process and performance dashboards, we were able to detect in real-time the new questions that callers were asking and immediately worked with our clients to maintain up-to-date information, while our Computational Linguists trained the language models based on the nuances of each new phrase.
The result was, concerned citizens and callers were met with an empathic virtual agent that provided up-to-the-minute information, and abstracted away the complexity of bot training from our customers.
